Linux kernel developer Linus Torvalds does not shy away from harsh words. He hopes for the end of Intel’s 512-bit vector extensions and admits that he has an “irrational hatred” of the floating point benchmarks (FP) that are to be verified.
Torvalds fired his criticism of the Intel Advanced Vector Extensions 512 (Intel AVX-512) in a mailing list chat. He was responding to an article by Michael Larabel. This criticized the support for AVX-512 in the compiler instructionsfor Alder Lake, which has enabled its 2021 desktop processors, in the GNU Compiler Collection 11. Intel’s future Xeon Sapphire Roads processors still support AVX-512.
“I hope that the AVX-512 dies a painful death and that Intel starts fixing real problems instead of trying tricks to create benchmarks they can look good on,” wrote Torvalds. Torvalds criticizes Intel’s focus on FP benchmarks and the performance of its processors on supercomputers or high-performance computers (HPCs).
“I hope Intel returns to the basics: getting their process back on track and more focused on normal code that isn’t HPC or any other pointless special case.” He notes that in the heyday of x86, Intel’s rival at FP -Have always cut off loads better.
“Intel’s FP performance was poor in relative terms, and one iota wasn’t important. Because absolutely nobody was interested in it outside of the benchmarks, ”explained Torvalds. “The same applies now – and in the future – largely also for AVX-512. Yes, you can find things that affect it. No, these things don’t sell machines on a large scale. ”The AVX512 has real disadvantages. “I’d much rather see the transistor budget used for other things that are much more relevant. Even if it’s still FP math (in the GPU, instead of AVX-512). Or just give me more cores (with good single-thread performance but without the garbage like the AVX-512) like AMD did. ”
“I want my performance limits to be reached with regular integer code, not with any AVX-512 power virus that takes away the peak frequency (because people ended up using it for memcpy!) And takes away cores (because of this useless junk Eating space) ”, continued Torvalds.
“Yes, yes, I’m biased. I absolutely loathe FP benchmarks, and I realize that other people are very concerned about them. I just think that AVX-512 is just the wrong thing. It’s a favorite nuisance of mine. It’s a prime example of something Intel did wrong, in part by only increasing the fragmentation of the market. Stop littering and let all the things everyone care about go as well as you humanly can. ”
In a separate email, Torvalds emphasized that he had an “irrational hatred of vector units and FP benchmarks”. “I think they’re mostly a complete waste of transistors and effort, and I think the amount of time spent on it – by both hardware people and software people trying to use them – was mostly a waste of time, ”he wrote.
Cloudflare: “Switch off AVX-512”
In addition to Torvalds and Phoronix, the web infrastructure company Cloudflare also dealt with AVX-512. After performance analysis, she recommends customers who do not need the instruction set extension for high-performance tasks to deactivate AVX-512 on the server and desktop in order to avoid “accidental” throttling.
This decision is based on Cloudflare’s observation that the Intel chip throttles its frequency when it executes AVX code, so that other applications run correspondingly slower. “It is difficult to say how much each core is throttled at a particular point in time, but during some tests with lscpu I found that running the openssl speed -evp chacha20-poly1305 -multi 48 benchmark shows 1199.963 MHz for I received OpenSSL with all AES-GCM connections at 2399.926 MHz and for OpenSSL with all ChaCha20-Poly1305 connections I received 2184.338 MHz. … If you don’t need AVX-512 for some special high-performance tasks, I suggest that you disable AVX-512 execution on your server or desktop to avoid accidental throttling by AVX-512 code. “
In principle, Cloudflare welcomes Intel’s instruction set extensions. At the same time, however, it also restricts: “But if the problem of frequency scaling has not been solved by then, the use of this for general cryptography libraries will (much) more harm than good.”