Intel ARC GPUs, Xe Architecture for dGPUs [2022-]

[Man from Atlantis]

Here, I’ll be looking at a Cyberpunk 2077 frame rendered on Intel’s Arc B580, with path tracing enabled. As always, I’m focusing on how the architecture handles the workload, rather than absolute performance.

Raytracing workloads are diverse, and engineers have to allocate their transistor budget between fixed function BVH traversal hardware and regular vector execution units. It reminds me of DirectX 9 GPUs striking a balance between vertex and pixel shader core counts. More vertex shaders help with complex geometry. More pixel shaders help with higher resolutions. Similarly, BVH traversal hardware deals with geometry. Hit/miss shaders affect on-screen pixel colors, though they operate on a sample basis rather than directly calculating colors for specified pixel coordinates

Intel Arc B580 GPU, significantly improves ray tracing over its predecessor with upgraded RT Accelerators, increased BVH traversal pipelines (from 2 to 3), and doubled BVH cache (8 KB to 16 KB). These enhancements reduce memory latency and boost performance in complex workloads like Cyberpunk 2077's path tracing. With a stronger shader array, the B580 handles RT calculations more efficiently.

Raytracing on Intel’s Arc B580

Edit: The article originally said Intel’s BVH nodes were 4-wide, based on a misreading of QuadLeaf.

chipsandcheese.com

 

[trinibwoy]

Great article as always. Intel’s profiler is really impressive. It always saddens me to see the terrible utilization rates on GPU hardware. Makes you wonder why these companies keep spending so many transistors on compute when memory is almost always the bottleneck.

 

[Cyan]

Intel and Microsoft showcase AI-powered texture compression at GDC 2025 Neural rendering and neural shaders are here, and AI-powered Neural Block Texture Compression is one thing that will bring big performance gains.

At GDC 2025, Intel's Anis Benyoub is joining Microsoft on stage as it showcases 'Cooperative Vectors on DirectX,' a new feature that will lead to a 10X improvement in inference performance for Neural Block Texture Compression on Intel Arc GPUs. Joe Rozek from AMD and Alexey Panteleev from NVIDIA will also be there to discuss how Cooperative Vectors will be used with Radeon and GeForce RTX hardware.

As described by Intel, "Cooperative Vectors allow the multiplication of matrices with arbitrarily sized vectors to be run on any shader stage. Therefore, they can be used not only for training AI models, e.g., using compute shaders, but also for real-time inference, specifically per-pixel inference, to accelerate the execution of neural rendering techniques on hardware with AI acceleration." Translation, Cooperative Vectors allow AI models to run during all stages of the rendering process with direct access to game and on-screen data.

Intel's Arc hardware includes powerful XMX units for AI workloads; Cooperative Vectors will help pave the way for neural rendering and neural shaders to become an integral part of game development and how games run on modern GPUs. Intel isn't alone in supporting Neural Block Texture Compression as a way to reduce texture sizes without losing quality; AMD is also all in on the tech thanks to RDNA 4's beefy AI upgrade, while NVIDIA is also deeply invested in a range of RTX Neural Shaders aimed at improving performance and reducing the VRAM footprint of modern games.

 

[trinibwoy]

Great to see all 3 IHVs pushing forward together. Maybe we’ll actually see some results soon.

 

[Nappe1]

Intel seems to have sense of humour... they just yesterday added another new battlemage pci id to Linux kernel as well as to mesa 25.1... only few days after rumours flying intel keeping B580/B570 as only battlemage options.

and of course, they did it on april fools day.

Now there's 5 unknown 7 unknown (original insert in Q2/2024 had 2 unknowns already) Bmg pci ids , from which 5 indentifies as bmg g21 and 2 as just bmg.

EDIT: corrected the unknown count.

 

[Nappe1]

24th of March, Shane McKee has done bugfix request to Ubuntu Plucky LTS. Description says "Several new Intel Battlemage cards are unsupported in Ubuntu Plucky and lower." All three "cards" are pci_id's added in Linux Kernel in end of January.

Bug #2104011 “[FFe] Several Intel BMG SKUs not supported in Pluc...” : Bugs : intel-gmmlib package : Ubuntu

[ Impact ] Several new Intel Battlemage cards are unsupported in Ubuntu Plucky and lower. This FFe adds the minimum necessary patches (Adding PCI-IDs) to enable new Intel Battlemage hardware across several packages. No other changes are necessary for this enablement, and impact to existing...

bugs.launchpad.net

You could think that there's going to be some sort of release of these cards (0xe210 BMG-G21, 0xe215 BMG and 0xe216 BMG) if intel-connected canonical guy makes a bugfix supporting them in Ubuntu Plucky LTS as well as downporting the support to earlier releases as well...

We'll see...

 

[Nappe1]

hmmh...
What's "on die gsc non-volatile memory"?

Something that's called as such got driver support in Discrete Xe GPU driver in linux just few days ago.
NVM support as auxiliary bus device was added in last year already, but this seems to be complete new addition:

[PATCH v8 10/12] drm/xe/nvm: add on-die non-volatile memory device - Alexander Usyskin

Note that Xe3p has (at least partially and at least for now) it's own driver... Xe is shared between Alchemist and Battlemage. However it is hard to say where this connects, but there seems to be something cooking.

EDIT: it seems to be on-die, but not on-die of GPU, but as on-die in child device...

EDIT2: Here's the first message of the patches messaging: https://lore.kernel.org/all/20250414093803.2133463-1-alexander.usyskin@intel.com/

 

[Kaotik]

Note that Xe3p has (at least partially and at least for now) it's own driver... Xe is shared between Alchemist and Battlemage. However it is hard to say where this connects, but there seems to be something cooking.

It was the same with initial Battlemage drivers, they'll get unified sooner or later.

 

[Nappe1]

https://twitter.com/x/status/1913505898653196473

This seems to be buiding nicely...

Also probably first spotting of Celestial Dgpu R&D shipping... which would indicate Q2/2026 release.

https://twitter.com/x/status/1913506845735477294

 

[Nappe1]

And more follows, this time g21 based workstation cards... B93 and B35. Both with clamshell memory configuration

https://twitter.com/x/status/1913605869205348614

As said, there is 7 unknown skus in Linux / mesa drivers, so there's plenty of indications that intel is preparing more bmg generation products while developing next gen simultaneusly.

 

[Cyan]

Intel is looking into CPU overhead associated with Arc GPUs on older chips

The B580's biggest weakness is older CPUs.

www.tomshardware.com

 

[Nappe1]

Sparkle has stated that they might be releasing B580 variant with 24GB memory in may/june.

https://videocardz.com/newz/sparkle-confirms-arc-battlemage-gpu-with-24gb-memory-slated-for-may-june

 

[Sxotty]

I sure hope they don't give up on the market. Glad to see that so far they have continued. My A770 is still doing fine and it will be a while till I upgrade again.

 

[snarfbot]

they gotta be kicking themselves for canceling big battlemage

 

[Kaotik]

they gotta be kicking themselves for canceling big battlemage

The one whose parts were again in fresh shipping manifests? Doesn't sound that cancelled to me.

 

[Nappe1]

The rumour about BMG-G31 being canceled as product in Q4/2024 was originated in just single source. Yeah, he (the source) has been right many times earlier, but yet everything else seems to show that at least Intel is keeping the option alive.

My prediction is that we'll see in Q2-Q3 timeframe two Battlemage Pro cards (codenamed B35-WS and B93-WS) based on BMG-G21 and one or two consumer cards based on BMG-G31.

For following generations, Celestial / Xe3p dGPUs are in works, but I doubt we see them in next 12 months. There's almost none R&D versions moving in shipping manifests and in case of G21, it took 16 months from test tools leaks to launch. Even if Intel will be able accelerate the product cycles, it still would mean Q2-Q3 of 2026. Release and especially availability of Panther Lake CPUs with Xe3 will most likely show the way.)

Time will show how far off I am this time. (last time I was wayyy off as only Parhelia came out. aaaaand it came out 10 months late. )

 

[Kaotik]

For following generations, Celestial / Xe3p dGPUs are in works, but I doubt we see them in next 12 months. There's almost none R&D versions moving in shipping manifests and in case of G21, it took 16 months from test tools leaks to launch. Even if Intel will be able accelerate the product cycles, it still would mean Q2-Q3 of 2026. Release and especially availability of Panther Lake CPUs with Xe3 will most likely show the way.)

Celestial is coming this year in Panther Lake (which is getting similar launch to Lunar Lake with limited availability this year and more early next year)

 

[Nappe1]

Celestial is coming this year in Panther Lake (which is getting similar launch to Lunar Lake with limited availability this year and more early next year)

Yeah. Difference with battlemage is that last year in april there had been several clear dgpu r&d shipping manifests from bmg-g21. Now there is only one that could possibly be Celestial C32 dgpu development prototype. However if they are planning to use Intel own 18A for xe3p, there might not be that many shipping manifests at all.

Another thing is the extremely long development cycle of bmg dgpu variants. Scrapped bmg-g10 prototype was reported to be running in intel Labs in August 2023. So is the bmg any kind of good metrics for making conclusions is questionable.

 

[Nappe1]

According to reports from Intel Foundry Direct Connect, 18A is suboptimal for HPC/AI applications, but 18A-P is designed for it. 18A-P is starting 2026. Just wild guess but, all this with the fact they added "p" in Xe3p might possibly nicely match with rumors Celestial dGPUs being done in Intel's own fab.

https://twitter.com/x/status/1917262905822916830

This all again means that there's easily time frame where BMG-G31 could be launched to compete with 5060Ti / 5070 before the Celestial arrives.