Intel Arc Graphics A750 + A770 Linux Gaming Performance

[cybereality]

So I spent the day reinstalling Ubuntu 22.04 and following the steps from Intel's page exactly. I did not install anything else of the system.

https://dgpu-docs.intel.com/installa...jammy-arc.html

I did every step and verified that it was done correct and working. Everything looked good, but games were still not working. Every game I tried crash after a few seconds, and I even got locks and hangs on the desktop doing nothing. This is really bad, unfortunately. I knew there were going to be bugs and stuff, but I expected it to be at least somewhat usable. Right now it's basically not even functioning, even with Intel's special kernel and drivers. Put the Arc A750 in the closet, will wait a few months before I bother testing again.

[cybereality]

I got my Intel Arc A750 a few days ago and I've been trying to get it working on Ubuntu. I think I did all the steps but the performance is bad and I'm getting crashes and glitches all over the place. I understand this is super alpha territory, but I'm hoping there is some way to fix this. You can read my review here with some details.

https://www.reddit.com/r/linux_gamin...w_and_testing/

My question is that I think Resizable Bar may not be working. I'm not sure if that's my problem. I tried this command from the ArchWiki.

dmesg | grep BAR=

That works with my AMD card on the same system and shows the full VRAM. But it displays nothing with the Intel A750. I also tried this command.

lspci -v​

Which shows this:

09:00.0 VGA compatible controller: Intel Corporation DG2 [Arc A750] (rev 08) (prog-if 00 [VGA controller])
Subsystem: Intel Corporation Device 1021
Flags: bus master, fast devsel, latency 0, IRQ 94, IOMMU group 19
Memory at fb000000 (64-bit, non-prefetchable) [size=16M]
Memory at fc00000000 (64-bit, prefetchable) [size=8G]
Expansion ROM at fc000000 [disabled] [size=2M]
Capabilities: [40] Vendor Specific Information: Len=0c <?>
Capabilities: [70] Express Endpoint, MSI 00
Capabilities: [ac] MSI: Enable+ Count=1/1 Maskable+ 64bit+
Capabilities: [d0] Power Management version 3
Capabilities: [100] Alternative Routing-ID Interpretation (ARI)
Capabilities: [420] Physical Resizable BAR
Capabilities: [400] Latency Tolerance Reporting
Kernel driver in use: i915
Kernel modules: i915​

That looks like it's working but the performance is still really bad. Any advice, did I do something wrong?

[GunpowderGuy]

Rather underwhelming, even in terms of performance-per-dollar. I'm curious how it performs with and without reBAR. Ostensibly, it seems the Linux team managed to decrease their dependency on system memory, so in theory reBAR shouldn't have a dramatic effect on performance.

could you point to the commits or messages that indicate that?

[coder]

I would say the current AM5 plattform isn't economic for consumer desktops either.

Well, yes. That's in line with my case against PCIe 5.

the whole AM5 plattform is an enthusiast only thing.

For now, it is. It needs to scale down, though. Entry-level is going to be BGA SoCs, but mid-range and slightly above needs to maintain a similar price point as before.

I've previously called AVX-512 a trap that Intel set for AMD. AMD was wise not to walk into it, until they could do so at no major deficit. PCIe 5.0 was another trap, and AMD walked right into it. Let's hope their low-end boards are PCIe 4.0-only.

[Anux]

You're being too dismissive of the practical and economic challenges involved in making faster SSDs. Since this entire exchange started with PCIe 5.0 not making much practical or economic sense for consumer desktops, such concerns are very relevant.

I would say the current AM5 plattform isn't economic for consumer desktops either.

I don't believe it simply didn't occur to anyone to add more channels to their controllers.

It has been done on the switch to PCIe 4 SSDs and therefore it will be done for PCIe 5, of that I'm sure.

As you said, it isn't for the regular consumer. But again the whole AM5 plattform is an enthusiast only thing.

[coder]

Why are you aways so stubborn?

You're being too dismissive of the practical and economic challenges involved in making faster SSDs. Since this entire exchange started with PCIe 5.0 not making much practical or economic sense for consumer desktops, such concerns are very relevant.

They could just build a controller with more channels or what ever gives them more parallel access. They build new controllers anyway.

I don't believe it simply didn't occur to anyone to add more channels to their controllers. You can actually find news coverage of new SSD controllers being announced, and the better reviews of SSDs delve into such details as which controller model is being used and how it's configured.

Born out of the (IMO, somewhat artificial) demand from PCIe 5.0 desktop platforms, there are PCIe 5.0 controllers and SSDs containing them on their way to market. Bolstered by faster flash, we should see speeds starting to peak above PCIe 4.0 limits. But, like we've been seeing just about everywhere else, it will come at the expense of significantly higher power consumption & heat output, not to mention cost. So, I don't consider it much in the way of progress.

[Anux]

Why are you aways so stubborn? They could just build a controller with more channels or what ever gives them more parallel access. They build new controllers anyway.

[coder]

Exactly and by just extending that one could easily saturate PCIe 5 bandwidth.

No, because they're already maxing the highest-end controllers. I don't know why you seem to think they just left obvious performance wins unexploited. Besides, you can only fit so many chips on a M.2 board.

[Anux]

That's what SSD controllers already do, and the high-end drives routinely max out the number of channels.

Exactly and by just extending that one could easily saturate PCIe 5 bandwidth.

[coder]

Some exciting news about Intel Arc support in Pytorch/Tensorflow: https://github.com/oneapi-src/oneDNN/issues/1465

I've been impressed with OpenVINO, on their iGPUs and CPUs. You can use it in GStreamer pipelines, using the elements provided in their DLStreamer project.

They're somewhat analogous to Nvidia's TensorRT & DeepStream, respectively.