Intel's ICC compiler once used CPUID vendor checks to route non-Intel chips, including AMD processors reporting AuthenticAMD, onto slower generic code paths even when SSE optimizations were available. The mechanism, discussed internally as a cripple-AMD feature, surfaced again through technical disassembly work and prompted Jeremy Howard to draw a parallel with recent Anthropic decisions.
Trevor was making a joke in reference to anthro, but jokes on him, ... Intel's compiler *did* intentionally generate worse code for AMD, called the "cripple AMD feature".
Remember when compilers would detect that someone was using it to build another compiler and silently inject bugs?
Reverse-engineering by Agner Fog showed the check prioritized the GenuineIntel string over actual instruction support, sending AMD hardware to inferior paths at runtime or in emitted libraries.
Lawsuits and a 2010 FTC order addressed the practice at the time, yet questions persist about whether similar vendor-specific logic lingers in any current libraries or benchmarks.
Many users condemned Intel's compiler as deliberately malevolent for using CPUID checks to cripple AMD performance out of ruthless competitive spite.
I’ve always thought the intel thing isn’t *quite* as bad as it sounds on first glance though.
Imagine BMW made an app that tune’s your car faster. It checks for a “genuine bmw” field before flashing.
Then, everyone get’s upset because it doesn’t work on a Toyota. They might even use the same engine!
Later, BMW says “oh, well we only tested it on BMW, we can’t guarantee behavior on a different platform.” I’m not sure…seems…somewhat reasonable? I know it’s not *quite* the same scenario as Intel vs AMD, but it’s pretty close.
If you're curious to read more but too lazy to prompt your llm: https://meta.ai/share/c/tTFyEhmDsW
@giffmana Yes, and the entire internet went batshit mad at them for years, and their reputation never fully recovered.
(And this was quite a bit less bad still than the reverted Anthropic sabotage attempt.)
Trevor was making a joke in reference to anthro, but jokes on him, ... Intel's compiler *did* intentionally generate worse code for AMD, called the "cripple AMD feature".
true, and to be clear, I do think that intel was absolutely trying to take advantage of the situation.
I'm more interested in the legal precedent though. Where exactly do you draw the line?
For example, would we also consider Intel's compiler "bad" for not taking advantage of a faster VP2INTERSECT on their competitors CPUs?
I'd imagine most people would say probably not. It'd be crazy to have an "Intel" compiler optimize for every scenario in a competitors architecture. It's an interesting grey area where I think we all tolerate *some* level of non-optimization on competitor's hardware.
but yes, "x87 path when SSE is right there" is likely a line too far.
@lauriewired @giffmana Compilers have long tables supporting detailed instruction selection ops, e.g. VP2INTERSECT is notably fast on Zen5 compared to any other CPU. However, those uarch variations are never in the scale of "better not enable this whole ISA subset".
@giffmana You shouldn't use AI as a substitute for human-written sources. I didn't trust that this was true until I looked it up and saw it on Wikipedia.
@jeremyphoward @giffmana I didn’t even know about this and wrote this as an example in my new article on Anthropic
@giffmana Not just the compiler. All of their old BLAS libraries that they aggressively upstreamed and turned on by default in OpenCV,Numpy etc!
@lauriewired @giffmana From the screenshot it reads as if “they could” make it work on AMD without extra effort, so the effort is actually spent on undercutting the competitor, which your version frames in the other way, where they just didn’t spend the effort to integrate to Toyota.
I used those specific brands on purpose ;)
sure, from a consumer standpoint it feels stupid. but would you really expect a bmw-specific tune (as in, a package/trim option from BMW itself) to be applied to a Toyota?
As in, drive up to the dealership, "hey, give me that z4 tune on my supra please".
I'm sure it's way beyond the legal headache they'd ever want to potentially deal with, and they'd tell you to go to Toyota instead.
I know I'm stretching it...but my overall point is how far are you supposed to optimize a competitor's execution path just because you share the ISA? Intel was over the top in this instance...but at the same time we all accept some level of "okay, yeah intel compiler works best on intel, big surprise".
@lauriewired @giffmana would be stupid when both the bmw & Toyota share the same platform & drivetrain
(& there is actually a current example for this, the bmw z4 & mkv toyota supra, both of which tune more or less the same)
its one thing if it isn't quite as fast but defaulting to the slowest path?
@giffmana This is quite a good analogy actually: if I recall correctly Intel was pilloried for this and for failing to clearly disclose it was doing this.
@giffmana What's funny is that the generated dispatch code (or library, I don't recall exactly now) is checking if CPUID family == 0x06, instead of, more logically, if family >= 0x06.
This forced Intel to give all their CPUs family ID of 0x06 for more than 30 years now.
@lauriewired @giffmana i feel like the cross licensing agreement with intel and amd for x86-64 makes it a bit more dirty than that example... though bmw and toyota also did cooperate with the supra so it would be funny if they actually did what you said
@giffmana I'm still annoyed the court decided they don't have to stop crippling AMD; they just have to document the behavior.
@lauriewired @giffmana Agreed. Intel was clearly (still is?) making excuses for giving themselves huge advantages. You don't need uarch/CPU/SKU-specific tuning for such major discriminations; the worst, un-tuned codegen with SSE2 will beat tuned-to-individual-die x87. CPUID was created for a reason.
@giffmana FTC investigated, considered it anticompetitive behavior and told Intel to knock it off.
I’m not saying this is my views of what should happen, just that there are similarities between the situations.
@lauriewired @giffmana depends on the intent. intel seemed pretty ruthless during that period in time and probably didn't like other firms butting in on x86.
this seems like the kind of thing you'd do just out of spite.
@lauriewired @giffmana I think better car analog would be the exact same engine producing more power if ECU identifier string says "Porsche" instead of "Audi". (Those cars share the same engine in some models.)
@lauriewired @giffmana Didn't Nvidia do something similar with kernels that include the name cutlass?
https://github.com/triton-lang/triton/pull/7298
