AVX-512 was a hotly discussed topic around the launch of the new Intel Alder Lake CPUs. At first it was said that the P cores supported it in principle, but in order to be able to offer a common instruction set with the E cores, the feature would be disabled at the silicon level. Later, however, motherboard manufacturers would have found the feature functional on the CPUs and would have made it re-activatable with hacks in the microcode via the BIOS. We’ll take a look at what the recovered feature really brings today.
If you don’t know what to do with “AVX-512” yet, I’d like to briefly pick you up before we go deeper into the matter. In principle, this is an instruction set that allows the CPU to perform special types of calculations particularly quickly. First integrated on Intel’s high-end desktop platform in 2013 with the “Skylake-X” CPU generation, the feature was most notable for its extremely high power consumption. Power and fuel, you know.
CPUs of Intel mainstream platforms with smaller sockets got the feature much later and only recently, with the 11th generation. Generation, also known as Rocket Lake. This generation suffered from many other problems as we know, which is why AVX-512 didn’t really come up here. But at least it was active by default and provided a bit more performance in one application or another.
Our British colleague Dr. Ian Cutress has reported extensively on the history of AVX-512 and Alder Lake on Twitter. So for those interested in the details, I highly recommend this thread.
Just to clarify in Roman's video here. Intel's engineers (Ari) specifically stated in Architecture Day (Aug) that AVX512 was fused. I explicitly asked the question if AVX512 would work in any instance, and they said no.https://t.co/jEgnrbKTEG
— 𝐷𝑟. 𝐼𝑎𝑛 𝐶𝑢𝑡𝑟𝑒𝑠𝑠 (@IanCutress) November 7, 2021
But back to today and our test: After the launch, the fact is, the feature is physically present and has been made available by all motherboard manufacturers in BIOS updates on just about every board. The only requirement for enabling AVX-512 is to disable the E cores, which do not support the instruction set due to their architecture.
So now we have this new weird hybrid platform, with 8 P and E cores each, but only the former supporting AVX-512. So naturally I asked myself: Which configuration actually performs best, all cores, only P-cores, only P-cores with AVX-512? We had already seen in a previous review that the E-cores are not always beneficial for the performance, for example because they put additional load on the memory system or lower the cache clock, but more on that in a moment.
I’m going to bang the door in your ace, just as it banged in my face when I played around with the settings during the first tests:
Prime 95 is a well-known stress test with very high CPU load, which still has its problems with Alder Lake and only supports 16 threads, even if the new CPUs actually have 24 by default. As a result, a third of the cores are bored and the power consumption is correspondingly lower. But that’s just an issue on the side. It gets interesting when we compare the P-cores with and without AVX-512. Because with the AVX-512 feature is active, the CPU has a lower power consumption than without. “Can’t be, AVX-512 has always been the hell inferno of instruction sets, there must be something wrong with HWInfo!”, I thought to myself. So I measured with the Elmorlabs (Power Measurement Device) once on hardware level.
The device is actually just a board with measuring resistors, which is wired between the power supply and the 12 V EPS connector of the mainboard. This allows it to measure the current delivered to the motherboard for the CPU power supply with relative precision. You can see the result in the diagram above. On the one hand, it confirms the trend that 8 P-cores with AVX-512 need less power than without, and on the other hand, a big difference between power measurements in hardware and software once again becomes clear. This is where the predefined AC and DC loadlineson the mainboards come into play, which Igor already discussed the other day.
But now I wanted to know, do the P-cores with AVX-512 just do less work or is the lower power consumption actually based on higher efficiency? For this I picked out two synthetic benchmarks that are known to benefit from AVX-512, Y-Cruncher and LinX. And for interest’s sake, I threw the game Far Cry 6 into the ring, as it was able to most clearly show differences between CPUs while gaming in our past CPU generation comparison. As always, the test hardware used can be found here:
|
Test systems |
|
|---|---|
| Hardware: |
|
| Cooling: |
|
| Housing: |
|
| Periphery: |
|
| Measuring devices |
|
49 Antworten
Kommentar
Lade neue Kommentare
1
Veteran
Veteran
Mitglied
Mitglied
1
1
Urgestein
Urgestein
1
Mitglied
Mitglied
Urgestein
Veteran
Veteran
Urgestein
Alle Kommentare lesen unter igor´sLAB Community →