However, as is normal with leaks and rumors, things change. Ever since the day the leaks for Arc Alchemist started pouring in, there was a clear indication that Intel is developing two different GPUs. Firstly, a flagship die that will power Intel’s top-end SKU and cut-down versions of that GPU would be used in SKUs below the flagship in the hierarchy. This was the DG2-512 (codename) GPU, also known as “SOC 1” (internal name) and now officially known as “ACM-G10” (public name). Secondly, an entry-level GPU that would compete with lower-end GPUs from AMD and NVIDIA in the market. This GPU was codenamed DG2-128 because of its 128 Execution Units. As you can tell, DG2-512 was named as such because the full die of that GPU features 512 Execution Units. Regardless, DG2-128 was internally known as “SOC 2” and at the Arc Graphics event, Intel officially revealed it as “ACM-G11“. Now, however, new reports are coming in contradicting the existence of just these two Intel Arc GPUs. There were leaks here and there mentioning that Intel was trying to produce a third GPU as well, but it was not their focus or priority. Rumors suggested that this would be either a 448 or 336 Execution Units-carrying GPU but this new report helps shed some light on and clarify what Intel might actually be planning instead.
SOC 3?
This third GPU was found named in the Intel Graphics Compiler (IGC) for OpenCL and Ubuntu operating system in a patch that was released on March 30th, the same day Intel announced their Arc A-Series GPUs. As mentioned before, Intel announced only two GPUs that day—ACM-G10 and ACM-G11—confirming that a third GPU does not exist, or perhaps that it is in development for the desktop Arc A-Series and the mobile series is limited to just those two. We say that because the evidence for a third GPU being in development is clear through its name being spotted on the IGC GitHub. This third variant was mentioned as “DG2 256” and “ACM-G12” in the compiler, both names were used interchangeably to refer to this third GPU. This suggests that either this is a desktop GPU that hasn’t been announced yet or possibly a workstation Arc GPU in development at Intel. Regardless, that DG2 256 name basically all but confirms that we’re looking at a GPU carrying 256 Execution Units, half of what the flagship DG2-512 GPU has. In fact, that allows us to figure out that pretty much all of its specs have that same correlation Therefore, this GPU will rock 16 Xe-Cores and 2048 FP32 cores, exactly half of the top-end DG2-512 GPU (ACM-G10). We’re unsure about the transistor count but, again, judging by that half apparatus, we expect the transistor count of DG2 256 to be around 10.85 billion. However, other things like the memory bus specs are unclear and cannot be calculated like that. Though we can make an educated guess that this GPU will most likely feature a 128-bit or 192-bit wide bus. Like I said, we’re not sure why Intel decided to not talk about this GPU at their recent Arc Graphics event, but all signs point towards the direction that this is simply just a part that Intel plans on launching later down the line. Public software development for ACM-G12 has evidently started just a few weeks ago, going by the GitHub compiler. AMD and NVIDIA spread out their SKUs over a generation instead of releasing them all at once, so it makes sense that Intel would also keep a unannounced GPU at bay for future releases.
