News AMD Processors No Longer Crippled with Latest MATLAB MKL Update（10980XE v. 3960X v. 3970X)

[csbin]

AMD Processors No Longer Crippled with Latest MATLAB MKL Update

MATLAB received an update that no longer cripples users of AMD processors. Back in November 2019, there was quite some controversy when it emerged that MATLAB, a popular computing platform popular with engineering firms, universities, and research institutes, wasn't working optimally with AMD...

www.techpowerup.com

MATLAB received an update that no longer cripples users of AMD processors. Back in November 2019, there was quite some controversy when it emerged that MATLAB, a popular computing platform popular with engineering firms, universities, and research institutes, wasn't working optimally with AMD processors. Specifically, the suite's Intel MKL (math kernel library) component was designed such that if it didn't recognize the "GenuineIntel" CPUID string, it would disable fast AVX2 code-paths and fall back to SSE. This would inflict anywhere between 20-300 percent performance penalties on "AuthenticAMD" processors.

Reddit user Nedflanders1976 developed a tweak back in November, which spoofs MKL into thinking AMD processors are "GenuineIntel," enabling it to leverage modern instruction sets such as SSE4, AVX, and AVX2. AMD processors have been supporting SSE4 and AVX since its 2011 FX-series, and AVX2 since 2017 Ryzen. With the latest R2020a version, MATLAB automatically enables AVX2 execution on AMD processors that support the instruction set. A quick set of tests by ExtremeTech confirms that the update does indeed leverage the faster code-path by default, with Ryzen Threadripper 3960X and 3970X gaining over 200% performance and beating the Core i9-10980XE (something that needed the Nedflanders1976 tweak earlier).

 

[ao_ika_red]

Correction: AVX2 since Excavator / Carrizo.

 

[NTMBK]

Correction: AVX2 since Excavator / Carrizo.

Yeah, but they're still crippled by the Bulldozer architecture

 

[ao_ika_red]

Yeah, but they're still crippled by the Bulldozer architecture

OP stated AVX since bulldozer.

 

[Atari2600]

Thanks for the info.

[and who cares if it was Bulldozer or a derivative? No-one interested in performance is gonna be using those now anyway]

 

[blckgrffn]

Thanks for the info.

[and who cares if it was Bulldozer or a derivative? No-one interested in performance is gonna be using those now anyway]

Forums would be 80% less fun if people weren't pedantic and said things like "I understand your point even if your words are a little wrong" and kept their replies to themselves

Yeah, but they're still crippled by the Bulldozer architecture

But, how much? I seem to remember Dresdenboy (sp?) pointing out how much stronger the FPU was moving from the Thuban cores to these newer cores with their oversold thread counts and all that. Now that OP said an 8350 (which I don't have) might be a lot faster at MATLAB (which I don't use, haven't used in 15+ years) I want to know by how much. For science?

 

[DrMrLordX]

But, how much? I seem to remember Dresdenboy (sp?) pointing out how much stronger the FPU was moving from the Thuban cores to these newer cores with their oversold thread counts and all that. Now that OP said an 8350 (which I don't have) might be a lot faster at MATLAB (which I don't use, haven't used in 15+ years) I want to know by how much. For science?

Excavator's trouble with AVX2 support is that it was only capable of 128-bit SIMD. XV was always fastest with XOP. If you took BD, PD, SR, and XV and benched them all on a single-module basis, XV would be the fastest in any SIMD operation per module. XV's problem was that we never got more than 2M in one chip. PD went higher.