Micron says GDDR7 will provide a 30% improvement in gaming — both ray tracing and rasterization

Micron GDDR7
(Image credit: Micron)

Micron has revealed that its next-generation GDDR7 graphics memory technology will offer a 30% performance improvement for all gamers. Discovered by Harukaze5719 on X (Twitter), The next-generation graphics memory standard is specified as having a 30% improvement in frames per second over its GDDR6 counterpart (GDDR6 and G6X), entailing both ray tracing-focused titles and pure rasterization-only games. 

Specifically, Micron is saying these changes will occur with its upcoming 32Gb/s GDDR7 modules, in comparison to its GDDR6 (non-X) 20Gb/s version, which is currently one of the fastest speeds on the market for standard GDDR6 memory. Micron neglected to share what benchmarks were run to get its 30% metric, but 30% is a huge improvement regardless, assuming these uplifts are from memory alone.

There's a lot of evidence to back up Micron's statement. Micron's new GDDR7 boasts an impressive 60% bandwidth improvement over GDDR6, and it does so with 50% improved power efficiency. Response times have also been reduced by 20%, which will help sensitive workloads such as machine learning and AI image generation.

The new modules are rated at a whopping 32Gb/s, which is almost 10Gb/s higher than the fastest GDDR6X modules ever created. With a test setup comprised of 12 Micron GDDR7 ICs connected to a 384-bit wide interface, Micron says its next-gen memory can offer more than 1.5TB/s of bandwidth. By contrast, the RTX 4090 with the exact same memory configuration, but with 21Gbps Micron GDDR6X modules, offers "just" 1.008 TB/s of memory bandwidth.

Micron's performance estimations are impressive and a testament to the kinds of improvements we could see in future graphics cards. But they must be taken with a grain of salt. Even though GDDR7 has massive performance implications, GPU makers such as Nvidia are inevitably going to cut the bus width of next-gen cards by a significant amount, which will handicap the full 30% performance improvement Micron is advertising for its 32Gb/s GDDR7 memory. We could see an RTX 5090 with GDDR7 and the same 384-bit bus interface as the RTX 4090, but we don't expect Nvidia to use the same memory interface down the whole RTX 50 series stack, as has been the case with previous generations of GPUs. AMD and Intel will likely do the same once both come to the table with GDDR7-equipped graphics cards.

GDDR7 is the next generation of graphics memory, succeeding GDDR6 and its more performant sibling, GDDR6X. GDDR7 will start at 28GT/s or 32GT/s on release, but plans are already underway from Micron, Samsung, and SK Hynix that will see the production of higher speed modules — as high as 40 GT/s in the future.

Nvidia is all but certain to be using these next-gen GDDR7 modules and is probably responsible for driving the test setup that Micron used to create its 30% performance improvement metric, using some sort of prototype RTX 50 series GPU. According to rumors and leaks, Nvidia will be the first to produce graphics cards with GDDR7 modules. AMD is expected to keep using older GDDR6 modules for their next-generation GPUs, and probably won't transition until the RX 9000 series. Intel's next-generation GPUs are less clear, but Team Blue could go either way and utilize GDDR6 and/or GDDR7 for Battlemage, or wait to go to GDDR7 until its third-generation Celestial debut.

Aaron Klotz
Contributing Writer

Aaron Klotz is a contributing writer for Tom’s Hardware, covering news related to computer hardware such as CPUs, and graphics cards.

  • According to rumors and leaks, Nvidia will be the first to produce graphics cards with GDDR7 modules.

    And even then so, only the flagship RTX 5090/5080 cards might sport these modules, that too utilizing 28Gb/s speed as the minimum, let alone using 32Gb/s and higher spec'd variants.
  • BTW, speaking of GDDR7 memory and Nvidia's Blackwell GPU lineup.

    There was an update that the next-gen RTX 5090 GPU will reportedly sport 28GB GDDR7, on a 448-bit bus !

    So it will not use all the available 512-bit memory bus of the GB202 die, and only 14 memory modules would be used out of 16 (see below). So that should give 28GB VRAM, out of 32GB total. More than enough for any gaming GPU.

    They could reserve the remaining VRAM for later Ti variants, if need be, or use it on a "ProViz" RTX GPU.

    Assuming Nvidia goes for 28 Gbps GDDR7 memory modules initially, then we could see a total bandwidth of up to 1568 GB/s on the RTX 5090, which is still 56% more bandwidth than the RTX 4090 (384-bit @ 21 Gbps).


    Also, it is rumored that the RTX 5090 would feature a 3 PCB design, on the Founders Edition/FE model. There are also four memory modules on the top, five each on the sides, and two at the bottom.

    So we get a total of 16 GDDR7 DRAM modules featured on the RTX 5090 PCB. But not clear as to whether that refers to three individual PCBs within a single card, or three unique designs of which only one should be finalized.

    The memory layout is very dense.
    5 5

    FE uses three PCBs to leave space for double-sided blowing. The 30 and 40 dovetails are single-sided blowing, you know what I mean.
    Because the bit width is increased, the PCB cannot hold the memory lying down, and the staggered layout like PG651 is not used, so the full blowing method of PG137 is not inherited.
    Chiphell Forums (Machine Translated)
  • thestryker
    The thing I'm most curious about is latency as the implication is that they've improved the latency which means to me mostly that they didn't have to blow it up to get the extra bandwidth. It's nice to see the signaling shift so we don't end up with Micron going it alone with GDDR7X with PAM4 for high bandwidth.
  • Amdlova
    30% on a 4060ti puts it on pair with a 4070 super...

    Another bull.hit claims

    They need tell on the fatest avaliable graphics card with gddr7 can have a uplift of 30%.

    If your cpu can do it... right now? No one can do it
  • hotaru251
    Nvidia: yes, but just means we have to gimp the bus to reduce performance by 30% to compensate on the xx60 gpu's in future!
  • Meanwhile, see this MOD in action till we dream of getting GPUs sporting some of the highest rated memory speeds. Enter RTX 4090 SUPER ! 🤠

    Don't fall for this though, 'cause such gains can only be observed under specific circumstances and test environment, and the app/software in question should be extremely bandwidth-starved/hungry as well.

    26Gbps GDDR6X memory mod
    Rno97ZCKGGEView: https://www.youtube.com/watch?v=Rno97ZCKGGE