Core Ultra 285K Bandwidth limitation of PCIe 5.0 dedicated SSD port

[cx-ray]

Hi,

I was testing my new Samsung 9100 Pro PCIe 5.0 SSD drive. I installed it in the top M.2_1 socket of an Asus Z890 Apex motherboard, that has a PCIe 5.0 connection to the CPU.

I immediately noticed that the max read bandwidth in CrystalDiskMark 8.0.6. was underperforming. It appears to be maxed out at around 12,400 MB/s.

The drive is specified to perform up to 14,800 MB/s. Independent online tests have confirmed this range of performance. For verification I then installed the 9100 Pro SSD in a PCIe 5.0 card (it just remaps M.2 to PCIe, no on-board controller logic present).

With the above setup on the same Z890 Apex motherboard the speed of the SSD is as expected at over 14,000 MB/s

I already contacted Asus about this. In short, they basically stated it was a CPU NVMe controller limitation.

Can you verify this is a hardware limitation of the PCIe 5.0 x4 SSD lanes of the 285K CPU?

Could it be a bugged interaction between the CPU and 9100 Pro SSD? I don't have a PCIe 5.0 SSD from another manufacturer to test.

If not a hardware limitation, can this issue perhaps be resolved by a microcode or related driver update?

See two attached screen shots of benchmark runs between native M.2 PCIe 5.0 slot on motherboard versus PCIe 5.0 card.

SSD isn't overheating during benchmarking. Max temp of the 9100 Pro didn't exceed 49C.

Best Regards,

david

[JedG_Intel]

Hello cx-ray,

Thank you for posting on Intel Community Forum.

It seems that you have concerns about the transfer speed of your drive. For me to address this properly, please share the information below.

1. Is the BIOS updated to the latest version?

2. Were you getting the supposed 14,800 MB/s previously?

3. Is it possible for you to try another processor to see if that makes any difference?

Additionally, to have a better understanding of your system configuration and components please generate System Support Utility (SSU) report. Please follow instructions here and send the report - How to get the Intel® System Support Utility Logs on Windows*

I look forward to your response.

Best regards,

Jed G.

Intel Customer Support Technician

[cx-ray]

Hi Jed,

1. Yes, the motherboard BIOS is updated to the latest publicly available version 1603 from Asus.

2. The dedicated PCIe 5.0 x4 SSD port that's connected directly to the CPU has never reached 14,800 MB/s. I only approach that speed when I connect the SSD to an M.2 port that will share  bandwidth with the PCIe x16 on the motherboard. The downside of this configuration is that the x16 slot will then drop to  x8 width.

3. I don't have another Core Ultra 9 processor with a dedicated PCIe 5.0 SSD port. The 9100 Pro SSD does approach the max advertised read speed on a Core i9 14900K system. On that platform the x16 PCIe slot will however drop to x8 width when a PCIe 5.0 x4 M.2 SSD is mounted, due to the limited number of PCIe 5.0 lanes available from the CPU..

Some other Asus owners have experienced the same issue with Arrow Lake CPUs here:

https://rog-forum.asus.com/t5/intel-800-series/z890-hero-samsung-9100-pro-speed-issues/m-p/1086372

(probably need an account on that forum to see the shared screenshots)

See attached SSU report to this reply.

[cx-ray]

Hi Jed,

1. Yes, the motherboard BIOS is updated to the latest publically available version 1603 from Asus.

2. The dedicated PCIe 5.0 x4 SSD port that's connected directly to the CPU has never reached 14,800 MB/s. I only approach that speed when I connect the SSD to an M.2 port that will share  bandwidth with the PCIe x16 on the motherboard. The downside of this configuration is that the x16 slot will then drop to  x8 width.

3. I don't have another Core Ultra 9 processor with a dedicated PCIe 5.0 SSD port. The 9100 Pro SSD does approach the max advertised read speed on a Core i9 14900K system. On that platform the x16 PCIe slot will however drop to x8 width when a PCIe 5.0 x4 M.2 SSD is mounted, due to the limited number of PCIe 5.0 lanes available from the CPU..

Some other Asus owners have experienced the same issue with Arrow Lake CPUs. I'm trying to post the URL but my post keeps getting deleted. Just search Google for  "asus 285K 9100 pro slow read speed"  and click on the rog asus forum result.

See attached SSU report to this reply.

[JedG_Intel]

Hello cx-ray,

Thank you for sharing all this information. This would really be helpful with the investigation that I will conduct. I will get back to you as soon as possible.

Best regards,

Jed G.

Intel Customer Support Technician

[Tim21]

I have the exact same issue on an MSI PRO Z890-A Wifi motherboard and a core ultra 285K processor. I have contacted both MSI and Samsung, but none of their suggestions have worked.

Following...

[JedG_Intel]

Hello cx-ray,

I'm writing to give you some suggestions and steps to follow to further identify the root cause of this issue. Since you mentioned that you do not have another Intel Core Ultra processor, kindly ask some PC shop for it or somebody who has ARL platform and Z-series motherboard.

Also we recommend doing cross testing : use this CPU and storage on other motherboard to find out if it's issue with original motherboard or not - since we know that CPU can handle full disk speed, but on different PCIe slot.

Furthermore, what's this PCIe card that you mentioned when tested NVMe in x16 PCIe slot?

Hello @Tim21,

I understand that you're also facing the same issue. To avoid confusion and for us to support you properly, please create a separate thread.

Best regards,

Jed G.

Intel Customer Support Technician

[cx-ray]

Hi Jed,

Thank you for your continued assistance with this issue. The additional steps you'd like me to take for troubleshooting make sense. However, I believe they go beyond the scope of what can be reasonably expected of an individual customer. Nowadays, many consumers purchase their components from online retailers, perhaps even the majority. Walking into a physical shop and asking for help without being a customer simply isn't an option.

Before posting here, I already contacted my motherboard manufacturer, Asus. They essentially stated that it's a CPU limitation, which is why I came here for further help.

In my troubleshooting, I confirmed that this problem affects multiple owners of Arrow Lake CPUs, across varying models and brands of Z890 series motherboards. Inserting the 9100 Pro M.2 SSD into other ports confirms that it's capable of performing up to its advertised potential. This slowdown issue specifically pertains to the onboard CPU PCIe 5.0 x4 SSD port (please see the attached image in this post, where the affected port is underlined in red).

This problem is likely to become more prevalent as PCIe 5.0 SSDs gain market share. We, as customers, need Intel to address this issue if possible, or at the very least, provide clarity regarding any limitations of the dedicated PCIe 5.0 x4 SSD port on the CPU.

>Furthermore, what's this PCIe card that you mentioned when tested NVMe in x16 PCIe slot?

It's a PCIe 5.0 card designed to house a single M.2 SSD, allowing you to insert it into a PCIe x4 motherboard slot. It was included with the Asus Z790 Apex, which was released about 2.5 years ago

Thanks,

David

[ArchieD_Intel]

Hello cx-ray,

Thank you for sharing all this information. It'll really help with my investigation about this. I'll get back to you soon.

Please feel free to reach out if you have any questions. Thank you.

Best regards,

Archie D.

Intel Customer Support Technician

[JedG_Intel]

Hello cx-ray,

Thank you for your patience.

After further review, kindly do a small compare test and the required software is: HWInfo. Please see the steps below.

1. Set in BIOS gen 5 value on M.2_1 and PCIex16 slots and switch to iGPU if possible. If not then move dGPU to other slot (just for test).

2. Mount drive in M.2_1 slot.

3. Run HWInfo (full mode-we need that summary window)

4. Go to hardware tree, then to Bus -] PCI Bus #0 -] find the one with disk connected and make a screenshot

5. Shutdown PC

6. Mount drive into this adapter and PCIex16 slot

7. Power on PC

8. Run HWInfo (full mode-we need that summary window)

9. Go to hardware tree, then to Bus -]find the one with disk connected and make a screenshot

10. Compare both - is there a difference?

11. Set PCIe x16 slot in BIOS according to GPU or to auto.

Once done with the steps above, please notify me about the results so we can further investigate this matter and ensure that we get to the bottom of this.

I hope to hear from you soon.

Best regards,

Jed G.

Intel Customer Support Technician

[cx-ray]

Hi Jed,

I followed all your suggested steps. Also, I used a dGPU instead of the iGPU, because this mobo doesn't have a native  HDMI port. It only can connect video directly through a Thunderbolt to DP/HDMI cable, which I don't have.

The only difference I see between the two setups are the performance impact and different location paths. In both instances the Samsung 9100 Pro SSD is configured as PCIe5.0 x4 at 32GT/s.

See attached screenshots for traceability of steps taken.

Best Regards,

David

[JedG_Intel]

Hello cx-ray,

This is much appreciated. I'll be continuing to investigate this matter and I'll give you an update at the earliest opportunity.

Best regards,

Jed G.

Intel Customer Support Technician

[KowalAn]

Hi,

@cx-ray  do you have any option to test some other gen5 NVMe drive?

I did a test but with Crucial disk (Crucial T700 1TB PCIe Gen5 NVMe M.2 SSD | CT1000T700SSD3 | Crucial.com) with 1GB and 8GB sample in CrystalDisk.

It's a bit slower but it was working as they wrote in specs (I thought about buying this Samsung drive but it's quite expensive).

As you mentioned disk is working fine when connected over PCIe x16, which means (logically thinking/in theory) that CPU is working fine. 

If drive was recognized and then "assigned" do disk controller but using different PCIe lanes then controller is working fine, because transfers were ok. Unless something happen to some lanes inside CPU and it can be seen in slower transfer speeds.

But on the other hand PCIe lanes are being assigned dynamically to devices during boot, so either GPU would cause some problems, drive, both or it's a crazy theory.

[NormanS_Intel]

Hello cx-ray,

I'd like to inform you that we are actively looking into the issue you reported about the Core Ultra 285K Bandwidth limitation of the PCIe 5.0 dedicated SSD port. Please be assured that we will post an update here as soon as we have more information.

Thank you for your patience and understanding.

Best regards,

Norman S.

Intel Customer Support Engineer

[RobertEnger]

Dear Intel:

I too purchased a new (very expensive) Z890 mobo and Ultra 9 285k CPU.  And I too see Gen5 M.2 SSD performance limited to 12GB and change.  Please add me to your list of disgruntled customers.

WILL THERE BE A REISSUE of the cpu in a form compatible with the z890 and 1851 socket  that will give us full throughput?

(If so, will we have to buy it, or will it be supplied as compensation for the bottleneck in the existing 285k?)

OR, will we have to wait for the next generation of Intel CPUs (using the new socket) and trash our motherboard too?  

(and if we have to wait for the next generation, will there be any attempt by Intel to release something SOONER rather than later?)

Bob Enger

[VonM_Intel]

Hello, cx-ray.

Thanks for providing the feedback to Intel, and we appreciate all the tests you have done.

The upstream memory read performance of the ARL S processor SOC tile root ports 13 and 14 (PCIe Lanes 1 to 16), configured as 1px16 or 2px8, may be reduced due to increased round-trip latencies during random upstream memory read operations. These latency variations depend on the specific workload and the capabilities of the PCIe endpoint device. Your results could be caused by the system configuration and/or SoC architecture.

At this point, we recommend contacting your motherboard manufacturer (OEM) for further assistance, as they are best positioned to support any board-level or BIOS-specific considerations that may be contributing to the behavior you're seeing. 

Additionally, I will close this inquiry now. If you need further assistance, please submit a new question as this thread will no longer be monitored. Should you have additional questions or require further clarification in the future, please do not hesitate to reach out.

@RobertEnger , @KowalAn 

I empathize with your situation and the issue you are facing. However, I kindly request that you create a new thread for your case. This will allow us to focus on your specific problem and provide you with better assistance.

Best regards,

Von M.

Intel Customer Support Technician

[RobertEnger]

For the record, I DID OPEN A TICKET OF MY OWN.  Ticket  06584461  handled by Agent Richard C.

He essentially BLEW ME OFF.  He asked ME to re-do all the research and testing that has been done by all the technical web sites and by the motherboard manufacturers (all of whom apparently point the finger at the new internal architecture of the 285k).

And Richard C spent a LOT OF TIME dealing with me.  That is time that COULD have been spent helping customers who need help.

I don't need help.  (And agent time is a valuable commodity!)  I just want to be put on a list to be notified when some "fix" is determined.

ALL THE INTEL AGENTS should be briefed on this situation.  They should NOT be asking regular end-users to do extensive tests.

That is a WASTE of AGENT TIME, and a burden (affront) to the customer.  WHAT IS INTEL THINKING?  Isn't AMD already stealing enough customers?  Why alienate your current customers???