Intel ACM-G12 GPU found to have 16 Xe cores and 256 EU

After the Intel Arc Alchemist ACM-G10 and ACM-G11, a budget graphics card with less than 2100 cores could be coming with the ACM-G12. The news comes from Linux patches released with support for this mysterious DG2 256, which could be a desktop or workstation GPU.

Whether it is a professional graphics card or a desktop card is not yet clear, but what is clear is that it exists. Right now we have 2 confirmed DG2 series GPUs, but we could see a third variant known as DG2 256. If it exists, we should expect it in Q3 2022.

Intel ACM-G12, an unknown GPU emerges

It might just be a testbed, but it certainly won’t be the top-of-the-line model. It will be built on the same 6 nm process from TSMC and will have the Xe-HPG architecture, and our colleagues at VideoCardz place it in Intel’s SOC3.

If we look at the IGC entries on Github, you’ll see that the first one confirms that the EU is 256 (Intel now calls them Xe Vector Engines). So we’re looking at 8 Xe cores and 2048 FP32 cores, which would put it somewhere between the G10 and the G11.

There are still some specs that need to be confirmed, such as the memory bus, and here Intel has been pretty tight-lipped about whether we’ll have 128-bit or 192-bit in its arc. That information has been left out. What we do know is that software development for SOC3 started a few weeks ago.

ACM-G12, Intel ACM-G12 GPU found to have 16 Xe cores and 256 EU, Optocrypto

This graphics card would be aimed at the low-end segment, and we dare not say which GPU it could compete against. At the end of the day, the hardware distribution between AMD, Intel, and NVIDIA is so different that it’s impossible to compare specs.

It’s not like Intel or AMD processors where the spec sheet makes it clear that the frequency is what it is and the core threads are what they are. With NVIDIA, you see a lot of CUDA cores compared to AMD Radeon Stream processors, which have similar performance.

That doesn’t mean NVIDIA spends more money or that CUDA cores aren’t efficient, it’s just a more complex industry where hardware and architecture organization is everything. Intel, for example, uses EUs (Execution Units), which are complexes that house other cores.