Reports and rumors of how Nvidia plans to prioritize production of its RTX 50-series graphics cards in 2026 are swirling after a CES with no new consumer GPU launches, followed by reports that the company is ending the production of some RTX 50-series gaming GPUs and moving them to end-of-life status.

We've received a comment from Nvidia on the matter. We also spoke with the CEO of Gigabyte during CES, and his comments provide context about the overall situation and outlined a rather simple calculation that Nvidia could use to determine which GPUs it will prioritize.

We asked Nvidia for comment on the recent news that some models are being cancelled and received the following statement: "Demand for GeForce RTX GPUs is strong, and memory supply is constrained. We continue to ship all GeForce SKUs and are working closely with our suppliers to maximize memory availability."

While that statement suggests that it's full steam ahead for existing RTX 50-series products, there's more to the story. Tom's Hardware's Paul Alcorn was able to sit down with Gigabyte CEO Eddie Lin at CES for a wide-ranging interview that suggests Nvidia will still prioritize production of some of its GPUs over others based on a rather straightforward calculation, and that we should expect some RTX 50-series products to be in relatively short supply as the year progresses.

Lin described Nvidia's potential GPU allocation strategy, which focuses on maximizing profit from the limited memory resource, as follows:

"They cannot produce only high-end or low-end [products]... but they can, for example, they have 1, 2, 3, 4, 5, five segments. They focus on 1, 3, and 5, and reduce the percentage on 2 and 4, because on 2 and 4, the revenue contribution per gigabyte of memory is lower. They will calculate how much revenue [each segment] contributes per gigabyte of memory." [Emphasis added]

Lin went on to share the example of a $300 GPU (like the RTX 5060), for which "the memory contributes $35 per GB of revenue, whereas for a $400 8GB GPU, that product would contribute $50 per GB of memory. For a $500 [card] with 16GB of memory, that puts you at only $32 of revenue per GB, then the [contribution] is lower."

Additionally, Lin noted that Gigabyte continues to receive bundled memory from Nvidia with its GPUs. Rumors have swirled that Nvidia is no longer providing board makers with memory, which would portend a dire situation for smaller-scale players if they were forced to buy memory on the open market. Other vendors could be subject to different agreements and conditions with Nvidia, but we haven't received any confirmation from vendors that Nvidia is no longer bundling memory.

Using Lin's framework, we can understand what graphics cards are likely to be prioritized and deprioritized for allocation in 2026. Consider the following table:

At the lower end of the market, the RTX 5060 Ti 8GB contributes $47.38 of gross revenue per gigabyte of GDDR7 compared to the RTX 5060's $37.38, meaning that the 5060 Ti will likely be prioritized for allocation despite its wildly underwhelming initial reception.

The RTX 5060 Ti 16GB is the most threatened card of the bunch by this measure, since as a byproduct of its MSRP and higher VRAM capacity, it contributes just $26.81 of revenue per GB of GDDR7 on board — the lowest of any RTX 50-series card.

Moving up the stack, the RTX 5070 and RTX 5070 Ti contribute the same gross revenue per gigabyte, meaning that the cheaper-to-produce 5070 will likely be favored over its Ti sibling (which uses a bigger, more power-hungry GPU and a more complex board design)—or that both cards are likely to be deprioritized in favor of the more profitable RTX 5060 Ti 8GB.

At the highest end, the RTX 5080 and RTX 5090 contribute nearly the same revenue per gig of VRAM, meaning that the RTX 5080 will likely take precedence for allocation of 2GB GDDR7 chips going forward due to its smaller GPU die (half the size of the RTX 5090's) and much less complex board design. It would also mean that the 32GB of VRAM needed to produce one large 5090 GPU would instead create two 16GB RTX 5080s, which would help with overall supply and possibly lead to more margin.

Just for fun, the RTX 5090 and RTX Pro 6000 Blackwell share the same GB202 GPU (albeit with differing SM counts), but even with 96GB of GDDR7 on board, the RTX Pro 6000 contributes a whopping 41% more revenue per GB of GDDR7 on board versus the 5090.

Of course, the RTX Pro 6000 uses 3GB GDDR7 chips in clamshell mode to achieve its memory capacity rather than the 2GB single-side parts on the RTX 5090, so it's not an apples-to-apples comparison.

But it does make it obvious why Nvidia may have foregone launching the RTX 50-series Super refresh at CES: the margins afforded by using 3GB packages on RTX Pro products are simply much more attractive than they would have been for GeForce cards that presumably would have sold for near the same MSRPs as non-Super cards did at launch.

Going forward, we expect that the RTX 5060 Ti 8GB, RTX 5070, and RTX 5080 may be the easiest cards to find on shelves, relatively speaking, while enthusiast favorites like the RTX 5060 Ti 16GB and RTX 5070 Ti will be in short supply. The writing is also on the wall for the RTX 5090 — we can already see the supply situation reflected in today's empty e-tail shelves and dramatically inflated prices from third-party sellers. We'll continue to monitor this situation and update our list of the best GPUs for gaming accordingly.

SK hynix has published its DRAM roadmap at its SK AI Summit 2025. While the plans are displayed in a very general way and do not reveal important specifics, they still show the direction of DRAM technology evolution and approximate timelines of the emergence of new technologies. Naturally, since SK hynix demonstrated the roadmap at an AI event, it has a clear bias towards AI servers.

Conventional types of DRAMs — such as DDR, GDDR, and LPDDR — will continue to serve the memory needs of AI servers for the foreseeable future, albeit for different applications. Meanwhile, HBM memory will continue to serve the bandwidth-hungry AI and HPC processors. SK hynix lists 3D DRAM as something that is expected to come in 2030, but details are scant, so we can only speculate what the technology will offer at its inception early next decade.

DDR5 will continue to offer the balance between cost, density, and performance for years to come, albeit in form factors like MRDIMM Gen2 that is set to arrive in 2026 – 2027 and support data transfer rates of 12,800 MT/s, or 2^nd Generation CXL memory expanders expected to hit the market in 2027 – 2028. As reported multiple times, DDR6 will only arrive in 2029 or 2030; before that, DDR5 will keep evolving.

LPDDR6 — which now features a host of data center-oriented capabilities on the silicon level — is projected to wed high-capacity, high-performance, and lowered power consumption towards the end of the decade. SK hynix expects SOCAMM2 modules based on LPDDR6 to arrive in the late 2020s, possibly when Nvidia rolls out its post-Vera CPUs, and will need a new memory subsystem. Interestingly, SK hynix plans to release LPDDR6-PIM (processing-in-module) solutions for specialized applications sometime in 2028.

As for GDDR7, it will remain a niche solution for inference accelerators, such as Nvidia's Rubin CPX, as it combines very high performance and relatively low cost (compared to HBM), but lacks much-needed capacity. SK hynix lists 'GDDR7-Next,' which probably means GDDR8, as the company is not known for developing Nvidia-specific solutions, unlike Micron, which has developed GDDR5X and GDDR6X for Nvidia.

The highest-performing DRAM solutions from SK hynix in the coming years will be HBM4, HBM4E, HBM5, and HBM5E memory solutions that will be released in 1.5 – 2-year cadences from now through 2031. Interestingly, it does not look like HBM5 — which presumably powers Nvidia's Feynman, that is set to land in late 2028 — will show up until 2029 or even 2030. For customers that need customized memory solutions, SK hynix will also offer custom HBM4E, HBM5, and HBM5E modules, though it remains to be seen how its customers plan to customize such devices.

As for high-bandwidth flash (HBF) products that promise to offer performance comparable to HBM and capacity comparable to 3D NAND, SK hynix does not expect them to arrive before 2030, as the company must develop all-new media as well as agree on the final specification with other makers of NAND memory, particularly SanDisk, which proposed the technology earlier this year.

Memory chip manufacturing giant SK hynix has confirmed that it’s preparing to roll out higher-capacity GDDR7 memory modules later this year. In its latest earnings call, the company said that it will be increasing the maximum capacity of its GDDR7 from 16GB (2GB per module) to 24GB (3GB per module). While the comment was made in reference to AI GPUs, it is hard not to dwell on the potential for these new memory chips in future consumer-grade GPUs.

Having more VRAM doesn’t necessarily mean better performance; however, in recent years, it has become clear that 8GB or even 12GB is just not enough for many modern games, especially at higher resolutions or when enabling ray tracing. Additionally, generative AI tools have become more common in creative and productivity workflows and benefit heavily from additional memory.

From a practical standpoint, this move could also help bring higher VRAM capacity to slightly more affordable GPUs, not just the flagship models. Using 3GB modules would potentially make it easier to build a 24GB card with just eight memory modules instead of 12, which would essentially help in reducing power, thermals, and cost.

SK hynix’s plans to introduce higher capacity GDDR7 memory modules coincide with recent rumors surrounding Nvidia’s next consumer GPU lineup. According to a recent report, the potential RTX 50 Super series could arrive with refreshed models featuring 24GB and 18GB of video memory. Meaning that these 3GB GDDR7 modules would be perfect if Nvidia wants to avoid stacking extra modules at the back of the PCB. Notably, Nvidia began using SK hynix GDDR7 chips only recently for its current-generation of RTX 50-series Blackwell graphics cards.

Of course, none of this is official confirmation. Nvidia hasn’t announced a Super refresh yet, and even if it does, there’s no guarantee whether it’ll make use of higher VRAM modules on every single SKU. But when your memory supplier starts talking about higher-capacity modules, it’s a strong hint that something bigger is on the table.

Beyond GDDR7, SK hynix says that it is pushing across its memory portfolio in response to the accelerating AI demands. It is preparing next-generation LPDDR memory modules for servers to enable energy-efficient AI inference in data center environments. It’s also continuing to lead in High Bandwidth Memory (HBM), with high-volume production of 12-layer HBM3E alongside plans to introduce HBM4 later this year. Meanwhile, the company is also on track to release enterprise-grade 321-layer NAND SSDs in the second half of 2025, targeting high-capacity storage solutions for AI and hyperscale workloads.

It has only been a few months since Nvidia officially announced its RTX 50-series graphics cards, but rumors are already pointing to a potential Super series refresh. According to a post on the Chiphell forums, a hit-and-miss source for hardware leaks, Nvidia might already be working on an RTX 5080 Super with 24GB and an RTX 5070 Super with 16GB of VRAM.

The key enabler behind these rumored memory configurations is said to be Nvidia’s use of 3GB GDDR7 memory modules. These modules essentially allow for more flexible VRAM amounts than the traditional 8GB or 16GB increments. By leveraging 3GB chips, Nvidia can gain the ability to offer configurations like 18GB (6×3GB) and 24GB (8×3GB) VRAM for its desktop class GPUs without making any major PCB redesigns.

This can also help Nvidia make refreshes or upgraded versions (like the aforementioned Super variants) easier to deliver without changing much hardware or firmware. With the higher amount of VRAM, these graphics cards will potentially have the ability to handle larger textures at higher resolutions, and more complex scenes with fewer performance drops.

Earlier this month, reports suggested that Nvidia has added SK hynix as a GDDR7 supplier for its RTX 50 series GPUs, after previously only sourcing from Samsung. On Chiphell, a user recently got their hands on a Gigabyte Gaming OC RTX 5070 Ti with SK hynix memory and found its overclocking potential similar to what we've seen from Samsung, via UNIKO's Hardware. While thermal and efficiency data are missing, this is nonetheless a positive indication.

Each generation, Nvidia casts a wide net to broaden its memory supplier base, ensuring multiple options. This is a common approach with many products, especially SSDs, many of which often ship with revised controllers or NAND flash modules later in their lifecycle. This shouldn't impact the average consumer as Nvidia likely validates all memory chips to run at a minimum of 28 Gbps (or 30 Gbps for the RTX 5080), at defined power, voltage, and temperature settings.

A user at Chiphell secured an RTX 5070 Ti from Gigabyte (SK hynix memory) earlier this week. Several days later, they attempted to increase the power limits by flashing a BIOS that was designated for the Aorus Master RTX 5070 Ti (Samsung memory), resultantly bricking the card. Fortunately, the damage wasn't permanent as their GPU was equipped with a dual BIOS, allowing them to switch to the secondary firmware.

Either way, to test the limits of their GPU's memory, the user achieved a memory clock of 2,125 MHz (34 Gbps) with overclocking, aligning with the majority of Samsung overclocks we've witnessed. This tells us that GDDR7 modules from SK hynix are just as capable as ones from Samsung when it comes to overclocking - with this sample.

One of the things that Nvidia does not exactly encourage its partners among add-in-board (AIB) manufacturers to do is to 'factory overclock' memory on graphics cards featuring its GPUs. However, it looks like MSI's Alexei 'Unwinder' has managed to unlock this possibility for GDDR7 memory on GeForce RTX 50-series products, as noticed by VideoCardz. Good news, it works for all GeForce RTX 50-series graphics cards. But, as always, there is a catch.

A new unofficial update for MSI Afterburner enables RTX 50-series GPU owners to push their memory overclocking limits by an additional 3 GT/s, achieving data transfer rates of up to 36 GT/s. While some may say that a mere 10% overclock — especially in the case of the GeForce RTX 5090 that has a wide memory interface and barely needs that extra bandwidth — is not a big deal, this is not as simple as it sounds.

GDDR7 memory ICs available today from Micron, Samsung, and SK hynix are rated for an up to 28 GT/s — 32 GT/s (Micron, SK Hynix, and Samsung have GDDR7, but the latter does not disclose its specs). Nvidia recommends reducing data transfer rates to 28 GT/s (except for RTX 5080, which is set to 30 GT/s), though the chips can handle higher speeds, which is where Unwinder's achievement represents a breakthrough as it 'unlocks' available capabilities of the memory chips.

This is particularly significant for GeForce RTX 5080-series boards that come with 32 GT/s chips that are 'down-clocked' to 30 GT/s by default. However, keep in mind that on the GPU side there are memory controller peculiarities and this represents risks in terms of overclocking.

The new update involves replacing a specific database file and is compatible only with MSI Afterburner version 4.6.6 Beta 5 Build 16555. Although this update might eventually be integrated into a future beta version, it remains to be seen when exactly this happens. Some users have instead turned to GPU Tweak III, Asus's software, which supports both AMD and Nvidia GPUs and avoids known bugs present in MSI Afterburner.

Ultimately, both MSI Afterburner and GPU Tweak III can now support GDDR7 memory overclocking on GeForce RTX 50-series graphics cards. This is especially beneficial for the GeForce RTX 5080-series users who aim to maximize performance beyond the default 30 GT/s limit. With the modified database, pushing speeds beyond 32 GT/s should now be achievable without issues.

Micron introduced its 16Gb DDR5 devices made on its new 1γ (1-gamma) fabrication process that uses EUV lithography, a first for Micron, today on March 25. The new IC not only delivers higher performance than its predecessor, but it also consumes less power and is poised to be cheaper to make. The company also said that its 1γ manufacturing technology (6th Generation 10nm-class node) will eventually be adopted for other DRAM products.

DDR5 at 9200 MT/s

Micron's lead 1γ product is the company's 16Gb (2GB) DDR5 IC that is rated for a 9200 MT/s data transfer rate at an industry-standard voltage of 1.1V. Compared to its predecessor — a 16Gb DDR5 IC made on 1β fabrication process — the new device consumes 20% less power and features a 30% higher bit density, which may translate into a comparable decrease in production cost once the new chips achieve yields comparable to that of 1β 16Gb DRAM devices. 

While Micron rates its latest 16Gb DDR5 ICs at 9200 MT/s, this speed bin is significantly higher than anything in the latest edition of the DDR5 specification. The company stresses that the chip can run at JEDEC-compliant speed grades just fine, and the higher speed bin will enable some future proofing and compatibility with next-generation CPUs. Micron also suggests that CUDIMMs or CXL-based memory modules could leverage higher-than-JEDEC speeds. DIMMs for enthusiasts will also likely adopt the new DRAMs for their post-10,000 MT/s modules.

Micron is currently sampling its 16Gb DDR5 ICs made on 1γ technology and products on their base (i.e., chips and modules) with laptop and server manufacturers and expects their qualifications to be completed in one or two quarters. That means we should see Micron's latest memory devices in retail products starting mid-2025. The company expects all types of memory modules — for desktops, laptops, and servers — to adopt its new memory chips.

Considering the fact that Micron's 1γ-based DRAMs offer a combination of valuable qualities for all market segments — enhanced performance for desktops as well as lower power consumption for notebooks and servers — we indeed expect the firm's latest 16 Gb DDR5 ICs to become quite popular when they hit the market.

Over time, Micron will use its 1γ fabrication technology to make other types of memory products, including GDDR7, LPDDR5X (at up to 9600 MT/s), and data center-grade products, so the node will become a workhorse for the company.

1γ manufacturing technology

Micron's 1γ manufacturing process is the company's first technology to adopt extreme ultraviolet lithography (EUV), something that other leading memory makers adopted years ago. It's been a while in coming and looks to offer significant benefits relative to existing product lines.

Micron did not disclose how many EUV layers the new production node uses, but we can speculate that the company uses EUV for critical layers that would otherwise require the usage of multi-patterning, which lengthens production cycles and can affect yields. Micron does say that 1γ uses EUV in conjunction with multi-patterning DUV techniques. Also, Micron's 1γ DRAM process technology adopts next-generation high-K metal gate technology and an all-new back-end-of-line (BEOL) circuitry.

" In addition to EUV adoption in 1γ, we have introduced our next generation high-K metal gate CMOS and advanced back-end-of-line processes, which together enable the 9200 MT/s [data transfer rate], a 15% performance improvement over 1β DRAM […] while reducing power by about 20% over 1β," said Shigeru Shiratake, senior vice president of DRAM Technology Development at Micron.

For now, Micron produces its 1γ DRAMs at its fabs in Japan, where the company's first EUV tool was installed in 2024. As the company ramps up production of 1γ memory, it will add more EUV systems to its fabs in Japan and Taiwan.

Despite the RTX 50 series debuting with newer and faster GDDR7 memory, almost every source indicates that Nvidia will stick with standard 16Gb (or 2GB) modules for these GPUs. However, according to a rumor from the renowned leaker Golden Pig Upgrade Pack, Nvidia may equip the RTX 5090 mobile with 24Gb (or 3GB) memory modules, resulting in 24GB of total video memory.

The RTX 50 series of GPUs from Nvidia codenamed "Blackwell" will be the first consumer offerings to launch with GDDR7 memory. This new type of memory offers several advantages over GDDR6 and GDDR6X, including higher speeds and improved densities. Samsung is developing 42.5 Gbps GDDR7 modules in 24Gb (3GB) flavors, expected to hit mass production soon. Word on the street is that RTX 50 mobile and desktop GPUs could be announced in parallel next week.

The RTX 50 product stack with the RTX 5090, RTX 5080, and RTX 5070 is expected to debut with less-dense 16Gb (2GB) GDDR7 memory modules. If this rumor holds water - something we'll probably find out next week - it begs the question: Why is Nvidia using denser 24Gb modules exclusively for the RTX 5090 mobile? A ransomware attack at Clevo many months back suggested the successor of the RTX 4090 mobile with 16GB of memory. Even though that report dates back to June, we suggest you still take this leak with a grain of salt.

There's a strong possibility that denser GDDR7 modules may be in limited supply, something that's expected to improve as the technology matures. At GTC 2024, Samsung and SK Hynix announced that the initial wave of GDDR7 memory will center around 16Gb capacities with plans to introduce 24Gb variants later on.

Echoing its predecessor, the RTX 5090 mobile is likely to be powered by the GB203 GPU die that's also rumored to be used for the RTX 5080 on desktop. The leaker reports that Nvidia is using higher-capacity modules to overcome the limitations of a 256-bit bus. This suggests that GB203 is physically limited to a 256-bit interface; enabling eight modules for either 16GB (16Gb / 2GB modules) or 24GB of memory (24Gb / 3GB modules).

A growing concern among desktop and laptop GPUs alike is the limited memory capacity. VRAM capacities in AD103 and GA103 GPUs were bound by the 256-bit interface. Consequently, even professional workstation GPUs namely the RTX A5500 Mobile (GA103) and the RTX 5000 Mobile Ada (AD103) offered just 16GB of memory.

With the RTX 5090 mobile rumored to launch with 24GB, expect much of the same for its workstation counterpart. Moreover, if Nvidia wants, budget Blackwell GPUs such as the RTX 5060 can easily come with 12GB of VRAM, even across a narrow 128-bit memory interface. Just food for thought but theoretically speaking, a 48GB RTX 5090 is entirely possible; not that Nvidia would release such a GPU as it'd likely eat into its data-center revenue.

Rumors suggest that Nvidia will launch the RTX 5080 with 30 Gbps GDDR7 memory, while the remaining Blackwell lineup will stick with slower 28 Gbps modules, via Benchlife as spotted by Harukaze on X. Do note that this also includes the flagship RTX 5090. Still, given the substantial disparities in CUDA core count and other specifications between the RTX 5090 and RTX 5080, a small memory nerf is unlikely to make a difference to the flagship.

Nvidia's RTX 50 series is expected to be the first GPU lineup to debut with GDDR7 memory. GDDR7 touts several advantages over its predecessor such as higher speeds and greater density. Upcoming GDDR7 offerings from Samsung promise speeds of up to 42.5 Gbps in 24Gb (3GB per module) capacities. Mainstream and budget GPUs suffer from narrow bus widths that limit VRAM capacities to a paltry 8GB or 12GB (128-bit/192-bit) using 16Gb modules. While newer 24Gb modules can theoretically improve these numbers up to 12GB and 18GB respectively, don't expect availability until 2025 or later, likely in the form of a SUPER refresh.

With that context in mind, Benchlife's report suggests that Nvidia will use slower 28 Gbps GDDR7 modules across all its Blackwell series of GPUs, except for the RTX 5080. Like its predecessor, the RTX 5080 is rumored to possess the fastest memory in the entire lineup, this time at 30 Gbps, allowing it to dish out 960 GB/s of bandwidth if we assume a 256-bit memory interface. Previous leaks suggest the RTX 5080 will ship with 16GB of VRAM and that configuration is only possible with 16Gb modules and a 256-bit interface, so our calculations stand consistent.

Adding to the mix, AMD's RDNA 4 allegedly uses a rather conservative memory setup with 18 Gbps GDDR6 modules. Likewise, Intel's Battlemage B580 adheres to a 12GB configuration featuring 19 Gbps GDDR6 memory. Nvidia's dominant position in the market gives them the leverage to charge a pretty penny for the latest tech, including GDDR7 VRAM.

Yesterday, a leak from Zotac confirmed that Nvidia is prepping to unveil the RTX 5090, RTX 5080, and RTX 5070 series at CES next month. What this suggests is that the RTX 5060 lineup could be pushed to Q2 25, aligning its release with the availability of 24Gb modules (at least from Samsung). This should allow for higher capacities, up to 12GB for the RTX 5060 but that's wishful thinking.

A shipping manifest has cropped up featuring different SKUs for Nvidia's upcoming Blackwell RTX 50-series GPUs. Discovered by harukaze5719 on X, the manifest shows seven variants, meaning we could see multiple iterations of the RTX 5090 and RTX 5080 at the very least, possibly with varying VRAM capacities.

The manifest showcases six GPU SKUs aimed at "development and testing purposes." All seven feature varying codenames: 699-1G144-0030-TS1, 699-1G144-0050-TS1, 699-1G144-0045-TS1, 691-1G145-2030-TS1, 699-1G147-0070-TS1, 699-1G147-0050-TS1, 699-1G147-0045-EB1.

It's impossible to tell what these models are specifically, but given Nvidia's past track history, it usually launches its highest-tier models first, those being the RTX 5090 and RTX 5080. As a result, these seven models could be multiple iterations of the RTX 5090 and RTX 5080 featuring either different variants and/or different board versions, with possibly an RTX 5070/5070 Ti SKU or two thrown in for good measure.

We expect at least two of the SKUs to potentially correlate to RTX 5080 24GB and 16GB flavors with the same bus width. To recap, preliminary specs for the RTX 5090 put it at 32GB of VRAM capacity, while the RTX 5080 will stick with 16GB. That would mean no 24GB card this generation, which has existed for the past three generations (Titan RTX, RTX 3090 / 3090 Ti, and RTX 4090).

It would be odd at best for Nvidia to ignore the 24GB capacity, as the gap between the RTX 5080's purported 16GB capacity and the 5090's 32GB is simply too big to ignore. An AI/prosumer-focused RTX 5080 with 24GB of VRAM would be the perfect middle-ground for users who might not need or cannot afford the RTX 5090 but still want the workflow capabilities that 24GB of VRAM offers.

Note that this isn't the same as what happened with the 'unlaunched' RTX 4080 12GB (which eventually became the RTX 4070 Ti). The 4080 12GB used a different die, AD104, with a 192-bit interface and fewer GPU processing cores. The RTX 5080 would likely have the same (or very similar) specs in both VRAM capacities, with only the amount of memory being different.

All this comes thanks to the GDDR7 standard and its new capacity options. GDDR7 will be the first graphics memory solution to offer capacities beyond 2GB per IC, with potential of 3GB variants coming in the future — and capacities up to 64Gb / 8GB could eventually become available. Higher capacities could prove hugely advantageous for next-gen GPUs, as it means manufacturers have more flexibility in choosing various memory capacities without being restricted by the bus width of a GPU's memory subsystem.

With 3GB ICs, the RTX 5080 would be able to offer 24GB of VRAM without adding or reducing the memory chip count, enabling it to retain its 256-bit wide interface. Depending on how Nvidia wants to operate in the age of GDDR7, it could also play the same trick with the RTX 5090 and give it 48GB of memory with 3GB ICs.

Of course, that's less likely as 48GB could put the 5090 into direct competition with Nvidia's future workstation graphics cards. More likely is that Nvidia will eventually offer a professional RTX GPU with 48GB on a 512-bit interface, and probably even 96GB by putting memory on both sides of the PCB. AI workload have proven to be voracious in their appetite for VRAM, after all.

SK hynix has just announced the introduction of its GDDR7 graphics memory, which it claims is the "industry’s best GDDR7." Offering 32 Gbps operating speeds, it performs 60% faster than current-generation GDDR6 memory which is limited to around 24 Gbps. It also promises a 50% improvement in power efficiency, potentially reducing the power requirements and heat generated by next-generation graphics cards that will use this new memory.

The company was able to achieve this jump in power efficiency by using new packaging techniques. For example, it increased heat dissipating substrate layers by 50% (from four layers to six), allowing the chip to release its built-up thermal energy much faster without increasing the PCB area it uses.

Even though the SK hynix GDDR7 graphics memory is great news for the next generation of GPUs, Samsung was actually the first chipmaker to complete the development of GDDR7 memory, in July 2023. However, on paper, SK hynix offers advantages in terms of power efficiency, as the latter only claims a 20% power efficiency improvement over GDDR6 (versus SK hynix’s 50% claim). Micron also revealed its GDDR7 chips early in July, which the company says will deliver a 30% improvement in both ray tracing and rasterization performance.

SK hynix plans to start mass production of GDDR7 memory in the third quarter of this year, putting it in line with rival memory manufacturers like Micron. The Korean company hasn’t dropped any hints about GDDR7 sampling by its partners, unlike Samsung and Micron, but we can assume that the company has already done that after showcasing its chips in Taiwan at Computex 2024 in June this year. Furthermore, the official launch of GDDR7 by SK hynix, alongside the third quarter commencement of mass production, means that it likely already has or is going through the final steps of testing.  

Aside from gaming, SK hynix also expects GDDR7 memory to be adopted in other industries, like 3D graphics, AI acceleration tasks, high-performance computing, autonomous driving, and more. Sangkwon Lee, head of DRAM Product Planning & Enablement at SK hynix says, “We will continue to work towards enhancing our position as the most trusted AI memory solution provider by strengthening the premium lineup further.” And with next-gen GPUs that sport GDDR7 memory slated to arrive in 4Q24, we’re excited to see how much power and performance we can extract from these graphic cards, like Nvidia’s 50-series GPUs.

What is GDDR7 memory? It's the next generation of graphics memory for GPUs like the upcoming Nvidia Blackwell RTX 50-series. It will be used in a variety of products over the coming years, providing a generational upgrade over the existing GDDR6 and GDDR6X solutions, which then boosts performance in gaming and other types of workloads. But there's a lot more going on beneath the name.

Ever since the second generation of GDDR memory—for "Graphics Double Data Rate," if you're wondering—rolled out, the pattern has been pretty clear. GDDR (formerly DDR SGRAM) arrived way back in 1998, and every few years, a new iteration has arrived, boasting higher speeds and bandwidth.

The current generation GDDR6 arrived in 2018 and was first used in the Nvidia RTX 20-series and AMD RX 5000-series GPUs, starting with speeds of 14 GT/s (giga-transfers per second — or alternatively Gbps for gigabits per second) and eventually topping out at 20 GT/S. There was also GDDR6X memory, used solely by Nvidia on higher tier RTX 30- and 40-series GPUs, with initial speeds of 19 GT/s that eventually reached up to 23 GT/s — at least in shipping products.

GDDDR7 has been on the horizon for several years now since Samsung first discussed the tech back in 2022, with the final JEDEC specifications released on March 5, 2024. All the major memory manufacturers — Micron, Samsung, and SK hynix — have already committed to supporting the standard, and chips should be in mass production now. We anticipate seeing the first retail products using GDDR7 this fall.

GDDR7 Speeds

GDDR7 will initially start at speeds of 32 GT/s — 60% higher than the fastest GDDR6 memory, and 33% higher than the fastest GDDR6X memory (though no products ever used 24 GT/s speeds). But that's just the baseline starting point.

Micron and Samsung have publicly disclosed plans to release GDDR7 with speeds of up to 36 GT/s, while SK hynix says it will have up to 40 GT/s. That last would double the bandwidth of the top GDDR6 solutions, and we will likely see such chips shipping in 2025.

Looking forward, plans are in place for GDDR7 to reach up to 48 GT/s. We don't expect to see such memory for at least another year or two, but conceivably we could eventually see even higher clocks, depending on what happens in the intervening years.

Actual memory clocks are not as high as the above numbers might suggest. As with GDDR6, GDDR7 will use a quad data rate (QDR) — so technically, it's more like GQDR7, but the naming people have settled on sticking with DDR. Data is also fetched from memory in chunks, and the base clocks are much lower than the names would suggest. In fact, even "QDR" is a bit of a misnomer, as GDDR6X has base clocks of 1188 MHz ("19Gbps") to 1438 MHz ("24Gbps"), while GDDR6 has base clocks of 1375–2500 MHz (11–20 Gbps).

Will there be GDDR7X memory?

The past two generations of GDDR have seen "X" variants, GDDR5X and GDDR6X. Both of these came from Micron, working in collaboration with Nvidia to create an even higher bandwidth variant of the base memory. GDDR5 topped out at 9 GT/s in the GTX 1060 6GB, while GDDR5X stretched to 12 GT/s. Similarly, GDDR6 maxed out at 20 GT/s, and GDDR6X pushed that to 24 GT/s. That's 33% faster with GDDR5X and 20% faster with GDDR6X.

So, could we see GDDR7X that pushes speeds beyond whatever level GDDR7 eventually reaches? We asked Micron about this at GTC 2024 and were told that nothing was officially in the works yet. Which does make sense—we didn't get GDDR6X until the second generation of graphics cards using GDDR6 came out, and GDDR7 isn't even shipping in products yet.

Micron isn't likely to discuss future plans for GDDR7X regardless of whether it's being worked on right now — just like Nvidia isn't talking about the future Rubin architecture for GPUs just yet. It wouldn't be surprising to see yet another Micron-Nvidia collaboration for GDDR7X, but we don't expect to hear about it one way or another for at least another two or three years. If it does become a thing, hopefully, it will provide at least a 20% boost in bandwidth, as with GDDR6X.

What products will use GDDR7?

At present, there is no official word on any products that will use GDDR7, but Nvidia's higher-tier Blackwell GPUs are widely rumored to be the first to use the new memory. We expect the first of those to arrive in the fall of 2024.

Earlier expectations were that AMD's RDNA 4 GPUs would also adopt the new memory type, but there are now indications that RDNA 4 will stick with GDDR6 memory. That's only a claim from a leaker, but there are also indications that AMD might be focusing on mainstream GPUs for its upcoming architecture. If that's correct, it wouldn't be too surprising to see a continued reliance on less expensive GDDR6 memory.

What about Intel's Battlemage GPUs? These are also targeting mainstream users, according to Intel representatives, and thus may go with GDDR6 as well. Or perhaps we could see a higher-end model with GDDR7 and mainstream solutions with GDDR6.

Whatever happens, current rumors suggest that AMD RDNA 4 may not arrive until 2025. There were whispers that Battlemage would ship this year, but now other rumblings say that it has also slipped into 2025. At present, there's no clear answer on when the various new graphics cards will ship.

GDDR7 could also be used on other devices, particularly with AI accelerators. The top AI accelerators have been using HBM memory types — mostly HBM3 and HBM3E now — but inference-focused designs could still benefit from the additional bandwidth that GDDR7 offers, even if it's not as dense a solution as HBM.

GDDR7 technical details

We've covered much about the speeds and bandwidth of GDDR7 as well as when and where we're likely to see it used, but what fundamental changes does the new memory bring?

One of the biggest changes will be a shrink in process node technology, from the current "10nm-class" to 21nm, down to 10nm to 15nm. Micron currently uses the smallest node for GDDR6/GDDR6X, calling it 10nm-class, but we anticipate it will move to a refined and/or smaller node with GDDR7. The same goes for SK hynix, which currently produces GDDR6 on a 21nm node, and Samsung, which also uses 10nm-class for GDDR6.

Current GDDR6 solutions typically use 1.35V, and GDDR7 will reduce that to 1.2V, with a potential 1.1V version also in the works (for lower clocks). This should reduce power requirements at equivalent performance, though the higher speeds may negate that advantage.

The biggest fundamental change with GDDR7 is that it will use PAM3 signaling, where GDDR6 uses NRZ (non return to zero) signaling — and GDDR6X uses PAM4 signaling. PAM3 (3-level pulse-amplitude modulation) reduces energy requirements compared to NRZ, while being less complex to implement than PAM4 (4-level PAM). That should make GDDR7 manufacturing equipment less complex and less expensive, though that doesn't mean it will be inexpensive.

GDDR7 will also support non-power-of-two configurations, and we expect to see 24Gb and, eventually, 48Gb memory chips. GDDR6 may also have 24Gb solutions coming, though no company has shipped a product using such a configuration at present. This means 50% more memory can be put on each 32-bit interface so that a typical 128-bit graphics card, for example, could have 12GB of VRAM instead of only 8GB.

Another change is that a 32-bit GDDR7 memory interface gets subdivided into four 8-bit channels, which helps facilitate fetching larger chunks of data. Where GDDR5 was an 8n prefetch, and GDDR6 was 16n, GDDR7 will have a 32n prefetch architecture. This is a way to pull larger amounts of data from DRAM while still operating at relatively low clocks.

GDDR7 also supports ECC (Error Correcting Code), which allows chips to continue functioning even if the occasional bit gets flipped. ECC can detect this and improve reliability, a critical factor as speeds and densities increase.

Looking beyond GDDR7

Just as sure as we've had GDDR2 through GDDR7 — as noted earlier, "GDDR1" was called DDR SGRAM — it's a safe bet we'll see GDDR8 in the future, probably four to seven years from now. The real question is what will happen after the seemingly inevitable GDDR9. Will we just add a digit and get GDDR10? Probably, though, we could also shift from DDR to QDR or ODR (Octal Data Rate) naming at some point instead.

GDDR2 from 2003 ran at a top speed of just 1 GT/s (Gbps), with a 32-bit chip yielding up to 4 GB/s of bandwidth. GDDR3 started seeing use the next year with a bandwidth of up to 8 GB/s per 32-bit chip. Nvidia skipped GDDR4 while AMD used it in a few GPUs from the X1000 and HD 2000 series in 2006–2007, with a top bandwidth of 9 GB/s. The jump to GDDR5 in 2009 brought a significant increase in bandwidth, with the slowest chips offering 16 GB/s, and later, GDDR5 would eventually double that to 32 GB/s.

Since GDDR5, the rate of introducing new variants has slowed down. GDDR5 stuck around for a good six years, and GDDR6 has done the same. It's likely that GDDR7 will be around for at least that long, coexisting alongside various forms of HBM (High Bandwidth Memory) used primarily in data centers and AI products. But at some point, even with up to 160 GB/s of bandwidth per 32-bit chip, even GDDR7 will eventually need to be replaced — and the engineers are likely already discussing ways to push even higher bandwidths for whatever comes next.

But right now, we're looking forward to seeing the first wave of GDDR7-equipped graphics cards. With more memory capacity and much higher bandwidth, GDDR7 will enable even higher levels of GPU compute. Those should arrive this fall if everything goes according to plan.

SK hynix was rather quick to retract a comment made by its booth representative at Computex who said that it plans to start volume production of GDDR7 memory in the first quarter of 2025. Roughly a quarter behind its rival Micron. The company told Tom's Hardware via email that its GDDR7 memory is on track for mass production in the fourth quarter of this year, when the relevant market opens up. This is in line with its rivals. 

SK hynix demonstrated its first GDDR7 memory chips at Computex 2024 as well as revealing its general plans for this type of SGRAM. The company's GDDR7 product line-up will include chips with 16Gb and 24Gb capacities and with data transfer rates of up to 40 GT/s. SK Hynix's first GDDR7 product will likely be a 16Gb SGRAM IC featuring a data transfer rate north from 30 GT/s, though the company yet has to formally announce its first GDDR7 devices, as unlike its competitors it still has not officially introduced them. 

Both Micron and Samsung have already said that they had begun sampling of their GDDR7 products with select partners. SK hynix, on the other hand, has not made such an announcement. Nonetheless, given the fact that the company has demonstrated its GDDR7 chips at Nvidia's GTC and Computex already, it is safe to say that it has working samples of these memory devices. 

GDDR7 comes in a 266-ball package, 86 pins more than GDDR6, as it features not two, but four, eight-bit channels per device and a different power delivery scheme. Yet, despite notable changes to packaging, the dimensions of GDDR7 memory chips remained unchanged and are at 14mm x 12mm. 

Micron claims that the first products featuring GDDR7 will be released in 2024. Although graphics cards are set to continue to be the largest application of GDDR memory, the AI accelerator market is expected to adopt this type of DRAM as well. As AI accelerators need both high memory capacity and bandwidth, GDDR7 promises to be an excellent match for inference accelerators, offering a more affordable alternative to HBM and better characteristics than commodity DDR5.

JEDEC has published the specification for the GDDR7 memory standard — next-generation memory that will be used for graphics cards — and AMD, Micron, Nvidia, Samsung, and SK hynix have all weighed in on the matter. We anticipate GDDR7 will be the memory of choice for high-end RDNA 4 and Blackwell GPUs, which are rumored to launch next year and vie for a spot on our list of the best graphics cards.

It has been nearly six years since the first graphics cards began supporting GDDR6 memory. That was Nvidia's RTX 20-series Turing architecture, which launched in September, 2018. The first RTX 2080 and RTX 2080 Ti GPUs with GDDR6 had the memory clocked at 14 Gbps (14 GT/s), providing 56 GB/s per device. Later solutions like AMD's RX 7900 XTX have clocked as high as 20 Gbps, with 80 GB/s.

Nvidia helped to create a faster alternative in GDDR6X, which started at 19 Gbps in the RTX 3080, and eventually went as high as 23 Gbps in the most recent RTX 4080 Super. Officially, Micron rates its GDDR6X chips as high as 24 Gbps, which would yield 96 GB/s per device.

GDDR7 is set to provide a massive generational increase in bandwidth. JEDEC's specification will eventually reach as high as 192 GB/s per device. That works out to a memory speed of 48 Gbps, double the fastest GDDR6X. It gets to that speed in a different way than prior memory solutions, however.

GDDR7 will use three levels of signaling (-1, 0, +1) to transmit three bits of data per two cycles, aka PAM3 signaling. That's a change from the NRZ (non-return-to-zero) signaling used in GDDR6, which transmits two bits over two cycles. That change alone accounts for a 50% improvement in data transmit efficiency, meaning the base clocks don't have to be twice as high as GDDR6. Other changes include the use of core independent linear-feedback shift register training patterns to improve accuracy and reduce the training time, and GDDR7 will have double the number of independent channels (four versus two in GDDR6).

None of this is new information, and Samsung revealed many of the key GDDR7 details last July. However, the publication of the JEDEC standard marks a key milestone and suggests the public availability and use of GDDR7 solutions is imminent — relatively speaking.

Nvidia's next-generation Blackwell architecture is expected to use GDDR7 when it launches. We will likely get a data center version of Blackwell in late 2024, but that will use HBM3E memory instead of GDDR7. The consumer-level products will most likely arrive in early 2025, and as usual, there will be professional and data center variants of those parts. AMD is also working on RDNA 4, and we expect it will also use GDDR7 — though don't be surprised if lower-tier parts from both companies still opt to stick with GDDR6 for cost reasons.

In either case, AMD or Nvidia, using GDDR7 at the highest speeds would potentially provide for up to 2,304 GB/s of bandwidth using today's widest 384-bit interfaces. Will we actually see such bandwidth? Perhaps not, as for example Nvidia's RTX 40-series GPUs with GDDR6X all use slightly lower than maximum clocks. Still, we could easily see double the bandwidth with the upcoming architectures.

When will those actually arrive? We're not ruling out a potential late-2024 launch. Nvidia's RTX 30-series launched in the fall of 2020, and the RTX 40-series came out in the fall of 2022. AMD's RX 6000-series likewise launched in late 2020, with the RX 7000-series arriving in late 2022. If both keep to the same two-year cadence, we could see GDDR7 graphics cards before the end of the year. But don't get your hopes up, as we still feel early 2025 is more likely.

On February 20th, 2024, Samsung is set to show off two versions of its latest GDDR7 graphics memory technology to other industry players at ISSCC 2024. The ISSCC, or International Solid-State Circuits Conference, is a global forum for manufacturers and other industry players to show off their latest SoCs and advances in solid-state circuits— including, yes, new iterations of DRAM and VRAM.

Samsung showing off GDDR7 at ISSCC this year isn't a huge surprise. After all, last year we heard announcements from both Micron and Samsung that they were planning to release GDDR7 soon, with Micron even specifying a 1H 2024 release window.

GDDR7 VRAM is expected to come with great improvements to not only bandwidth, but also power consumption (at the same performance level of GDDR6/X) thanks to the adoption of PAM3 signaling over traditional signaling methods. Of course, making the most of GDDR7 could still see it using comparable power to modern GDDR6 configurations, just with a higher degree of performance-per-watt. The next revision of USB4 is also expected to adopt PAM3 signaling for reduced power consumption, as well.

Two versions of GDDR7 VRAM are expected to be seen at ISSCC this year: a low-power 35.4 Gb/s per-pin GDDR7 from SK hynix, and a higher-power 37 Gb/s per-pin GDDR7 from Samsung. For your reference, GDDR6X's bandwidth per pin is roughly 19-24 Gigabits, according to Micron.

The low-power version is most likely being targeted at laptops, and its presentation, which is officially titled "A 35.4Gb/s/pin 16Gb GDDR7 with a Low-Power Clocking Architecture and PAM3 IO Circuitry" leans toward this interpretation. 

Meanwhile, the high-power GDDR7 presentation is titled "A 16Gb 37Gb/s GDDR7 DRAM with PAM3-Optimized TRX Equalization and ZQ Calibration." This most likely corresponds to the version of GDDR7 we can expect to see in desktop GPUs — perhaps even later this year, if Micron's past comments on its introducing GDDR7 in 1H 2024 still hold water.

Only time will tell how long it actually takes for us to see GDDR7 in shipping products, embedded into a graphics card or laptop for us end users to enjoy. Considering past comments and the timing of Samsung's upcoming presentation at ISSCC, though, GDDR7-equipped GPUs will likely find their way to us before the end of the year.

Micron on Thursday introduced its new 128GB DDR5-8000 RDIMM memory modules based on monolithic 32Gb DRAM ICs aimed at servers and shared its vision for future high-performance and high-capacity memory technologies set to arrive over the next five years. The company also shared a roadmap that includes several technologies not previously discussed publicly, including 256 GB DDR5-12800 sticks, HBM4E, CXL 2.0, and LPCAMM2.

Bandwidth Hungry Applications

When it comes to bandwidth-hungry applications, Micron expects them to keep using HBM and GDDR types of memory in the coming years. 24GB 8-Hi HBM3E stacks are set to arrive in early 2024 and offer bandwidth of over 1.2 TB/s per stack, which is set to tangibly increase the performance of AI training and inference. The company plans to refine its HBM3E lineup with 36GB 12-Hi HBM3E stacks by 2025, which will further increase memory capacity for HBM-supporting processors, but will not boost bandwidth. For example, Nvidia's H100 could use 216GB of HBM3E memory if the company had these stacks. In 2025, they would be used by a post-Blackwell GPU for AI and HPC compute. 

The true revolution for HBM will be HBM4, which is set to arrive by 2026. These stacks will feature a 2048-bit interface, requiring memory, processor, and packaging makers to work together closely to make it work properly. Yet, the reward promises to be quite tangible as each stack is expected to feature bandwidth of over 1.5 TB/s. As for capacity, Micron envisions 36GB to 48GB 12-Hi and 16-Hi stacks in the 2026 to 2027 timeframe. HBM4 will be followed by HBM4E in 2028, according to Micron. The extended version of HBM4 is projected to gain clocks and increase bandwidth towards 2+ TB/s and capacity to 48GB to 64GB per stack. 

HBM will remain the prerogative of the most expensive and bandwidth-demanding applications. Cheaper products, such as AI accelerators for inference and graphics cards, will keep using GDDR. Based on Micron's roadmap, the next version of GDDR — GDDR7 — is set to arrive by late 2024, boasting a data transfer rate of 32 GT/s (128 GB/s bandwidth per device) and capacities of 16 to 24Gb. As GDDR7 becomes more mature, it will speed up to 36 GT/s (144GB/s bandwidth per device) and gain per IC capacities higher than 24Gb (think 32Gb, 48Gb) sometime in late 2026.

Capacity

Servers need memory, and Micron's 128 GB DDR5-8000 RDIMM is just the beginning of Micron's 32Gb DRAM ICs. Expect the company to come up with more products based on these devices. 

In a bid to bring together capacity and performance, Micron expects to offer 128GB – 256GB MCRDIMM modules with a data transfer rate of 8800 MT/s in 2025, and then MRDIMMs with capacities of over 256GB and a data transfer rate of 12800 MT/s in 2026 or 2027.

For machines that need further memory expansion, Micron is poised to ship CXL 2.0-supporting expanders featuring 128GB – 256GB capacities and up to 36 GB/s of bandwidth over a PCIe interface. These will be followed by CXL 3.x-compliant expanders with over 72 GB/s of bandwidth and capacities of over 256 GB.

Low Power

For low-power applications, the industry will keep using LPDDR memory and based on Micron's roadmap, LPDDR5X with 8533 MT/s or 9600 MT/s data transfer rates will stay with us for a while. Meanwhile, for laptops and other applications that benefit from memory on modules, Micron is set to offer LPDDR5X-8533 16 to 128GB LPCAMM2 modules starting in 2025, and then LPDDR5X-9600 LPCAMM modules with capacities of 192GB and higher starting in mid-2026.

China's Commerce Minister Wang Wentao last week had a meeting with Sanjay Mehrotra, chief executive of Micron, expressing support for the company's expansion in China, reports Reuters. Meanwhile, there is no word whether the People's Republic is going to lift a ban on Micron's memory devices that are used for PCs used by government-controlled agencies and critical infrastructure.

"We welcome Micron Technology to continue to take root in the Chinese market and achieve better development under the premise of complying with Chinese laws and regulations," Wang said.

Micron is one of the companies that does not produce chips in China, but which has vast packaging operations in the country. The company's 3D NAND and DRAM memory ICs, made Singapore and Taiwan, are then packaged in China and distributed to makers of memory modules and solid-state drives, including the best SSDs with a PCIe 5.0 x4 interface.

Micron's China operations are vast. Back in June the company announced plans to invest some additional $600 million in its facilities and add 500 more jobs, increasing its headcount in the country to more than 4,500.

But the massive facilities that that Micron runs in China did not stop the People's Republic's government from banning its chips following a security assessment conducted by Cyberspace Administration of China (CAC). The organization concluded that Micron's products "have relatively serious potential network security issues, which pose a major security risk to my country's critical information infrastructure supply chain."

But CAC did not elaborate on the exact security threats that Micron's products may pose in particular and how a memory IC can affect reliability of supply chains. Given the concerns over national security, the ban on Micron memory devices was executed promptly, which naturally hit the company's revenue.

China's action toward Micron is largely interpreted as a retaliation to the United States' attempt to limit Chinese access to advanced manufacturing technologies. This context of geopolitical maneuvering also includes U.S. efforts to persuade its allies to limit exports of wafer fab equipment to China. 

Against this backdrop, the conversation between Wang Wentao and Sanjay Mehrotra reflects a softening of tensions, occurring in the run-up to a planned dialogue between the U.S. President Joe Biden and Chinese President Xi Jinping at the forthcoming Asia-Pacific Economic Cooperation summit in San Francisco, signaling a potential diplomatic warming.

Being the only maker of GDDR6X, Micron has been the primary supplier of GDDR memory for Nvidia since mid-2020 and will continue to be for at least a year. But with next-gen GDDR7 around the corner, the situation may change, and the company is already evaluating Samsung's GDDR7 components with the aim to use it with its next-generation GeForce RTX 50-series graphics products, reports BusinessKorea.

Samsung has been particularly vocal about GDDR7 in the last year or so and introduced the industry's first GDDR7 chips in mid-July. Samsung has already begun to ship GDDR7 samples to Nvidia "for the verification of its integration into next-gen systems," the report asserts. SK hynix is not far behind, with plans to finalize its own GDDR7 technology within this year, BusinessKorea reports. By contrast, Micron only said that it would introduce its first GDDR7 offerings in the first half of next year. 

The report suggests that given the timeline of GDDR7 introduction by leading DRAM makers, Nvidia may lean towards Samsung's and SK hynix’s offerings for its next-gen GeForce RTX 50-series graphics processors. There is a catch, though. Nvidia is interested in mating as many GDDR IC models with its GPUs as possible in order to encourage competition between DRAM suppliers and ultimately make GeForce graphics cards cheaper.

For now, Micron will remain the sole supplier of GDDR6X, which is used by Nvidia's range-topping GeForce RTX 4080 and 4090-series products, which are among the best graphics cards money can buy today. But the focus on GDDR7 by Samsung and SK hynix signals a shift in the competitive landscape of the GDDR market as there will be three suppliers to offer GDDR memory for Nvidia's next-generation high-end GPUs codenamed products, not one, like in the case of the GeForce RTX 30 and RTX 40-series.

GDDR SGRAM is crucial for gaming graphics cards, and these types of memory are particularly lucrative for DRAM makers like Micron, Samsung, and SK Hynix. The total available market of GDDR memory is expected to expand from $3.2 billion in 2018 to $4.8 billion by 2030, according to Industry Growth Insights. 

Micron was the first company to introduce 24Gb DDR5 memory devices as well as start shipping actual modules on their base last fall. As it turns out, as part of Micron's HBM3 Gen2 announcement, the company wants to maintain the lead and is prepping to mass produce 32Gb DDR5 ICs as well high-capacity memory modules in the first half of 2024, the company revealed today.

Micron's monolithic 32Gb DDR5 IC will be made on Micron's 1β (1-beta) fabrication technology, which is the company's most advanced production node and also the last fabrication process that does not use extreme ultraviolet lithography. For now, Micron has not disclosed the data transfer rates it expects from its 32Gb devices, though given the fact that these will be Micron's 3rd Generation DDR5 ICs, expect them to be reasonably fast.

One of the things that modern DRAM fabrication technologies allow memory makers to do is to build high-capacity monolithic memory devices. After Micron built its 24Gb memory IC on its 1α (1-alpha) process technology last year, it was logical for the company to proceed with a 32Gb device with its 1β node and this is exactly what it is doing. 

32Gb DDR5 DRAM ICs will be particularly useful for datacenter-grade memory modules as they obviously benefit from high-capacity DRAMs. But while 32Gb devices open doors to 1TB DDR5 modules (which use 32 8-Hi 32Gb stacks), Micron is not jumping the gun here and will only offer 128GB DDR5 modules based on these ICs next year. Going forward the company plans 192GB and 256GB DDR5 modules. 

Meanwhile, 512GB and 1TB memory sticks are not currently listed in the roadmap, possibly because Micron still considers such memory sticks as niche devices that it might make available for select clients. 

Speaking of niche types of memory products: Micron's roadmap published on Wednesday reiterates Micron's plans to mass-produce 16Gb and 24Gb GDDR7 memory chips with a 32 GT/s data transfer rates sometimes in mid-2024 as well as HBMNext memory in 36GB and 64GB stacks offering 1.5TB/s 2+ TB/s bandwidth per module sometime in 2026 or later.

In a quite unexpected twist, Samsung late on Thursday said that it had completed the development of the industry's first GDDR7 memory chip. The new device will feature a data transfer rate of 32 GT/s, use pulse-amplitude modulation (PAM3) signaling, and promise a 20% power efficiency improvement over GDDR6. To achieve this, Samsung had to implement several new technologies.

Samsung's first 16Gb GDDR7 device features a data transfer rate of 32 GT/s and therefore boasts a bandwidth of 128 GB/s, up significantly from 89.6 GB/s per chip provided by GDDR6X at 22.4 GT/s. To put it into perspective, a 384-bit memory subsystem featuring 32 GT/s GDDR7 chips would provide a whopping 1.536 TB/s of bandwidth, which by far exceeds GeForce RTX 4090's 1.008 TB/s.

To hit unprecedentedly high data transfer rates, GDDR7 uses PAM3 signaling, a kind of pulse amplitude modulation featuring three distinct signaling levels (-1, 0, and +1). This mechanism enables the transfer of three bits of data within two cycles, which is more efficient than the two-level NRZ, which is the method used by GDDR6. However, it is important to note that PAM3 signals are more complex to generate and decode than NRZ signals (which means additional power consumption), and they can be more susceptible to noise and interference. Meanwhile, it looks like the benefits of PAM3 outweigh its challenges, so it is set to be adopted by both GDDR7 and USB4 v2.

In addition to higher performance, Samsung's 32 GT/s GDDR7 chip is also said to feature a 20% improvement in power efficiency compared to 24 GT/s GDDR6, though Samsung does not specify how it measures power efficiency. Typically, memory makers tend to measure power per transferred bit, which is a fair thing to do, and from this point of view, GDDR7 promises to be more efficient than GDDR6. 

Meanwhile, this does not mean that GDDR7 memory chips and GDDR7 memory controllers will consume less than today's GDDR6 ICs and controllers. PAM3 encoding/decoding is more complex and will require more power. In fact, Samsung even goes on to say that it used an epoxy molding compound (EMC) with high thermal conductivity and a 70% lower thermal resistance for GDDR7 packaging to ensure that the active components (the IC itself) do not overheat, which is an indicator that GDDR7 memory devices are hotter than GDDR6 memory devices, especially when working at high clocks.

It is also noteworthy that Samsung's GDDR7 components will offer a low operating voltage option for applications like laptops, but the company does not disclose what kind of performance we should expect from such devices.

Truth be told, Samsung's announcement is a little bit shy of details. The company does not say when it plans to start mass production of its GDDR7 components and which process technology it is set to use. Given the cadence of new GPU architecture announcements by AMD and Nvidia — every two years — it is logical to expect next-generation graphics processors to hit the market in 2024, and they are more than likely to adopt GDDR7. 

Meanwhile, Samsung expects artificial intelligence, high-performance computing, and automotive applications to take advantage of GDDR7 as well, so perhaps some sort of AI or HPC ASICs may adopt GDDR7 ahead of GPUs.

"Our GDDR7 DRAM will help elevate user experiences in areas that require outstanding graphic performance, such as workstations, PCs and game consoles, and is expected to expand into future applications such as AI, high-performance computing (HPC) and automotive vehicles," said Yongcheol Bae, Executive Vice President of Memory Product Planning Team at Samsung Electronics. "The next-generation graphics DRAM will be brought to market in line with industry demand and we plan on continuing our leadership in the space."

Micron said on Wednesday that it would introduce its first GDDR7 memory chips in the first half of next year. The new type of memory promises to offer higher performance than GDDR6 and GDDR6X, but it will require brand-new memory controllers and therefore GPUs.

"We plan to introduce our next-generation G7 product on our industry-leading 1ß node in the first half of calendar year 2024," said Sanjay Mehrotra, chief executive of Micron.

GDDR7 SGRAM will be the next generation memory for GPUs that will be used for some of the best graphics cards as well as other devices that require high bandwidth, but do not necessarily need expensive HBM3 memory. Samsung envisions that GDDR7 will offer data transfer speeds in the range of 36 GT/s, though it remains to be seen when the novel type of SGRAM will offer this level of performance.

Earlier this year Cadence revealed that GDDR7 memory will use PAM3 signaling, which promises to let it boast with higher bandwidth compared to GDDR6 (which uses PAM2 or NRZ encoding) without complications and higher power consumption imposed by GDDR6X (which uses PAM4 signaling). 

It should be noted that formal introduction of a new type of memory does not necessarily mean its immediate commercial adoption. Since GDDR7 uses a completely different encoding than GDDR6 or GDDR6X, it will require all-new memory controllers and therefore GPUs. While it is logical to expect AMD, Intel, and Nvidia to introduce their next generation GPUs in 2024 or early in 2025, only these three companies know when exactly these graphics processors will ship.

Cadence already has its GDDR7 verification solution, so adopters can ensure that theire controllers and physical interfaces will be compliant with the GDDR7 specification eventually.

Samsung is looking to capitalize on its deep and varied revenue sources and intends to double-down on memory technology investments for the future. The company recently announced that it won't be reducing its capital expenditure on new memory technologies, including GDDR7 that will ultimately end up in the best graphics cards of the future and ultra-high-layer V-NAND for the best SSDs. Nor is it looking to reduce manufacturing, despite the slowing market conditions.

The company is confident that its spending plans are necessary to ensure technological dominance over its competitors, while cementing its position as the leading memory chip manufacturer. To do so, Samsung is betting on next-generation fabrication processes that will allow for scaling memory density in DRAM, VRAM, and the latest, densest NAND that can power the data-intensive workloads of the future.

In the DRAM field, the company announced that its latest, fifth-generation 10-nm class manufacturing (1b) will start pushing out volume chips by 2023. Exploratory research is being done in sub-10nm DRAM manufacturing employing new patterning, materials and architectural designs that include High-K gates. Samsung is leagues ahead of most of its rivals, which generally manufacture DRAM chips on 14nm class nodes.

Samsung's next-generation GDDR7 memory, which is sure to grace future graphics accelerators, is already in production. It offers speeds up to 36 Gbps, a doubling over GDDR6's 18 Gbps. At those data rates, a 384-bit memory bus would be able to deliver around 1.728 TB/s bandwidth — a large increase from the upcoming RTX 4090's 1 TB/s bandwidth. This will ensure GPU manufacturers have an adequate reservoir of bandwidth left without having to increase bus widths, as that would lead to more expensive PCBs and potentially and even worse impact on pricing.

As for NAND, where Samsung stands as the undisputed king of the hill, the company is now designing and prototyping its 9th and 10th generation V-NAND, with appropriate increases in layer density compared to the technology of today. Samsung is now shipping its 7th generation 176-layer V-NAND, with plans to release V-NAND chips based on its 8th generation 230-layer design by the end of the year. The latter will offer a 42% density increase with 512 Gb chips.

But Samsung is eyeing even more significant jumps in density, and expects to achieve a 1,000-layer V-NAND design by 2030. Samsung also continues to work on QLC (Quad-Level Cell) tech, hoping to boost performance while also increasing storage bit density.

“Even if the current situation is not good, we won’t change the course we have already charted,” said Han Jin-man, executive vice president and head of Samsung’s global memory sales and marketing division. “No one at Samsung is talking about cutting chip production for now.”

Several other memory manufacturers have announced cuts to their memory production in response to slowing demand, a situation that's partly attributable to the general macroeconomic situation. The slowing demand has already led to an oversupply in the memory market, which has been described by Micron's CEO as "unprecedented," leading the company to reduce capital spending in fiscal 2023 by more than 30% (about $8 billion), with a further 50% cut in spending on wafer-fab equipment.

In Japan, Kioxia also announced a 30% reduction in memory wafer output in hopes that it will be enough for stockpiled inventory to be absorbed. Both companies are attempting to reverse the downward pressure on NAND pricing, which has taken a beating throughout 2022 with more reductions on the horizon before the end of the year.

Samsung in contrast is choosing to double-down on its technology leadership even as its competitors are playing the safe game by reducing their output. To be fair, Samsung's multiple divisions and revenue streams somewhat shield it from cyclical market and demand conditions, so it's more of a case of who can do it rather than simply, blindly staying course. Samsung may end up helping DIY enthusiasts, thanks to its apparent refusal to lower manufacturing output. The company also believes that it can widen the technological gap with its competitors by staying the course on its investment roadmap.

As reported by Computerbase, Samsung used the Tech Day 2021 event to showcase its future memory technology roadmap. Unfortunately, Samsung did not allow slides or photographs of the event to be captured and shared, which is a return to the note-taking days of old. However, the company shared plans on memory development pursuits for the coming years, with expected revisions and developments of standard technologies such as DDR RAM, GDDR graphics memory, and HBM3.

While standards for the next-generation DDR6 memory still haven't been set by JEDEC (DDR5 is currently still in its initial adoption phase), work is already ongoing in establishing next-generation memory technologies. Samsung announced that standard DDR6 speeds are expected to hit 12,800 MTs - while overclocked DDR6 memory (as in, any operational frequency above JEDEC's standard) could hit a ceiling at around 17,000 MT/s. Paired with the expectation of doubled memory channels for each DDR6 stick (quad-channel memory sticks compared to DDR5's dual-channel connection) and quadrupled memory bank size (64 compared to DDR5's maximum 16), DDR6 should enable an incredible jump in pure throughput and memory capacity. The low-power version of DDR6, DDR6LP, will achieve the same 17,000 MT/s operational speeds, but at 20% lower energy consumption.

Samsung is also currently developing extensions on the technology that would enable GDDR6+ chips to operate at up to 27 Gbps. NVIDIA's all-powerful RTX 3090 and its GDDR6X memory, co-developed in partnership with Micron, hits 21 Gbps, while conventional GDDR6 modules top out at 18 Gbps. As a result, real improvements will have to wait for GDDR7. While the technology doesn't currently have a date for its debut, Samsung expects the technology to be engineered at up to 32 Gbps throughput, slightly less than double the highest bandwidth available with GDDR6.

Lastly, Samsung has confirmed that HBM3 development is running as scheduled. The company didn't confirm speeds or stack density in its Tech Day presentation and mentioned only that market availability is expected for Q2 2022. Samsung is taking a different approach to HBM3 than other memory manufacturers, however. SK Hynix, for instance, is developing 24GB memory chips with a 6.4 GT/s data transfer rate and a 1024-bit interface, providing a bandwidth of up to 819 GB/s. Samsung, on the other hand, has been exploring its Hybrid-Substrate Cube (H-Cube) technology; it aims mostly to reduce fabrication costs for HBM3 manufacturing and packaging technologies - particularly in stacks of six or higher HBM3 memory chips.