[GSD-13331] XMX regression performance of Vulkan cooperative matrix extensions after 8531 driver & oneAPI MKL 2026.1

[NikosDi]

Pre-submission Checklist

• I am using the latest GPU driver version (download center)
• I have searched for similar issues and found none

GPU Hardware

Intel Arc A380

GPU Detailed Information

PS> Get-WmiObject Win32_VideoController | Where-Object {$_.Name -like "Intel"} | Select-Object Name, DriverVersion, DriverDate, VideoProcessor, AdapterRAM, PNPDeviceID | Format-List

Name : Intel(R) Arc(TM) A380 Graphics
DriverVersion : 32.0.101.8531
DriverDate : 20260221000000.000000-000
VideoProcessor : Intel(R) Arc(TM) A380 Graphics Family
AdapterRAM : 2147479552
PNPDeviceID : PCI\VEN_8086&DEV_56A5&SUBSYS_60041849&REV_05\8&2724E0D0&0&00080000020A

Driver Installation Details

Device Manager manual installation in order to keep latest firmware from latest driver 8974

Windows Build Information

PS C:\Users\Nikos> Get-ComputerInfo | Select-Object WindowsProductName, WindowsVersion, OsHardwareAbstractionLayer | Format-List

WindowsProductName : Windows 11 Pro
WindowsVersion : 25H2
OsHardwareAbstractionLayer : 10.0.26100.1

Actual Behavior

Hello.

It seems that the last good known Vulkan runtime is in the driver 8531, regarding full performance of XMX engine using cooperative matrix extension.

After that driver, even the latest one, drops performance to around 2/3 of 8531 as measured by vkpeak for fp16-matrix and int8-matrix microbenchmarks.

Did Intel drop performance of XMX engine for Vulkan coopmat on purpose or is it a bug ?

Are you going to restore max speed of Vulkan coopmat in the next driver ?

Benchmark results:

vkpeak
device = Intel(R) Arc(TM) A380 Graphics
driver = Intel Corporation / 101.8531

fp16-matrix = 25855.50 GFLOPS
int8-matrix = 51721.49 GIOPS

device = Intel(R) Arc(TM) A380 Graphics
driver = Intel Corporation / 101.8974

fp16-matrix = 16165.81 GFLOPS
int8-matrix = 32335.90 GIOPS

Worth mentioning, I have kept latest FW from 8974 driver and I have manually installed 8531 driver.

Also using latest oneAPI MKL offline installer 2026.1 I saw the same regression like Vulkan ICD which unfortunately is not restored even using 8531 driver.
It's independent.

I would definitely try older oneAPI versions like 2026.0 or 2025.3 if you provide me the installers since they are not easily found.

BR,
Nikos

Expected Behavior

The coopmat extension of Vulkan and joint matrix & oneMKL should restore their full speed using XMX engine

Reproduction Rate

Always reproduces - 100%

Steps to Reproduce

Install 8531 driver and run vkpeak.
Then install any driver newer than that and run vkpeak.

See the fp16-matrix and int8-matrix results.

Is this a regression?

• Yes, this is a regression - functionality that previously worked is now broken

Last Known Working Driver Version

8531

First Known Failing Driver Version

No response

API Call Logs

No response

Windows Event Viewer Logs

No response

Crash Dumps / Minidumps

No response

Source Code / Reproducer

No response

Command Line / Application Details

No response

oneAPI Version (if applicable)

No response

System Configuration

No response

Screenshots / Video

No response

Intel System Support Utility Report

No response

Additional Notes

No response

[kgibala]

Hi @NikosDi,

Thank you for bringing this issue with the templates to our attention.

We will investigate the regression internally and provide an update here as soon as we have meaningful information to share.

Thank you again for your cooperation.