The AVX-512 supports Intel in the personal computer segment from the first 10nm architecture, Cannon Lake. Then Ice Lake, Tiger Lake a Rocket Lake. In fact, it is supported by large processor cores Alder Lake called Golden Cove. But…
With a generation Alder Lake merged Intel in the desktop into a single large core processor (Golden Cove) and small (Atoms Gracemont). However, they do not support AVX-512. Because it would be difficult for Intel to address AVX-512 support using software on such a hardware configuration, the situation should be concluded that when small cores are active, the processor does not report AVX-512 support and it is not used.
In fact, it was a little more complicated. The pre-release test platforms that Intel provided to some partners had the AVX-512 turned off completely (more in the link below). Subsequently, prior to release, Intel allowed board manufacturers to activate AVX-512 support to the extent described above (unless small cores are active). But now, according to igor’s Web site, Intel has ordered board manufacturers to turn off AVX-512 support altogether, regardless of the state of the small cores.
igor’sLAB is trying to find reasons for this decision in an effort by Intel to increase sales of server processors (which have the AVX-512 active). Personally, however, I have a different opinion. Intel has been pushing the AVX-512 into the personal computer segment for about five years and has been trying to push it beyond its servers (it doesn’t make sense to address how successful these efforts have been). So, in the long run, he has taken steps to make the AVX-512 available to a wider audience (especially developers), which is the exact opposite of the AVX-512’s server wording (which is a thing of the past).
The reason will be different and igor’sLAB basically represented it with his performance chart, he just lacked the correct interpretation. In the graph you can see the results of the test benefiting from AVX-512 (less is better), which is (in order from above) performed on Alder Lake with 8 + 8 cores (ie without AVX-512), then with switched off small cores (ie 8 large + AVX-512) and then the same without AVX-512. Comparing the third result with the first, we find that small cores can only (accurately) compensate for the power lost by shutting down the AVX-512. In other words, 8 + 8 cores without AVX-512 gives the same performance as 8 + 0 cores with AVX-512.
So far, Intel has only had models on the market Alder Lakwith large and small cores. However, he is now preparing to release models built only on large cores. If he left support for AVX-512 (for processors without small cores) active as before, paradoxes could arise in the reviews, when in applications supporting AVX-512 a cheaper model built only on large cores is just as fast (or even slightly faster) than the much more expensive model combining both types of cores. Which is a bit stupid for Intel marketing. A cheaper model simply has to be slower than more expensive.
This explains not only why Intel shuts down the AVX-512 Alder Lakebut at the same time why he’s doing it right now, before releasing models without small cores.
The second but related thing is AVX2 support. You Intel and Alder Lake offers (AVX2 can do small nuclei, Atoms). However, this generation of processors has a special limit: AVX2 can only be used by large cores up to a clock frequency of 5.1 GHz. Although the Core i9-12900K, for example, has a 5.2GHz boost, the maximum processor frequency will be limited to 5.1GHz when using the AVX2 instructions, regardless of temperature, regardless of power consumption, simply fixed.
If there was a problem with temperature or consumption, monitoring and temperature and consumption limits would address it implicitly. So the reason will probably be elsewhere. igor’sLAB assumes that it may be the lifetime (electron migration) of the chip. It is not yet clear whether this limit will also apply to the upcoming Core i9-12900KS, which is to differ from the standard model (without S) by supporting 5.5GHz single-core and 5.2GHz all-core boost. If the limitation remained, it would basically only mean a 5.1GHz boost in AVX2 applications, which would no longer be significantly different from the 5.0GHz boost of the classic model.
–
