Author Topic: Intel Arc A770 - failed experiment  (Read 7408 times)

tle

  • Sr. Member
  • ****
  • Posts: 467
  • Karma: +53/-0
    • View Profile
    • Trung's Personal Website
Intel Arc A770 - failed experiment
« on: November 01, 2022, 10:40:24 pm »
My local computer shop has finally got the first batch of A770 and A750!

After two years of waiting, I finally have my hand on this dGPU.

I am going to spend a day or two to get this working on Fedora 37. AFAIK few things must be prepared:

* Latest kernel 6.1
* Mesa 22.2 or newer
* Latest GuC firmware
* Explicitly declare i915 module probe in grub menu entry

Code: [Select]
0000:03:00.0 VGA compatible controller: Intel Corporation DG2 [Arc A770] (rev 08) (prog-if 00 [VGA controller])
Subsystem: Intel Corporation Device 1020
Device tree node: /sys/firmware/devicetree/base/pciex@600c3c0000000/pci@0/pci@0/pci@1/vga@0
Flags: fast devsel, NUMA node 0, IOMMU group 0
Memory at 600c000000000 (64-bit, non-prefetchable) [disabled] [size=16M]
Memory at 6000000000000 (64-bit, prefetchable) [disabled] [size=256M]
Expansion ROM at 600c001000000 [virtual] [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

I’ll report back to you all my results on gaming and 4K/8K AV1 decoding
« Last Edit: November 03, 2022, 08:37:12 am by tle »
Faithful Linux enthusiast

My Raptor Blackbird

tle

  • Sr. Member
  • ****
  • Posts: 467
  • Karma: +53/-0
    • View Profile
    • Trung's Personal Website
Re: Intel Arc A770
« Reply #1 on: November 02, 2022, 04:18:12 pm »
The first hurdle is to get i915 kernel module

Code: [Select]
config DRM_I915
        tristate "Intel 8xx/9xx/G3x/G4x/HD Graphics"
        depends on X86 && PCI

As we could clear see that the driver is only for x86 platform. From what I know, Intel recommends users to pair the card with 10th gen Intel CPU or latest AMD CPU for REBAR feature. However Intel also said that the card would work with older CPU though likely suffers performance due to lacking of REBAR. In theory you don't have to have REBAR to get going. I am thinking of removing the X86 condition check to see if the module could get compiled and loaded.
« Last Edit: November 03, 2022, 08:38:57 am by tle »
Faithful Linux enthusiast

My Raptor Blackbird

MauryG5

  • Hero Member
  • *****
  • Posts: 774
  • Karma: +22/-1
    • View Profile
Re: Intel Arc A770
« Reply #2 on: November 02, 2022, 04:19:29 pm »
What I'm wondering is if these cards can work well on our Power systems, when we are having difficulty with AMD's RDNA2 which has always worked on our systems as GPU products ...

tle

  • Sr. Member
  • ****
  • Posts: 467
  • Karma: +53/-0
    • View Profile
    • Trung's Personal Website
Re: Intel Arc A770
« Reply #3 on: November 02, 2022, 07:54:43 pm »
Unfortunately the i915 could NOT be compiled for ppc64le due to coupling with X86 codes.

I will get in touch with Intel driver team to see what could be done.

I am glad Michael Larabel at Phoronix has mentioned about this https://www.phoronix.com/news/Intel-Arc-Graphics-No-POWER, let's hope Intel devs would see it
« Last Edit: July 24, 2023, 05:37:28 pm by tle »
Faithful Linux enthusiast

My Raptor Blackbird

Hasturtium

  • Full Member
  • ***
  • Posts: 155
  • Karma: +10/-0
    • View Profile
Re: Intel Arc A770 - failed experiment
« Reply #4 on: November 03, 2022, 10:04:54 am »
Thanks for doing so much investigating here. I'm sad to hear Arc isn't working for Power yet, but hopefully Intel's driver team will clean up the issues soon. To contribute a little that I've learned so far:

- ReBAR's reportedly a PCIe 3.0 feature, so I'm a little surprised Blackbird and Talos don't support it. The feature is important for maximizing hardware performance on Arc / Alchemist, but I don't know how drastic a difference it would make for the less game-centric 3D use cases of a typical Power9 build.

- The video encoding and decoding are not impacted by ReBAR at all. The A750 in my Windows machine is able to transcode a 1080p MKV from Blu-ray quality to a rough equivalent of x264's Slow quality at north of 325 frames per second. Better still, the video encoder is a constant for the product line - to my knowledge the A380 is just as proficient for encoding and decoding video as an A770. Given Power9's relatively pokey SIMD and the agonizing wait we've endured to see x264 adopt optimized vector code into the codebase, a $140 USD Arc A380 could improve video encoding on the platform by a factor of ten, and more in the case of h.265 and AV1. Even with ReBAR disabled, that card would probably deliver performance north of a Radeon RX 560.

- Arc also benefits from PCIe link state power management - would you need to pass i915.aspm=0 until someone at Intel manages to get that covered as well?
« Last Edit: November 03, 2022, 09:15:20 pm by Hasturtium »

tle

  • Sr. Member
  • ****
  • Posts: 467
  • Karma: +53/-0
    • View Profile
    • Trung's Personal Website
Re: Intel Arc A770 - failed experiment
« Reply #5 on: November 03, 2022, 11:56:20 pm »
- ReBAR's reportedly a PCIe 3.0 feature, so I'm a little surprised Blackbird and Talos don't support it. The feature is important for maximizing hardware performance on Arc / Alchemist, but I don't know how drastic a difference it would make for the less game-centric 3D use cases of a typical Power9 build.

I was wrong. @q66 has confirmed ReBAR is supported. Now it is all about waiting for Intel team to refactor the driver to be independent of X86 platform.

It's good to see 3 companies NVIDIA, AMD and Intel have released GPU that supports AV1 decoding/encoding. Soon I believe Apple will incorporate that function into the next A16 or M3 SoC.
Faithful Linux enthusiast

My Raptor Blackbird

tle

  • Sr. Member
  • ****
  • Posts: 467
  • Karma: +53/-0
    • View Profile
    • Trung's Personal Website
Re: Intel Arc A770 - failed experiment
« Reply #6 on: November 24, 2022, 05:07:17 am »
There's seems progress upstream in removing x86 dependency codes

ref: https://www.phoronix.com/news/Intel-Mesa-Build-For-Non-x86

Let's hope by the end of next year we would have a fully decoupled i915 driver
Faithful Linux enthusiast

My Raptor Blackbird

MauryG5

  • Hero Member
  • *****
  • Posts: 774
  • Karma: +22/-1
    • View Profile
Re: Intel Arc A770 - failed experiment
« Reply #7 on: November 24, 2022, 02:17:41 pm »
Yes, I also read the news that they are working to make the new Arc GPUs usable also by other alternative architectures to X86. We hope that they will be well supported and that a user like us at Power can choose which GPU to buy, without being forced by the support issue... Let's see what happens in the coming months... Meanwhile, the Radeon 7000 series are about to arrive...

tle

  • Sr. Member
  • ****
  • Posts: 467
  • Karma: +53/-0
    • View Profile
    • Trung's Personal Website
Re: Intel Arc A770 - failed experiment
« Reply #8 on: December 04, 2022, 07:29:39 pm »
@q66 has been attempting to get the i915 compiled. Though he did not have success (yet)

Ref: https://twitter.com/octaforge/status/1598428172684787713?s=46&t=83FbSCd9Lfk_0C7UCEZiDA
Faithful Linux enthusiast

My Raptor Blackbird

Hasturtium

  • Full Member
  • ***
  • Posts: 155
  • Karma: +10/-0
    • View Profile
Re: Intel Arc A770 - failed experiment
« Reply #9 on: January 01, 2023, 01:24:53 pm »
It looks like post-Skylake GPUs will be supported in a new Xe/Arc driver in Linux going forward. At least some of that is likely tied to removing x86isms that would keep Intel out of datacenters running ARM64, Power, and RISC-V servers in the datacenters Intel is eager to grow into. So let's hope for good things!

Hasturtium

  • Full Member
  • ***
  • Posts: 155
  • Karma: +10/-0
    • View Profile
Re: Intel Arc A770 - failed experiment
« Reply #10 on: February 09, 2023, 03:33:22 pm »
Out of morbid curiosity, has there been any movement for Intel GPUs on ppc64le since the last update?

edit: Checking the PCIe compatibility list on the Wiki, it looks like it fails to compile. Guess we'll wait for the Xe driver before we try again...
« Last Edit: February 12, 2023, 01:54:38 pm by Hasturtium »

Hasturtium

  • Full Member
  • ***
  • Posts: 155
  • Karma: +10/-0
    • View Profile
Re: Intel Arc A770 - failed experiment
« Reply #11 on: May 10, 2023, 09:10:55 pm »
There's a bit of hubbub over on Phoronix at the moment regarding Intel's driver maintenance decisions that's making people a little nuts. To wit:

- The current i915 driver, which is the official DG2 support release, has full support for the HuC microcontroller that is chiefly responsible for the low-power video encoding built into the chip. This driver presently only works with x86/x86_64, and is also not getting support for VM_BIND / sparse residency features in Vulkan. In other words, its full 3D performance capabilities will never be realized. Why, you ask? Well...

- The planned move to the Xe driver, which supports non-x86 parts, will have modern Vulkan support, but it sounds like Alchemist will not receive a port of the HuC support, as its handling has been described as "relatively annoying" and would require special effort. Meteor Lake IGPs and future discrete graphics cards have a different solution in place that has been implemented, and which will be consistently utilized going forward.

So, without access to the integrated video encoding, and with no FP64 support in hardware, Arc becomes a whole lot less interesting on Power. Looks like I'll be selling my A750 soon, along with the i3 it's been in to date, but that's the current status. Intel's let me down on this one. Ah well - at least Fedora's moving to support Rusticl for my RX 6600 soon...
« Last Edit: May 11, 2023, 12:06:26 pm by Hasturtium »

tle

  • Sr. Member
  • ****
  • Posts: 467
  • Karma: +53/-0
    • View Profile
    • Trung's Personal Website
Re: Intel Arc A770 - failed experiment
« Reply #12 on: May 14, 2023, 09:04:06 am »
There's a bit of hubbub over on Phoronix at the moment regarding Intel's driver maintenance decisions that's making people a little nuts. To wit:

- The current i915 driver, which is the official DG2 support release, has full support for the HuC microcontroller that is chiefly responsible for the low-power video encoding built into the chip. This driver presently only works with x86/x86_64, and is also not getting support for VM_BIND / sparse residency features in Vulkan. In other words, its full 3D performance capabilities will never be realized. Why, you ask? Well...

- The planned move to the Xe driver, which supports non-x86 parts, will have modern Vulkan support, but it sounds like Alchemist will not receive a port of the HuC support, as its handling has been described as "relatively annoying" and would require special effort. Meteor Lake IGPs and future discrete graphics cards have a different solution in place that has been implemented, and which will be consistently utilized going forward.

So, without access to the integrated video encoding, and with no FP64 support in hardware, Arc becomes a whole lot less interesting on Power. Looks like I'll be selling my A750 soon, along with the i3 it's been in to date, but that's the current status. Intel's let me down on this one. Ah well - at least Fedora's moving to support Rusticl for my RX 6600 soon...

Thanks for the update. With limited resources, I don't think Intel would care much about non-X86-64 architecture.
Faithful Linux enthusiast

My Raptor Blackbird