Spectre Haunts SSDs; QLC & 3D NAND Aren’t Afraid
ONE OF THE BIGGEST STORIES of 2018 so far has been the Meltdown and Spectre exploits, and the fallout keeps coming. Recently, while testing a new batch of NVMe SSDs, some oddities started popping up. In short, performance was far worse than expected.
The poor results weren’t just with the new drives: All the fastest NVMe SSDs had lost performance. Sequential read/write speeds were still high, but random IO had dropped, and more critically, it took a significant hit for low queue depths.
That last point is important, because while SSD manufacturers love to show maximum random IO performance at queue depth (QD) of 32 or even 64, most real-world work happens at QD1, and over 99 percent of reads and writes happen below QD4. The lower performance first came up in early May, right after Microsoft rolled out the April 2018 update to Windows 10. Digging a bit more, it looks like a combination of Spectre firmware patches for Intel platforms and the latest Win 10 update are to blame.
Using a Z270 system without Spectre firmware, running Windows 10 Creators Update (version 1709), initial QD1 read/write results are 75.9/236.3MB/s. Updating the firmware to incorporate Intel’s Spectre fixes dropped performance to 70.1/181.2MB/s. Applying the April 2018 update (version 1803) to Windows 10 resulted in a further drop to 59.9/177.6MB/s. That’s a 21 percent drop in read speeds, and a 25 percent drop in write speeds. The kicker is that similar performance (using Samsung’s 970 Evo 500GB as the guinea pig) showed up on Z270, Z370, and X299 Intel platforms with Spectre firmware patches and Win 10 April 2018. Switching to an AMD X470 platform delivered 72.8/206.2MB/s; still a bit slower, but better than any Intel system.
To be fair, most real-world workloads don’t run purely random IO, but the reduced performance shows that earlier statements suggesting the impact to system performance would mostly be seen in data-center workloads aren’t accurate; even home PCs can be affected. These results also show that we’re not yet clear of the reach of Spectre and similar exploits, with new variants being announced on a regular basis. In other words, these aren’t the sort of patches where disabling them to regain lost performance is a viable option. There is also some hope that affected software can be changed to work better within the post-Spectre ecosystem, and future updates to
both firmware and Windows 10 may alleviate some of the pains.
It’s not all doom and gloom, though. After over a year of climbing SSD prices, things are returning to 2016 levels, and sub-$100 500GB drives are now common. We’re not likely to stop there either, as many manufacturers are pursuing QLC NAND. That increases the complexity of reading and writing data to NAND, with four bits per cell (compared to TLC’s three bits, and MLC’s two bits), but improvements in 3D NAND and controllers make the approach viable, especially for read-heavy workloads.
Intel, Micron, Samsung, and Toshiba are all talking about QLC NAND, and while the names vary, the idea is the same: 33 percent more data in the same space compared to TLC, or twice as much data as MLC. One of the driving forces of QLC NAND adoption in SSDs appears to be 3D NAND, and specifically the move to 96-layer designs. Toshiba’s BiCS4 QLC will allow for up to 1.33Tb per chip (and multiple chips in a package), meaning we could see multi-terabyte SSDs with only two NAND packages.
More importantly, Samsung is talking about terabyte-class SSDs becoming mainstream by 2020. We’re already seeing 1TB SSDs in the $200 range, which means in the next couple of years, we could see prices drop by half. The days of the spinning disk may finally be coming to an end, and 20TB and larger SSD capacities are in the works.
With up to 1.33Tb per chip, we could see multi-terabyte SSDs with only two NAND packages.
Jarred Walton has been a PC and gaming enthusiast for over 30 years.