Intel Xe2 Lunar Lake Graphics Compute / OpenCL Performance Looking Great

[phoronix]

Phoronix: Intel Xe2 Lunar Lake Graphics Compute / OpenCL Performance Looking Great

Now that Linux 6.12 has a fix for the Lunar Lake performance with the ASUS Zenbook I have been using for my Core Ultra 200V series Linux testing as well as there recently being an updated Intel Compute Runtime with Lunar Lake fixes, I have been working on some fresh Lunar Lake Xe2 graphics benchmarks using the very latest upstream open-source code. In today's article is exploring how the Xe2 Lunar Lake graphics is performing for OpenCL / GPU compute relative to the prior Meteor Lake Arc Graphics that were already a nice step-up over earlier Intel integrated graphics.

https://www.phoronix.com/review/intel-xe2-lunarlake-compute

[coder]

Aside from the regressions on Blender, this is great to see! I'm definitely looking forward to picking up a Battlemage (Xe2) dGPU.

Does anyone know the nominal specs on these two iGPUs? I'm just curious how well the actual results align with the scaling predicted by a superficial comparison.

[sophisticles]

As impressive as these results are, if i was serious about GPU Compute in a laptop I would buy one with an NVIDIA GPU and not play games.

I have no doubt that even an older mobile RTX 2050 would smoke these results.

For a workstation, I do think that the combination of low prices and decent amounts of vram makes Intel dGPU a compelling buy but I would still probably pick up a low cost NVIDIA card as the primary card and use the Intel GPU was a secondary workhorse.

[coder]

As impressive as these results are, if i was serious about GPU Compute in a laptop I would buy one with an NVIDIA GPU and not play games.

Quite the word choice, there. A lot of people buying Nvidia-equipped laptops would get one primarily for playing games!
😄

That aside, I hate dGPU laptops. They're big, heavy, loud, and have terrible battery life. Things that, IMO, almost defeat the point of them being laptops.

Yes, if someone needs the most compute possible, they will obviously have to go for a dGPU. However, that doesn't mean iGPUs are worthless. If you compare the Classroom benchmark, an Intel Core Ultra 256V (Lunar Lake) renders it in 636 seconds, while its iGPU can do it in just 184 seconds. That's a 3.46x speedup you get almost for free, probably while using less power as well. And, maybe with some further tuning, the speedup will improve even more.

So, if someone just wants to run DarkTable or do some light Blender work, I think it's great to have these iGPUs as a viable option.

[oleid]

Michael How do these compare to AMD's Ryzen AI series?

[Svyatko]

Aside from the regressions on Blender, this is great to see! I'm definitely looking forward to picking up a Battlemage (Xe2) dGPU.

Does anyone know the nominal specs on these two iGPUs? I'm just curious how well the actual results align with the scaling predicted by a superficial comparison.

Intel Arc B580 in December 2024, B770 will be late (2H 2025?).

https://wccftech.com/intel-arc-b580-battlemage-gpu-leak-confirms-12-gb-memory-standard-power-connectors-up-to-2-8-ghz/

https://videocardz.com/newz/intel-arc-b580-gpu-has-tdp-lower-than-225w-and-pcie-5-0x8-interface

[coder]

Intel Arc B580 in December 2024, B770 will be late (2H 2025?).

Thanks, but I was wondering about the specs of the iGPUs that were compared in the article. I'm wondering how the real world results aligned with how they compare on paper.

[elliott]

The clpeak global memory bandwidth and ViennaCL sAXPY performance look too good to be true -- peak memory bandwidth should be 136.5 GB/s, right?

[pWe00Iri3e7Z9lHOX2Qx]

They've done nice work with their GPU compute stack over the last 2 years, but if they'd merge their SR-IOV work on iGPUs I'd seriously consider an Arrow Lake laptop.

[Svyatko]

The clpeak global memory bandwidth and ViennaCL sAXPY performance look too good to be true -- peak memory bandwidth should be 136.5 GB/s, right?

8 MiB L2 cache for GPU.