Samsung 990 Pro SSD Review: The Return of the King

The Samsung 990 Pro is the much-anticipated successor to the popular 980 Pro. It is poised to be the very fastest drive we’ve tested, outdoing even the SK hynix Platinum P41 and the Western Digital (WD) SN850X. Initially available at 1 and 2TB for $169 and $289, respectively, the 990 Pro will also have a 4TB model out next year. Available bare or with an RGB heatsink, the 990 Pro is using Samsung’s new Pascal controller and it's new V7 TLC NAND flash, a combination that promises to be faster and more efficient.

The 990 Pro has the rest of the Samsung trimmings, including software support with the updated Magician application and the option for encryption. The drive is designed to fit in a wide range of devices, from desktops to laptops to the PlayStation 5 (PS5) console. As Samsung’s new top-of-the-line consumer PCIe 4 SSD, the 990 Pro is designed to take a place among the best SSDs in terms of performance and, according to our tests, it delivers. It also takes the crown of the fastest SSD on our list of Best SSDs for the PS5. 

Samsung 990 Pro Specifications

The Samsung 990 Pro arrives in the 1TB and 2TB capacities at launch. The 4TB model is coming later, in 2023; it will be nice to see a larger capacity option from Samsung. The 990 Pro promises up to 7450/6900 MBps, sequential read and write, with up to 1.4 million / 1.55 million read and write IOPS. This is more than competitive, exceeding the previous-gen, 980 Pro on all counts but also promising numbers higher than the SK hynix Platinum P41 and WD SN850X.

The Samsung 990 Pro does support TCG Opal encryption - an unsupported feature for many consumer SSDs - and offers 600TB of warrantied writes per TB of capacity. The endurance rating follows the JEDEC JESD218 standard which is largely irrelevant. The TBW ratings are unexceptional in any case, but should be sufficient for the drive’s intended use.

Samsung is offering a variant at each capacity with a heatsink and RGB, taking a page out of the WD SN850X’s playbook. It appears that the 4TB model will also have this option which was not the case with the SN850X. The Platinum P41 is therefore perhaps a bit less attractive by not having a heatsink variant. The Samsung 990 Pro’s heatsink meets the PCI-SIG D8 standard, which means that it comes in at less than a 8.8mm height to ensure it fits into a wide variety of devices, including the PS5.

The Samsung 990 Pro launches with a price tag of $169.99 for 1TB and $289.99 for 2TB. The heatsink and RGB add another $20 to each price, which matches the premium of the SN850X. One complaint we had about the SN850X is that its launch price was too high - something that has since changed drastically. Samsung also had a hefty MSRP for its T7 Shield launch but the drive was never actually sold for launch prices. We therefore expect the 990 Pro to be discounted, which is reasonable given the current real-world pricing of its direct competitors.

Software and Accessories of the Samsung 990 Pro

Image 1 of 2

Samsung released version 7.2 of its Magician storage software on October 4th, in advance of the 990 Pro’s release. This software now allows for RGB control for the heatsinked versions of the drive. It also enables data migration and offers diagnostics, firmware updates and driver updates. Other functions are also present such as a PSID revert, which performs a secure erase and unlocks an encrypted drive. Samsung’s SSD software is, in general, the standard for the industry.

This software also has a Performance Optimization panel which allows for a Full Performance Mode. This is reminiscent of WD’s original Game Mode, which effectively disabled lower power states. In our tests, we benchmarked with the “Full Power Mode” both on and off.

A Closer Look at the Samsung 990 Pro

Image 1 of 3

The 990 Pro has unexceptional looks, with a label on both front and back providing basic drive information. The heat spreader back label helps reduce load temperatures. Under the top label there are two NAND packages, a DRAM package, and the controller. The controller is nickel-plated to improve heat dissipation, but the 990 Pro also uses Samsung’s Dynamic Thermal Guard (DTG) to improve thermal consistency. We’ve seen this before in the 980 Pro but this is also similar to the SN850X’s adaptive thermal management feature. This technology is nothing new, but may help with gaming workloads.

The Samsung 990 Pro sports a new controller, Pascal, as a successor to the 980 Pro’s Elpis. The Pascal is produced in Samsung’s 8nm process node and is still ARM-based. There does not appear to be much difference between the controllers superficially, but the Pascal has significantly higher performance specifications. Samsung states that this is thanks to an optimized NAND data path through “hardware automation technology” with reduced processing time via a “cache algorithm.” When I asked about the changes, Samsung stated that this architecture was meant to have more effective low-power mode coverage and that it was capable of taking advantage of read caching.

Hardware automation involves the storage data path: the flow of data from the host interface to the flash. As SSDs scale up to greater levels of parallelization with more complex addressing, IOPS bottlenecks begin to arise. Hardware automation helps overcome these and also offers power savings which is becoming more important as drives get faster. One example is hardware-accelerated flash map management which can increase IOPS versus managing the flash translation layer (FTL) - which translates between physical and logical memory addresses - in software. I/O queues and data transfers also benefit from acceleration. Separately, parallel processing improvements in firmware can also improve garbage collection/scheduling as with Phison’s I/O+ Firmware.

Volatile memory like SRAM is used by the controller to cache mapping information and buffer data to commit to the non-volatile NAND flash, so optimization of algorithms can improve overall performance. Write performance often benefits more from such advances, but changes may also be relevant for DirectStorage. Some manufacturers, like Solidigm, have also decided to introduce a form of caching to hold specific data in the cache via a specialized NVMe driver. In that case, knowledge about the type of data and how it is utilized - metadata - can bolster performance with intelligent caching. Samsung did not elaborate on what it meant by “read caching,” but it is not uncommon to keep some hot data in pSLC to improve forthcoming reads.

A typical cache algorithm would be least recently used (LRU), where the least recently accessed data would be evicted first from a full cache. Such an algorithm is constrained if it does not take into account spatial locality, that is, knowledge of adjacent and nearby memory locations. The inevitable increase in FTL overhead with more complex algorithms can create a performance bottleneck. Specific improvements to the controller can increase maximum IOPS by offloading or automating some of this work. This can also generally lead to lower latency, although the full benefits are unlikely to be realized on a consumer device.

Samsung has certainly made advances with the controller but other components are more familiar. The memory, or DRAM, is Samsung’s LPDDR4 which has been used on many of its SSD products. It offers a bit of power savings over traditional DDR. This module is 2GB in size, which matches the 2TB of flash with an ideal ratio. It’s expected that the 1TB model would have 1GB. The power components are also standard for Samsung. We would expect the promised power savings, in comparison to the 980 Pro, to come from the new flash plus controller optimizations.

The flash packages are labeled K9DVGY8JRD-DCK0 which denotes that it is Samsung’s 176-layer TLC NAND. With the 2TB SKU, this flash has 512Gb dies as indicated by “V” - although the upcoming 4TB may require 1Tb “X” dies. This is V7 in Samsung’s V-NAND nomenclature which was explored in some detail at ISSCC 2021. Samsung showed off its 8th generation TLC at ISSCC this year but recent trends in the flash market have generally left manufacturers in a holding pattern with regard to the production of higher-layer flash. In any case, this is the first time we’ve seen this flash, a direct upgrade to the V6 128-layer TLC found on the 980 Pro.

Samsung made many improvements with this generation of flash, the most significant being the use of a four-plane design and a Cell-on-Periphery (COP) implementation. More planes means more parallelization which translates at least to higher bandwidth. COP is similar to Micron’s CMOS-under-Array (CUA) but this is the first time Samsung is utilizing such technology. This involves moving control circuitry away from the peripheral or side of the array. This can greatly improve power efficiency and also reduces die surface area by placing peripheral circuitry under the data cell array.

Placing this circuitry under the array introduced some problems for Samsung who solved it by using innovative capacitor design. The result is improved power delivery with reduced surface area requirements. Samsung also uses a dual-scheme termination approach that allows for increased power efficiency when full I/O speeds are not required. This should help Samsung back up its claim of better power efficiency through the use of a low-power-aware architecture.

Samsung is the one manufacturer that was able to avoid string-stacking, that is the use of multiple NAND array decks, with higher flash layer counts. Its approach comes with many challenges but also avoids having to deal with melding decks. Etching so many layers leads to an increased aspect ratio which, among other things, can increase the level of voltage threshold deviation between cells in different layers. More layers in the same space does translate to an effectively smaller cell capacity which can impact performance and endurance, as well.

Samsung’s solution introduces an additional latch - a type of dynamic buffer, as with page buffers - to overcome this issue. The extra data is utilized with bit and word line forcing schemes to produce a refined cell charge, essentially having an extra verify stage to keep bit values tighter. While a manufacturer could simply keep adding decks - and it’s theoretically possible to hit 800 or more layers with that technique - it’s not without its own challenges. Samsung has instead decided that its 176-layer NAND provides the optimal balance between performance and power efficiency.

MORE: Samsung Promo Codes

MORE: Best SSDs

MORE: Best External SSDs and Hard Drives

MORE: How We Test HDDs And SSDs

MORE: All SSD Content