Discussion Reduced SATA performance on X570?

[tamz_msc]

https://www.reddit.com/r/Amd/comments/fwh7q0

This might be worthy of further investigation.

 

[VirtualLarry]

If you think about it, really, it's not all THAT surprising. AMD themselves developed (supposedly?) the X570 chipset. Previous B450 chipset was from ASMedia.

One thing though - have they compared the 4K random scores, between the SoC (on-CPU-die) SATA ports, and the "Chipset" SATA ports, between the B450 and X570? That might be interesting, to see if the SoC SATA ports actually have different performance, between B450 and X570, using the SAME CPU (model). If true, that COULD potentially point to a BIOS bug, or other bug, somewhere.

Traditionally, AMD's own chipsets, have had worse AHCI performance than Intel, although ASMedia SATA port add-on cards were often worse still.

Seems possible that this is a real issue, but where are the big headlines about Intel's consumer chipsets and PCI-E 3.0 x4 NVMe slowdown, due to the fact that Intel platforms DON'T have a dedicated PCI-E 3.0/4.0 x4 lane bundle for the NVMe slot, like AMD's AM4 platform does.

Thus, Intel consumer platforms have to share the DMI bandwidth (roughly PCI-E 3.0 x4 itself) between the chipset-attached NVMe slot, and the REST of the I/O, including SATA ports, and chipset PCI-E x1 slots. (And network and sound onboard PCI-E x1 devices, as well.)

IMHO, the NVMe "slowdown" on Intel platforms, is a WORSE issue, than AMD's SATA performance. Who uses SATA for anything but large spinners these days anyways, with most higher-end X570 boards having TWO PCI-E 4.0 x4 NVMe slots for primary storage.

I would rather have full-speed dual NVMe 4.0 slots, and slightly-gimped SATA ports, than full-speed SATA ports, and really gimped NVMe slots.

 

[tamz_msc]

One thing though - have they compared the 4K random scores, between the SoC (on-CPU-die) SATA ports, and the "Chipset" SATA ports, between the B450 and X570?

These tests were all done on the chipset SATA ports, according to the OP.

IMHO, the NVMe "slowdown" on Intel platforms, is a WORSE issue, than AMD's SATA performance. Who uses SATA for anything but large spinners these days anyways, with most higher-end X570 boards having TWO PCI-E 4.0 x4 NVMe slots for primary storage.

That's a bit disingenuous; SATA SSDs are more commonplace than NVMe, and Intel's limitation is a known one due but this "issue" on X570 if it is indeed a limitation has no explanation whatsoever.

 

[VirtualLarry]

That's a bit disingenuous; SATA SSDs are more commonplace than NVMe, and Intel's limitation is a known one due but this "issue" on X570 if it is indeed a limitation has no explanation whatsoever.

Not really. They're both chipset limitations (if the AMD chipset SATA issue is true, which I suspect that it probably is). Just that the Intel ones has been known for some time, whereas, the AMD SATA limitation may be kind of out of the blue here.

But knowing AMD's "history", with implementing their own AHCI ports, I wonder if it's simply that, an AMD implementation that was in some way lacking compared to Intel. (Didn't Intel "invent" AHCI?) I would personally find that explanation "enough".

But they really need to test the SoC SATA ports too, on both B450 and X570 boards, and compare all around. (*)

(*) Having SATA ports on a board, with varying SATA performance profiles / capabilities, is nothing new for Intel owners, they've had to deal with it for years... thinking of boards with Intel SATA6G ports and SATAII ports on the same board, additional chipset Marvell or ASMedia ports, etc. It has only been with Skylake and newer Intel chipset / platforms, that their SATA performance, across all ports, was basically top-notch. And again, still, it has the limitation that it is shared over DMI bus, which means that a heavily-loaded system with an NVMe SSD is going to perform worse, compared to AMD, even with AMD's supposedly-inferior chipset SATA ports, simply because AMD's AM4 platform has dedicated x4 PCI-E lanes for the primary NVMe slot.

 

[VirtualLarry]

What CPU was this tested with? The 3900X is the lowest CPU that has full write bandwidth, the 3600, 3600X, 3700X, 3800X have half write bandwidth, due to the single chiplet.

Maybe that's the only explanation, and why this hasn't been noticed before?

Edit: This was already a known aspect of AMD's Zen2 on AM4, although a bit technically-obscure, and generally only relevant to benchmarkers and the HPC / DC / mining crowd.

 

[tamz_msc]

What CPU was this tested with? The 3900X is the lowest CPU that has full write bandwidth, the 3600, 3600X, 3700X, 3800X have half write bandwidth, due to the single chiplet.

That's memory write bandwidth. What does it have to do with SATA performance?

 

[VirtualLarry]

That's memory write bandwidth. What does it have to do with SATA performance?

Because AHCI ports DMA data to / from main memory, AFAIK. (Edit: I do believe that chipset memory writes, since the memory-controller is on the CPU, are affected by this issue likewise.)

I'm not saying that this IS the reason for the "AMD X570 chipset SATA performance anomaly", but only that it COULD be just a side-effect of having half the write bandwidth with a lesser CPU.

This can of course be tested, if someone has a 3600 and a 3900X CPU, and an X570 board or two.

Edit: I noticed that no-where in that Reddit "article", does the author state WHICH AMD CPU he was using, even though he mentions owning "several". Couldn't be a hit-piece, could it? Nah...

Edit: I also find it really curious, how he didn't check the SoC SATA ports for some reason. Is the author un-aware that not all SATA ports are provided by the chipset, on some boards, they fan-out the SoC SATA ports too, as some of the onboard SATA ports? My B450-F ROG STRIX does that, I believe. There are six chipset SATA and two SoC SATA, I think. Or maybe 4 and 2.

Edit: The more I think about it, I think my theory about the half write bandwidth is wrong. Because, wouldn't that still be the case when a 3900X is paired up with a supported B450 mobo too? Hmm. More to think about. (I guess the only way that would make sense, would be if he had the 3900X in a B450 board, and a 3600/3700X in a X570 board, to get the results that he is, if my theory was correct.)

 

[SPBHM]

how does it compare to previous AMD chipsets?
the southbridge on AM3+ (SB900?) or the FM2 chipset

 

[Hi-Fi Man]

I was curious, so I did I comparison with my 4TB 860 evo connected to X570 SATA and my LSI 9212i (SAS2 2008 Falcon) HBA with IT firmware. The HBA does not have DRAM cache onboard and with IT firmware acts only as an HBA (no RAID). One thing I noticed with the HBA is that CPU usage was MUCH lower than with the X570 SATA. During the test one CPU core was nearly maxed out on the X570 SATA for most of the run whereas the HBA produced single digit CPU load for half of the test and when things did ramp up the load was distributed across two cores with one averaging around 40% load and the other 15%. All results were gathered with the MS AHCI driver for the X570 SATA and driver version 2.0.79.82 for the HBA. The HBA is equipped with a 700MHz(?) PowerPC CPU so that might explain the lower CPU load.

Results with LSI 9212i HBA:

-----------------------------------------------------------------------
CrystalDiskMark 6.0.2 Shizuku Edition x64 (C) 2007-2018 hiyohiyo
Crystal Dew World : https://crystalmark.info/
-----------------------------------------------------------------------
* MB/s = 1,000,000 bytes/s [SATA/600 = 600,000,000 bytes/s]
* KB = 1000 bytes, KiB = 1024 bytes

Sequential Read (Q= 32,T= 1) : 557.537 MB/s
Sequential Write (Q= 32,T= 1) : 514.883 MB/s
Random Read 4KiB (Q= 8,T= 8) : 298.450 MB/s [ 72863.8 IOPS]
Random Write 4KiB (Q= 8,T= 8) : 230.655 MB/s [ 56312.3 IOPS]
Random Read 4KiB (Q= 32,T= 1) : 298.087 MB/s [ 72775.1 IOPS]
Random Write 4KiB (Q= 32,T= 1) : 229.075 MB/s [ 55926.5 IOPS]
Random Read 4KiB (Q= 1,T= 1) : 40.911 MB/s [ 9988.0 IOPS]
Random Write 4KiB (Q= 1,T= 1) : 93.923 MB/s [ 22930.4 IOPS]

Test : 1024 MiB [D: 52.1% (1941.3/3726.0 GiB)] (x5) [Interval=5 sec]
Date : 2020/04/09 10:59:53
OS : Windows 10 Professional [10.0 Build 18363] (x64)


Results with onboard X570:

-----------------------------------------------------------------------
CrystalDiskMark 6.0.2 Shizuku Edition x64 (C) 2007-2018 hiyohiyo
Crystal Dew World : https://crystalmark.info/
-----------------------------------------------------------------------
* MB/s = 1,000,000 bytes/s [SATA/600 = 600,000,000 bytes/s]
* KB = 1000 bytes, KiB = 1024 bytes

Sequential Read (Q= 32,T= 1) : 548.056 MB/s
Sequential Write (Q= 32,T= 1) : 486.584 MB/s
Random Read 4KiB (Q= 8,T= 8) : 233.260 MB/s [ 56948.2 IOPS]
Random Write 4KiB (Q= 8,T= 8) : 197.195 MB/s [ 48143.3 IOPS]
Random Read 4KiB (Q= 32,T= 1) : 233.024 MB/s [ 56890.6 IOPS]
Random Write 4KiB (Q= 32,T= 1) : 197.233 MB/s [ 48152.6 IOPS]
Random Read 4KiB (Q= 1,T= 1) : 43.056 MB/s [ 10511.7 IOPS]
Random Write 4KiB (Q= 1,T= 1) : 104.322 MB/s [ 25469.2 IOPS]

Test : 1024 MiB [D: 52.1% (1941.3/3726.0 GiB)] (x5) [Interval=5 sec]
Date : 2020/04/09 10:44:12
OS : Windows 10 Professional [10.0 Build 18363] (x64)

 

[tamz_msc]

I was curious, so I did I comparison with my 4TB 860 evo connected to X570 SATA and my LSI 9212i (SAS2 2008 Falcon) HBA with IT firmware. The HBA does not have DRAM cache onboard and with IT firmware acts only as an HBA (no RAID). One thing I noticed with the HBA is that CPU usage was MUCH lower than with the X570 SATA. During the test one CPU core was nearly maxed out on the X570 SATA for most of the run whereas the HBA produced single digit CPU load for half of the test and when things did ramp up the load was distributed across two cores with one averaging around 40% load and the other 15%. All results were gathered with the MS AHCI driver for the X570 SATA and driver version 2.0.79.82 for the HBA. The HBA is equipped with a 700MHz(?) PowerPC CPU so that might explain the lower CPU load.

Results with LSI 9212i HBA:

-----------------------------------------------------------------------
CrystalDiskMark 6.0.2 Shizuku Edition x64 (C) 2007-2018 hiyohiyo
Crystal Dew World : https://crystalmark.info/
-----------------------------------------------------------------------
* MB/s = 1,000,000 bytes/s [SATA/600 = 600,000,000 bytes/s]
* KB = 1000 bytes, KiB = 1024 bytes

Sequential Read (Q= 32,T= 1) : 557.537 MB/s
Sequential Write (Q= 32,T= 1) : 514.883 MB/s
Random Read 4KiB (Q= 8,T= 8) : 298.450 MB/s [ 72863.8 IOPS]
Random Write 4KiB (Q= 8,T= 8) : 230.655 MB/s [ 56312.3 IOPS]
Random Read 4KiB (Q= 32,T= 1) : 298.087 MB/s [ 72775.1 IOPS]
Random Write 4KiB (Q= 32,T= 1) : 229.075 MB/s [ 55926.5 IOPS]
Random Read 4KiB (Q= 1,T= 1) : 40.911 MB/s [ 9988.0 IOPS]
Random Write 4KiB (Q= 1,T= 1) : 93.923 MB/s [ 22930.4 IOPS]

Test : 1024 MiB [D: 52.1% (1941.3/3726.0 GiB)] (x5) [Interval=5 sec]
Date : 2020/04/09 10:59:53
OS : Windows 10 Professional [10.0 Build 18363] (x64)


Results with onboard X570:

-----------------------------------------------------------------------
CrystalDiskMark 6.0.2 Shizuku Edition x64 (C) 2007-2018 hiyohiyo
Crystal Dew World : https://crystalmark.info/
-----------------------------------------------------------------------
* MB/s = 1,000,000 bytes/s [SATA/600 = 600,000,000 bytes/s]
* KB = 1000 bytes, KiB = 1024 bytes

Sequential Read (Q= 32,T= 1) : 548.056 MB/s
Sequential Write (Q= 32,T= 1) : 486.584 MB/s
Random Read 4KiB (Q= 8,T= 8) : 233.260 MB/s [ 56948.2 IOPS]
Random Write 4KiB (Q= 8,T= 8) : 197.195 MB/s [ 48143.3 IOPS]
Random Read 4KiB (Q= 32,T= 1) : 233.024 MB/s [ 56890.6 IOPS]
Random Write 4KiB (Q= 32,T= 1) : 197.233 MB/s [ 48152.6 IOPS]
Random Read 4KiB (Q= 1,T= 1) : 43.056 MB/s [ 10511.7 IOPS]
Random Write 4KiB (Q= 1,T= 1) : 104.322 MB/s [ 25469.2 IOPS]

Test : 1024 MiB [D: 52.1% (1941.3/3726.0 GiB)] (x5) [Interval=5 sec]
Date : 2020/04/09 10:44:12
OS : Windows 10 Professional [10.0 Build 18363] (x64)

Random 4K at higher QD seems to be lower.

 

[VirtualLarry]

Is there a difference in terms of total number of outstanding NCQ tags/requests that can be in-flight at once, in Intel's, ASMedia's (B450), and AMD's (X570) AHCI implementation? That could possibly affect random 4K I/O bandwidth.

 

[13Gigatons]

With the prices of M.2 falling I think AMD got it right. Sata has finished it's development with no new speed increases.

Just need larger M.2 drives to be priced right. How about 4TB for $200?

 

[DrMrLordX]

With the prices of M.2 falling I think AMD got it right. Sata has finished it's development with no new speed increases.

Most x570 users still on SATA will use the SoC SATA ports instead of the chipset ports. You only get a few from the chipset anyway. You'd have to be using a lot of SATA devices for the chipset to even matter. I only have one SATA drive just for the hell of it. Otherwise I've moved over to NVMe.

 

[Fanatical Meat]

With the prices of M.2 falling I think AMD got it right. Sata has finished it's development with no new speed increases.

Just need larger M.2 drives to be priced right. How about 4TB for $200?

Yeah, seems like drives are fast enough now. Probably explains why nobody is talking about x570 SATA drives...

 

[undertaker101]

Is this still an issue on x570? Gimme SATA speeds? I have a lot of SATA drives and only one nvme and will be pissed if performance was half on the new platform. Good thing I caught this before I assembled my x570+5800x system.

 

[Leeea]

I do not believe this issue exists in any significant way.

Need something more then a random reddit post. I have seen to many deliberate falsifications on reddit to believe it.

 

[tamz_msc]

Is this still an issue on x570? Gimme SATA speeds? I have a lot of SATA drives and only one nvme and will be pissed if performance was half on the new platform. Good thing I caught this before I assembled my x570+5800x system.

Performance isn't halved, but using the chipset SATA ports certainly results is somewhat reduced performance compared to Intel chipset ports.

 

[undertaker101]

How bad is the drop? 5%, 10% or greater...disappointed since x570 is the pro solution and there still isn't a fix. With all that PCI-E 4.0 bandwidth getting sata ports to work at full hilt should not be an issue.

 

[Fanatical Meat]

How bad is the drop? 5%, 10% or greater...disappointed since x570 is the pro solution and there still isn't a fix. With all that PCI-E 4.0 bandwidth getting sata ports to work at full hilt should not be an issue.

I think it isn’t a noticeable difference either way assuming you are not doing some sort of odd thing with your machine.
To my understanding there is no noticeable difference for home use.

 

[undertaker101]

So far, at least on my x570 TUF PRO, the performance of the native SATA controllers is as expected: 500 mbps+ R/W for the ssds.

 

[parhamsan]

Hi Everyone,

Just wanted to share my SATA issues with x570 MB with everyone here.
I preordered the AMD Ryzen 5900x form a retailer (Canada Computers in Canada) and while waiting for my CPU, decided to give the GIGABYTE X570 AORUS ULTRA and the 3900x a try before my 5900x would arrive.
I updated the MB to the latest BIOS (F30 at that time) and installed a fresh copy of the Windows 10.

To give you more info on my setup here are my PC specs:

CPU: AMD Ryzen 3900x
MB: X570 AORUS ULTRA Latest F30 BIOS
GPU: EVGA RTX 3080 FTW3 Ultra (running at x8 Gen4 because of the 10gb card)
M.2 (First slot): Samsung 970 PRO 512 NVME (OS)
M.2 (second Slot): Sabrent SB-ROCKET-2TB
SSD: Crucial MX500 2TB (Connected to SATA 0 on MB)
10Gige Card: Mellanox ConnectX-2 Ethernet Adaptor (SingleSFP+)

Now as a comparison, I had the exact same hardware on a Gigabyte Z390 MB with an Intel 9900k so basically the X570 was going to replace the Intel and the Z390 Chipset.

After the fresh installation of the latest windows 10 (20H2) and all the AMD drivers from the Gigabyte website I started my daily routine jobs on my PC.
I do tons of video encoding so most of the time my CPU load is at 80% and above. I use handbrake and Staxrip for my video encodings.

After some time when I was transferring files from the Sabrent to the MX500 I realized that the speed of copying files is around 130MB/s. Now from my previous system I knew that the MX500 is capable of speeds of 560MB/s Read and 510 MB/S write so this was surprising.

I instantly stopped all the my work and tested the transfer speed with no load on the CPU and the results were the same (using Crystaldiskmark 8 x64).
I even relized that speed goes down on my M.2 Sabrent drive with no cpu load.

After some troubleshooting I realized that after the computer boots, my SSD transfer speed was capped at 520/430 MB/s (compared to the 560/490 MB/s that I had my on Intel Z390 system) but as soon as the CPU went under some load, the speed would go down to around 200MB/s (write) so I did more troubleshooting.

Speed SSD MX500 on Intel Z390 system (CPU at 90% load):


Speed SSD MX500 on X570 system (CPU no load after boot):


Speed SSD MX500 on X570 system (CPU at some minor load, or after 5min of idle system load):


Speed Sabrent on Z390 system (CPU at no load):


Speed Sabrent on Z390 system (CPU at 90% load):


Speed Sabrent on X570 system (CPU no load):


Speed Sabrent on X570 system (CPU some minor load):


I took out the Sabrent M.2 and the 10G card form the MB and did the same test. At some random times the speed would go down to 200/140 MB/s even after startup.

I even tested the speed on different SATA ports and with older drivers (even no driver) and the result was the same. Tried older BIOS and other AMD drivers as well.

I even played around forcing the MB to work on GEN3 both on the GPU and the M.2 slots but after sometime the speed of the SATA port on the SSD would go down again. I read a lot about the PICe lanes on the x570 and the CPU.

I know that the CPU has 24 PCIe lanes (4 lanes shared between CPU and the x570 for interfaces) so 20 available and x570 has 16 lanes.

My system and PCIe lanes:

GPU x8 Gen4 (CPU) (running at x8 because the 10Gige card is also running at x4 and the last PCIe running at x4 so basically x8/x4/x4 ; I even tried with no 10gige lan and x8 on the GPU)
10Gige x4 (CPU)
Samsung M.2 x8 Gen3 (CPU)
Sabrent M.2 x8 Gen3 (x570)
SSD from Sata on x570

As I mentioned i even tried taking out the 10gige and the Sabrent M.2 and the SSD speed was still unually slow at 200/140 MB/s.
I tried the SSD on a z390 Intel 9700K system running at 100% CPU load and the speed was just normal so this could not be related to PICe lane. The other PC that I have is using almost identical specs (970EVO M.2, Sabrent SB-ROCKET-2TB, 1050TI, and the exact same Mellanox ConnectX-2 Ethernet Adaptor (SingleSFP+)).

Currently I switched from the X570 MB and used the exact same hardware on a GIGABYTE Z490 10900k and and I don't see any SATA speed issues on the Z490 Intel MB so this must be either a driver issue or a X570 Chipset issue.

I have read on many forums and youtube video that many people are having speed issue on M.2 or SATA SSD speed:

https://www.reddit.com/r/Amd/comments/czg84a

Anyone with a ASUS TUF X570-Plus getting slow SSD Speeds

so I trouble shooting my ASUS TUF BOARD I have a 1TB SSD that I was getting in the 400s and faster write speeds etc I delt with WD and they RMA'd it I tested the new 1tb ssd and still poor speeds I also swaped port cables and no change.. the newer SSD does get better speeds but not really and the...

linustechtips.com

Slow NVMe speeds after upgrading to x570 + R5 3600

Hey guys just bought a 3600 + x570 taichi, updated to latest bios and all AMD drivers installed (I think). I have 3 NVMe ssd's: Adata sx7000 (M2_1) Adata sx8200 (M2_2) Adata sx7000 (M2_3) Both of the Adata sx7000s hit their rated speeds, however the sx8200 tops out at 2,400 MB/s read and 1,600 MB...

linustechtips.com

So this is the reason I had to return my x570 MB and 3900X and still deciding to go with B550 or another brand of x570.
Currently I have a GIGABYTE AORUS ULTRA Z490 with Intel 10900k and the exact same hardware and have not seen any SSD or M.2 issue at all with CPU running at 90% and above.
Please let me know if anyone else is having the same issue.

I think this is very true:

"""""""""""""""""""
UPDATE:
B550 has solved this issue compared to the X570 chipset. It really seems AMD messed up with X570.
""""""""""""""""""""""

Thanks.

 

[AmericanLocomotive]

I currently have a 3900X in an ASUS TUF Gaming x570 with three drives:

- 1TB Samsung 970 Pro connected to the CPU nvme slot
- 1TB Samsung 970 Evo connected to the x570 nvme slot
- 1TB WD Black HDD connected to SATA

My system has been on for several hours and using Crystal Disk Mark 8.0.1, my drives get the following read speed:
- 970 Pro : 3500 MB/s
- 970 Evo: 3300 MB/s
- WD Black: 132 MB/s

With Cinebench R20 MT running:
- 970 Pro: 3528 MB/s
- 970 Evo: 3314 MB/s
- WD Black: 135 MB/s

So, I'd say it's more of a motherboard/system issue than an x570 issue.

 

[parhamsan]

Hi AmericanLocomot,

Thanks for sharing your benchmark.

The issues might be related to the GIGABYTE X570 AORUS ULTRA MB.
Unfortunately, I have returned both my MB and 3900x but I have my 5900x coming soon so I can probably order MBs from ASUS (X570) and GIGABYTE (B550) and test the CPU and see if the SATA speed issue persists.

Can you mention what GPU you are using and if you are using any other full length PCIe device?
Also did you install the AMD chipset driver 2.10.13.408 or you let the drivers to be detected by windows?

As you saw (at the time of my testing) I have tried taking out the second M.2 (Sabrent) and the 10gb card so I don't think the slow SATA speed was related to exceeding the PCIe lanes on the MB (the test was done only with 970 pro as boot drive, EVGA 3080ti, and the MX500 SSD).

The very odd thing was that I was experiencing very slow performance on the Crucial MX500 SSD compared to other people seeing minor speed drops on their SSD.
I have found another post so was definitely not the only one with this issue:

Anyone with a ASUS TUF X570-Plus getting slow SSD Speeds

so I trouble shooting my ASUS TUF BOARD I have a 1TB SSD that I was getting in the 400s and faster write speeds etc I delt with WD and they RMA'd it I tested the new 1tb ssd and still poor speeds I also swaped port cables and no change.. the newer SSD does get better speeds but not really and the...

linustechtips.com

I still believe there is an issue with the x570 chipset but might be limited to certain MB brands.

Thanks.

 

[AmericanLocomotive]

I use a 5700 XT in PCIE 4.0 mode.

 

[Kenmitch]

As you saw (at the time of my testing) I have tried taking out the second M.2 (Sabrent) and the 10gb card so I don't think the slow SATA speed was related to exceeding the PCIe lanes on the MB (the test was done only with 970 pro as boot drive, EVGA 3080ti, and the MX500 SSD).

There's your problem. Get a real GPU and try again! lol