Intel Xeon Cascade Lake Compiler Performance - GCC 9/10 vs. LLVM Clang 8/9

[phoronix]

Phoronix: Intel Xeon Cascade Lake Compiler Performance - GCC 9/10 vs. LLVM Clang 8/9

At least for the newest Intel Xeon "Cascade Lake" processors, the LLVM Clang compiler is running incredibly well compared to the long-standing GNU Compiler Collection (GCC). Overall, LLVM clang is now nearly at performance parity to GCC 9 and the in-development GCC 10 compilers. Here are some Linux compiler benchmarks using the dual Intel Xeon Platinum 8280 server built around the Gigabyte S3461-3R0.

http://www.phoronix.com/vr.php?view=28006

[AsuMagic]

Huh, clang 9 seems to have improved compilation times quite noticeably. Anyone has an idea what this could be due to?
Anyways, nice results. I recall GCC 10 getting a new optimization from a redhat compiler dev, is this in yet?
Edit: Also, what could be going on with the MKL-DNN tests? Is there missing parallelization and/or vectorization on gcc?

[tildearrow]

Typo:

With the HMmer molecular biology benchmark

[milkylainen]

Michael: Saying that something is "performing better" when it's x40 the performance is quite the understatement.
In MKL-DNN for example. Performing better to me is within a resonable margin.

40x is not "performing better". It's broken in one direction or the other, somehow.

[carewolf]

Michael: Saying that something is "performing better" when it's x40 the performance is quite the understatement.
In MKL-DNN for example. Performing better to me is within a resonable margin.

40x is not "performing better". It's broken in one direction or the other, somehow.

Yeah, and probably shouldn't be used as a benchmark until we understand if it is the benchmark or compiler's fault.

[FireBurn]

Are you absolutely sure the geometric mean is calculating things correctly - taking into account the less is better results?

[Cape]

> tons of munny > tons of manpower
> extremely shitty license
>STILL PERFORMS WORST than GCC after a decade of development

Is LLVM a meme compiler now?

[eberdonn]

40x is not "performing better". It's broken in one direction or the other, somehow.

Note that the same test back in April, on the same processor but with gcc 9.0.1, yielded results more in line with llvm :
Harness: Convolution Batch conv_alexnet - Data Type: f32 => 55.35ms instead of 1890.16ms

https://www.phoronix.com/scan.php?page=news_item&px=Intel-MKL-DNN-GCC9-CKL-Tests

There is either something broken in gcc >= 9.1 or a problem with the test.

[AsuMagic]

> tons of munny > tons of manpower
> extremely shitty license
>STILL PERFORMS WORST than GCC after a decade of development

Is LLVM a meme compiler now?

Too bad this is an ignorant statement AND performance is very far from the only worthwhile features of a compiler. clang is still way ahead of gcc for compiler diagnostics. gcc barely caught up to clang in many situations. In particular C++ template instantiation errors are much less cryptic with clang than with gcc.

[F.Ultra]

> tons of munny > tons of manpower
> extremely shitty license
>STILL PERFORMS WORST than GCC after a decade of development

Is LLVM a meme compiler now?

"It's better to keep your mouth shut and appear stupid than open it and remove all doubt"
- Hávamál