How to Bypass Matlab’s ‘Cripple AMD CPU’ Function
extremetech.comI'm glad I clicked on this story, because the title kind of buries the lede. Apparently Intel was asked to suggest a real world application for comparing its top desktop processor to AMD's top Threadripper for the desktop AM4 socket. Intel suggested Matlab.
Then this reviewer discovered via Reddit that Matlab includes an Intel library that doesn't generate SIMD optimized code for AMD processors. Forcing the library to generate the optimized code in an environment variable results in AMD crushing Intel with both tested processors.
It's remarkable both for AMD becoming clearly dominant in the desktop performance market, and for Intel promoting a dishonest benchmark, whether they knew it or not.
Intel has been accused of many things over the years, unfortunately honesty wasn't one of them.
Whats even more remarkable is MathWorks not fixing this, which suggests some Intel deal under the table. Nowadays its no longer in a form of clear $cash money$ bribe, instead Intel offers cross promotion (including you in their marketing, pushing your product to their clients), bundle deals (buy hiend motherboard, receive CPU and/or SSD for free), or bonuses (sell X amount at $MSRP, receive free product bumping your margins by as much as 20-30%).
This is hardly the first time this has happened.
In this case their math library didn't do a CPUID feature test, but rather a vendor name test. It would be equivalent to testing website compatibility by doing a user agent check with just the browser name (e.g. navigator.userAgent.indexOf("Netscape"), etc)
ICC also historically generated intentionally bad code for non-intel platforms - deliberately making function loaded depend on intel vs. everyone else rather than by feature set.
Yes, which is why I posted this story to raise awareness. It's shocking to me how even many software developers don't know about this.
Nvidia is also pretty good at this, shipping FPU optimized PhysX library when SSE was widespread.
Nice!