AMD Ryzen 7040 Series Shows Great AVX-512 Performance For Laptops / Mobile / Edge

[phoronix]

Phoronix: AMD Ryzen 7040 Series Shows Great AVX-512 Performance For Laptops / Mobile / Edge

Similar to what I have shown with the Ryzen 9 7950X AVX-512 desktop performance and AMD EPYC 9004 series AVX-512 server performance, the new Ryzen 7040 series mobile processors are exhibiting great AVX-512 performance for laptops. In today's article is a look at the performance impact when toggling AVX-512 capabilities for a Ryzen 7 7840U "Phoenix" SoC compared to toggling AVX-512 with prior Intel Tiger Lake and Ice Lake laptops that offer AVX-512.

https://www.phoronix.com/review/amd-ryzen7040-avx512

[ddriver]

Amd beating intel at its own game in avx 512 and smt.

Meanwhile, intel mocked amd's glue (which captured billions worth of market), then went on to pimp their own super-glue, then evidently super-glued their fingers together.

[coder]

The results don't look quite so impressive, when you realize they're showing an 8-core CPU vs. two 4-core CPUs.

My next thought was: "wow that Tiger Lake laptop is really smoking the Ice Lake". But, then I noticed the difference in power consumption and it seemed to explain a lot of that. In fact, the Tiger Lake laptop's power consumption went up by even more than its performance, which is why it's consistently worse than Ice Lake on the perf/W graphs.

Finally, looking at power consumption, it's pretty impressive to see the 7840 is much closer to Ice Lake than Tiger Lake. On all the perf/W metrics, the new Ryzen really smokes them both.

[dkokron]

libxsmm with MNK=64 is a bandwidth limited operation and therefore not well suited for this comparison. You need a much larger matrix size to get the arithmetic intensity up high enough to make it compute bound. Only then will you see any impact from avx-512. I also notice in the first libxsmm plot that it was compiled with march=core-avx2!?

[ddriver]

The results don't look quite so impressive, when you realize they're showing an 8-core CPU vs. two 4-core CPUs.

The article investigates avx 512 performance, it doesn't compare apples to oranges. But the perf per watt is core count agnostic, and intel is getting decimated there.

Of course those are older gen chips as well, I suspect 12 and 13 gen will to a tad better.

[Anux]

Of course those are older gen chips as well, I suspect 12 and 13 gen will to a tad better.

We at least know that 12th gen doesn't fare much better: https://www.phoronix.com/review/amd-ryzen7-7840u
But since 12th and 13th gen have no AVX512 we can't compare them in this context.

[avis]

And I've a got a Ryzen 7 7840HS based laptop, perhaps the only one in my country. I guess almost 8 years is enough for my previous laptop.

Haven't run Linux on the new one yet because SystemRescue doesn't support its WiFi card, it has no NIC/LAN and I have no large enough USB flash drive to copy my previous laptop Linux partition to it, so it will take time to figure it all out. I'll check if Fedora Live ISO works on it and then will try to send the data via Wi-Fi.

AMD laptops unfortunately have some obscure Wi-Fi cards which don't always work nicely with Linux. Intel laptops with Intel AX201 Wi-Fi work near perfectly.

[stormcrow]

We at least know that 12th gen doesn't fare much better: https://www.phoronix.com/review/amd-ryzen7-7840u
But since 12th and 13th gen have no AVX512 we can't compare them in this context.

That's not a 7040. It would be relevant to know if AVX-512 makes much of a difference between the 7040 and 12th or 13th gen Intel equivalent sans the extension. My hot take is that it may only make a practical difference in cases where there's some heavy optimization going on for very particular work loads. Otherwise, it's probably a non-issue in day-to-day road warrior office work so it doesn't matter if 12th & 13th gen desktop/laptop chips have it for most people. That's not to say I'm letting Intel off for its artificial differentiation of products, only that from a practical standpoint most consumers aren't going to care.

[yump]

But the perf per watt is core count agnostic, and intel is getting decimated there.

It's not. With the same microarchitecture and power budget, more cores at lower frequency get more work done than fewer cores at high frequency.

[stormcrow]

It's not. With the same microarchitecture and power budget, more cores at lower frequency get more work done than fewer cores at high frequency.

Only true if a task is capable of being broken down into parallel threads that can run at the same time. If it's single threaded and synchronous, multiple cores aren't going to help and the only instruments you care about are any helper OPs to minimize the number of cycles needed for each instruction. The faster the CPU can dispatch those instructions, the better... at least till you hit externals like writing to storage. Like in most cases in life "it depends", especially when CPU makers are artificially changing available CPU instructions in the same microarchitecture.