Samsung debuted its new Z-NAND at the Flash Memory Summit this week in an obvious ploy to undercut IMTF 3D XPoint before it comes to market, but the company also has another initiative to attack in the memory space.
On the surface the new 1 TB Z-SSD, which is built with Samsung's new Z-NAND, is very similar to other existing PCIe SSDs. Samsung claims that this new specialized version of NAND provides a radical performance increase compared to existing NAND-based SSDs.
Samsung Attacks 3D XPoint Storage Applications
Samsung presented slides that listed the Z-SSD's performance specifications, but a note in the bottom right-hand corner of the slides indicates that it derived the benchmarks from an emulator. This may indicate that the company doesn't actually have physical products yet and is relying on simulations to derive performance specifications. The company compared the SSD to both NAND-based and PRAM-based SSDs. Phase-Change RAM (PRAM) is a glass-type of non-volatile chalcogenide memory that many speculate Intel/Micron uses with 3D XPoint. It appears that Samsung is implying that it can offer similar performance and latency to 3D XPoint.
In either case, the slides indicate much higher throughput than a normal SSD (1.6x better than a Samsung PM963 NVMe SSD) and radically reduced latency that is on par with the PRAM-based SSD. The fact that Samsung has a PRAM-based SSD for comparison, however, is telling. The company, like many other fabs, constantly has new technologies under development, and it's possible that we may see a PRAM-based SSD in the future.
Samsung representatives informed us that the new NAND would have SLC-like endurance specifications, which implies that the new Z-NAND is simply using the entire capacity of an MLC NAND product programmed as SLC. The company also notified us that it based the new technology upon its existing 3D NAND technology, so it is likely using a 64-layer architecture. One of the key differentiators, beyond performance, is that the Z-SSD promises incredible scaling and maximum IOPS under low QD workloads. This is similar to the 3D XPoint performance profile, and like its competitor, this is likely due to an increased number of planes.
Normal NAND-based SSDs require an excessively heavy workload--rare in normal use--to extract the maximum performance. This is because the NAND die need to respond en masse to incoming requests. NAND vendors boost the performance of individual die by splitting the die into separate regions that can respond to commands separately (increase parallelism). The majority of current generation NAND are dual-plane designs, but Micron 3D NAND employs a quad-plane design to increase performance.
3D XPoint doesn't have the traditional plane architecture, but the company disclosed that there are 64 regions that can respond independently, which is similar to a 64-plane architecture. Samsung may have boosted the number of planes per Z-NAND die to achieve a similar level of performance and parallelism. Coupling the extra die with SLC, which is much faster and endurant than MLC and TLC NAND, would provide tremendous performance benefits.
Samsung provided results, again generated with an emulator, that show higher performance (up to 1.2X in a big data analytics workload) and much lower power consumption than the PRAM-based SSD. The NAND-based SSD trails far behind the PRAM and Z-NAND solutions, but Samsung did not provide a direct power comparison between NAND and Z-NAND. The company also did not provide any random performance specifications during the announcement, but did note that the new SSD will use the NVMe protocol and will be faster than existing NVMe SSDs. Most NVMe SSDs top out in the range of 750,000 IOPS with a x4 connection and 1 million IOPS with a x8 connection. Samsung also divulged to us that the Z-SSD would offer up to a 7x improvement in 99.99th percentile latency (worst-case) over existing NAND-based SSDs.
Samsung representatives noted that because it based Z-NAND on existing NAND technology it will be much cheaper to produce, while its 3D XPoint competitors are investing billions in new fabs and tooling. The impact of the increased investments have already shown up in Intel's financial reports, and the extra cost will push the price of 3D XPoint up beyond NAND-based SSDs.
Samsung Attacks 3D XPoint as Memory
Samsung is also squeezing 3D XPoint, which can be used as either storage or memory, on the RAM axis. Samsung's partner Netlist also announced its new HybriDIMM products, which use HyperVault technology.
Samsung invested a total of $23 million in Netlist last year to develop the HyperVault technology. The result of the investment is the new HybriDIMM products, which connect into a DDR4 LRDIMM slot and allow the system to address large repositories of NAND (up to 512 GB with the first-gen product and 1 TB with the second-gen product) as either storage or memory.
The DIMMs work without a BIOS change, which will be unlikely with 3D XPoint DIMMs, and use PreSight predictive algorithms to pre-fetch and stage data from NAND into a small amount of DRAM (8-16 GB) on board the DIMM, thus alleviating performance limitations.
These new DIMMs will allow users to place terabytes of RAM into a single server in a much more cost-effective manner than current DRAM and 3D XPoint. We have deeper coverage of the technology here. HybriDIMMs use Samsung NAND, but it would be logical to assume that we might also see Z-NAND on them in the future.
Samsung's line of attack with the Z-SSD might help price 3D XPoint out of the market for storage applications. Combining the storage assault with lower-cost HybriDIMM memory alternatives will also provide additional pressure. The kicker is that Samsung based all of the new products on proven and mature technologies, which is important for the lucrative data center market that Intel and Micron are chasing.
Samsung's strategy capitalizes on the simple fact that in most cases good-enough and cheap will win, but we will have to wait until products hit the market to see if it pans out.