Question NVMe SSDs perform better on Intel platform

[igor_kavinski]

Storage Showdown - Intel Alder Lake vs. AMD Zen 4 | TweakTown



AMD needs to fix whatever is causing the slowdown. Otherwise, they will be at a disadvantage when the gaming world moves to DirectStorage.

 

[Billy Tallis]

Otherwise, they will be at a disadvantage when the gaming world moves to DirectStorage.

The primary purpose of DirectStorage is to reduce the CPU overhead of IO in Windows, which should have the effect of reducing or eliminating this kind of storage benchmark disparity.

It would be interesting to see this comparison done on Linux, which already has good options for low-overhead IO.

 

[Justinus]

Now that they finally have tested side by side, the difference is much smaller than I thought. Makes me wonder why my drives perform so much worse than theirs on my 7950X

 

[igor_kavinski]

Makes me wonder why my drives perform so much worse than theirs on my 7950X

Which benchmark did you run? Maybe trying to go for better mem clocks punishes the PCIe link on X670E?

 

[mikeymikec]

Storage Showdown - Intel Alder Lake vs. AMD Zen 4 | TweakTown

View attachment 68998

AMD needs to fix whatever is causing the slowdown. Otherwise, they will be at a disadvantage when the gaming world moves to DirectStorage.

IMO your conclusion is premature because the benchmarks don't correlate; you've got AMD winning some by like a gigabyte per second and others with Intel doing the same in return, and many won by a nose / close enough to be considered within a margin of error based on background system activity. At this point I'd assume that at least some of these benchmarks benefit a particular architecture, maybe because of the compiler that was used.

If there was a correlation between all these benchmarks, the next thing I'd do is to test both hardware platforms on Linux, and if there's a similar correlation there too, then maybe you can start talking about the hardware. Even then, I'd do additional tests on Windows, say testing Samsung SSDs with Samsung's NVMe driver and look for a similar correlation (I don't know if other SSD makers also have their own favoured drivers), because it could be an issue with the Windows NVMe driver, for example. I'd also test Win10.

I haven't checked the figures for this, but the only correlation I saw in them was that the SSD models seem to come out in the same order (when comparing the Intel figures to the AMD figures), which is a start, I guess.

Furthermore, in my experience it's typical for the ATTO benchmark to absolutely hammer the CPU at the low end of the transfer spectrum (e.g. 512B/sec). I've seen a fair bit of variation on different machines for the same SSDs over the years.

 

[Justinus]

Which benchmark did you run? Maybe trying to go for better mem clocks punishes the PCIe link on X670E?

I have had limited time with the system, but I am running 3100UCLK/6200 mem just like TweakTown is on their zen 4 review system. I only ran 3dmark SSD bench, scored ~4250 while they scored 4801.

They did explicitly say they opted to use a 7600X because there's some kind of additional latency for dual ccd chips for I/O. I need to update the BIOS (new version out) and maybe I can try with and without the slow CCD enabled.

 

[igor_kavinski]

I need to update the BIOS (new version out) and maybe I can try with and without the slow CCD enabled.

Great idea!

 

[Justinus]

Updated bios, updated chipset drivers. Tested my 980 Pro and SN850X again.

I tested both with and without CCD1 disabled. Results were identical in both cases. from 3DMark SSD Bench:

~3000 pts for 980 Pro (Tweaktown sees 3800)
~4175 points for SN850X (Tweaktown sees 4850)

No clue what could possibly be causing the performance to be so low on my system. The scores on the 7950X/X670E system are almost identical to the results I get on my 5950X system, which makes even less sense to me.

I am confused.

 

[mikeymikec]

No clue what could possibly be causing the performance to be so low on my system. The scores on the 7950X/X670E system are almost identical to the results I get on my 5950X system, which makes even less sense to me.

It seems reasonable to me. You've bought into a bleeding-edge, brand-new platform, it's not going to be as mature/optimized as the previous generation. Some manufacturer's optimisation routines are going to be better than others depending on what the issues are for a given generation. If you were using exactly the same hardware as the reviewer including BIOS updates, then a significant performance difference would be a concern.

 

[In2Photos]

Updated bios, updated chipset drivers. Tested my 980 Pro and SN850X again.

I tested both with and without CCD1 disabled. Results were identical in both cases. from 3DMark SSD Bench:

~3000 pts for 980 Pro (Tweaktown sees 3800)
~4175 points for SN850X (Tweaktown sees 4850)

No clue what could possibly be causing the performance to be so low on my system. The scores on the 7950X/X670E system are almost identical to the results I get on my 5950X system, which makes even less sense to me.

I am confused.

I may be grasping at straws here but which slots are you using during the test? Are the lanes tied to the CPU or chipset? Maybe it doesn't matter, just a thought.

 

[mikeymikec]

I may be grasping at straws here but which slots are you using during the test? Are the lanes tied to the CPU or chipset? Maybe it doesn't matter, just a thought.

and are those slots sharing bandwidth with another device.

 

[VirtualLarry]

Power management, did you disable it? All those tiny micro-sleep / wake on the NVMe interface can cause performance losses.

 

[Justinus]

It seems reasonable to me. You've bought into a bleeding-edge, brand-new platform, it's not going to be as mature/optimized as the previous generation. Some manufacturer's optimisation routines are going to be better than others depending on what the issues are for a given generation. If you were using exactly the same hardware as the reviewer including BIOS updates, then a significant performance difference would be a concern.

I am using the same MB, I think it's highly likely they were using the release bios version as the article came out before the next version was available. I tested both. I'm running the same ram speed with better timings. Only real difference is I have a 7950x instead of a 7600x.

I may be grasping at straws here but which slots are you using during the test? Are the lanes tied to the CPU or chipset? Maybe it doesn't matter, just a thought.

and are those slots sharing bandwidth with another device.

The board has two direct CPU M.2 slots. I have tried testing in both, both with and without a dGPU installed (that wasn't really intentional, and it made no difference since they are all direct CPU lanes.)

Power management, did you disable it? All those tiny micro-sleep / wake on the NVMe interface can cause performance losses.

I'll take a look at that.

 

[tcsenter]

AMD's integrated performance of I/O devices such as IDE/ATA, SATA, RAID, and USB have perennially been inferior to Intel all the way back to the South Bridge days.

 

[igor_kavinski]

Samsung 990 PRO 2TB SSD Review - A Higher Level | TweakTown

Only test where AMD excelled:



But overall, Intel wins:

 

[IntelUser2000]

Updated bios, updated chipset drivers. Tested my 980 Pro and SN850X again.

I tested both with and without CCD1 disabled. Results were identical in both cases. from 3DMark SSD Bench:

~3000 pts for 980 Pro (Tweaktown sees 3800)
~4175 points for SN850X (Tweaktown sees 4850)

No clue what could possibly be causing the performance to be so low on my system. The scores on the 7950X/X670E system are almost identical to the results I get on my 5950X system, which makes even less sense to me.

I am confused.

The way they test could be a factor. It doesn't look like it for the particular Tweaktown review, but review systems tend to be very "clean". They only have test applications installed and in some cases test an empty drive.

You'll be hard pressed to get the results they get for HDD and SSD(read: Non-Optane) systems and will always be to a degree, off. That's because SSDs rely on buffers and algorithms to make up for the weaknesses(the NAND flash itself) and performance varies vastly based on what you are doing and with what file size and how frequently, etc.

 

[igor_kavinski]

@Justinus

If you are using Windows 11, can you check if VBS is turned on? It can slow down processes considerably.

 

[igor_kavinski]

SSTC PCIe 5.0 SSD hits 10GB/s speeds on Intel and AMD platforms - VideoCardz.com

SSTC preps Tiger Shark SSD with PCIE Gen5 support A storage company from Vietnam is set to launch its first PCIe Gen5 storage very soon. The first samples of the new Tiger Shark SSD have just been sent out for reviews. This is next-gen storage based on PCIe Gen5 interface and NVMe 2.0 protocol...

videocardz.com

At least one PCIe 5.0 SSD also working faster on Intel platform.

Hmmm...AMD needs to look into this seriously.

 

[igor_kavinski]

Phison E26 Max14um ES 2TB SSD World Exclusive Preview - Fastest in history

Join us for this exclusive of the fastest flash-based consumer SSD in history, delivering 14,600 MB/s, and cooled by Frore Systems AirJet and passively.

www.tweaktown.com

9% performance loss with AMD chipset.

 

[Justinus]

View attachment 85171

Phison E26 Max14um ES 2TB SSD World Exclusive Preview - Fastest in history

Join us for this exclusive of the fastest flash-based consumer SSD in history, delivering 14,600 MB/s, and cooled by Frore Systems AirJet and passively.

www.tweaktown.com

9% performance loss with AMD chipset.

It's not really the AMD Chipset at fault here, AM5 thus far doesn't offer PCI-e 5 on the chipset. It's down to the PCI-e 5.0 implementation in the IOD on the CPU.

 

[Shmee]

I haven't kept up on the PCIe 5 SSDs as I don't have a PCIe 5 system yet, but is heat and therefore giant cooling solutions still a big issue? As far as platform support, is it possible that new revisions of motherboards and say, Zen 5 CPUs in existing motherboards, could resolve bottlenecks like this?

It seems like all this PCIe 5 tech is still pretty new anyway. I personally would be more interested once more refined options come out for SSDs that are easier to cool, and chipsets/IOD controllers that have more lanes and bandwidth available.

 

[igor_kavinski]

It's down to the PCI-e 5.0 implementation in the IOD on the CPU.

Good to know. It means that Zen 5 could fix this issue with a slightly revised IOD.

 

[ikjadoon]

It's not really the AMD Chipset at fault here, AM5 thus far doesn't offer PCI-e 5 on the chipset. It's down to the PCI-e 5.0 implementation in the IOD on the CPU.

Even weirder are these PCIe 4.0 tests: here is 990 PRO 2 GB with a heatsink. AMD & Intel are neck and neck in almost every test, except:

12.3% lead for Intel

12.1% lead for AMD

Weird, weird, weird. I wish TweakTown would do an exploratory, rigorous comparison on why either platform is preferred on either test, besides these lines in the OP article:

Sequential throughput numbers have been solidly in AMD's favor since Zen 3 first landed.

...

Intel has a higher IPC even with factoring in our AMD chip's 100 MB/s higher clock speed, so it comes as no surprise that Alder Lake wins here.

...


AMD really destroys Intel here simply because this is one chunk of data being served to the host, meaning it's sequential in nature and right up AMD's alley.

Don't ask why it says "100 MB/s higher clock speed", haha.

 

[Shmee]

Yeah, their summary of this doesn't seem to make a lot of sense.