PCI-Express 4.0 NVMe SSD Performance on Ryzen 3000 & X570 60

PCI-Express 4.0 NVMe SSD Performance on Ryzen 3000 & X570

(60 User Comments) »

Conclusion

We were really curious to see what performance gains can be enjoyed when using a blazing-fast PCI-Express 4.0 NVMe SSD, so we ran extensive tests using actual real-life applications, not just synthetic benchmarks that don't take into account how applications work today. While the synthetic numbers are truly impressive for PCI-Express 4.0, showing over 5 GB/s actual transfer rates, which is 43% faster than the same SSD running at PCI-Express 3.0, the actual real-life gains are small. Averaged across our test suite, we only see a 1% improvement, nothing you'd ever notice in real-life.

It seems the problem is not with sequential transfer rates, which are high enough, but what matters in real applications—random IO. As the synthetic testing on page 1 shows, 4K random is essentially identical between all the PCIe modes, at least on this Phison-based SSD. There is only a single test in our test suite that shows significant gains, and that's the "ISO file copy test", which copies a 4 GB file. If you do copy large files all day then this is definitely a use case for PCIe 4.0 SSDs. I do wonder where your data is coming from or going to, though. In order to get those speeds, both your source and destination must be able to sustain those rates, which rules out networking, SATA drives, USB, Thunderbolt, and everything else I can think of. If you copy between NVMe SSDs, you had better shell out the big bucks for MLC drives because TLC won't be able to sustain such rates. Due to its design, TLC is very slow during writes, and SSD manufacturers work around that by running part of the flash in pseudo-SLC mode, which is faster, but limited in capacity because it triples the cell usage compared to TLC. Another bottleneck could be SSD temperatures, which no doubt will go up with Gen 4 because the controller has to process more data in a shorter timeframe, ultimately resulting in throttling because its temperature will go too high. Gigabyte includes a very beefy heatsink with their SSD, which works great and looks beautiful at the same time.

Our sequential testing data shows that the Phison E16 doesn't make full use of the new interface's capabilities, even for sequential reads. A doubling of the interface bandwidth should ideally result in a doubling of the sequential data transfer rate over PCIe 3.0, so 7 GB/s. While 5 GB/s is definitely impressive, it's not double, which is what some people hint at when they talk about "PCIe 4.0 doubling the speeds". I'm sure we will eventually see much faster SSD controllers that push transfer rates even higher and improve IOPS rates at the same time—that's when actual performance could go up.

The future for PCIe 4.0 looks bright, and we're sure one of the bigger manufacturers, like Samsung, will release a drive that better utilizes this bandwidth. We're also optimistic for low-cost drives that use PCIe 4.0 x2 controllers, which lowers cost of the controller, yet performance should be comparable to today's PCIe 3.0 x4 drives. Unfortunately, the remaining two PCIe lanes will sit unused as they can't just magically switch to other slots or devices. While it is possible in theory for motherboard manufacturers to route only two PCIe lanes to an M.2 slot and use the remaining lanes elsewhere, that approach seems unrealistic considering the vast majority of the market will demand "x4" and will stay away from boards offering "just x2". The more realistic option would be PCIe switches, like those used in today's x16/x8x8 graphics slot configurations, but these switches are not cheap and will drive up motherboard cost.

Looking at our "Intel" data point, which uses the same PCIe Gen 3 SSD on both Intel and Zen 2, shows that there's definitive performance gains to be had when switching from a slower CPU to a faster CPU, even without touching storage. While not an exact apples-to-apples comparison, it is still solid evidence that spending money wisely will pay off as the money saved on the SSD could let you buy a faster processor, more RAM, or a more powerful graphics card.

Current PCI-E 4.0 SSD pricing is high; the tested 2 TB Gigabyte drive is listed online for $460, and the 1 TB version costs $260. In this article, we included performance results from the PCIe 3.0-based ADATA SX8200 Pro 1 TB, which just costs $145—almost half that of the Gen 4 SSD, and it's just as fast as the latter, mostly due to a better controller that delivers higher random IO performance. Personally, I could see myself spending maybe 10% or $10-$20 more for a Gen 4 SSD, but everything beyond that is just the early adopters tax. On top of that, you have to consider the higher cost of Gen 4 capable motherboards—only AMD X570 supports PCIe Gen 4 at this time.

Do let us know in the comments for this article if you actually have a use case where such drives can result in substantial performance improvements.
Discuss(60 User Comments)