680 karma · joined May 19, 2020
So the H100(which costs vastly more than a Zen5..), has 14592 32 bit SIMD lanes, not cores.
A Zen 5 has 16x4(64) 32 bit SIMD lanes per core, so scale that by core count to get your answer. A higher end desktop Zen5 will have 16 cores, so 64x16 = 1024. The Zen5 also clocks much higher than the GPU, so you can also scale it up by perhaps 1.5-2x
While this is obviously less than the H100, the Zen5 chip costs $550 and the H100 cost $40k.
There is more to it than this, GPUs also have transcendental functions, texture sampling, and 16 bit ops(which are lacking in CPUs). While CPUs are much more flexible, and have powerful byte & integer manipulation instructions, along with full speed 64 bit integer/double support.
Dems basically handed this election to Trump; Biden running when he had memory issues, then handing it to Harris whose numbers were questionable against Trump(with the braindead excuse that only she could get the funds), along with generally not responding to issues voters obviously care about(border), trying to ignore issues like excessive wokeness/DEI instead of admitting it is a problem.
Trump is going to lie and obviously only cares about himself, some % of rubes will vote for that, but you don't have to make it so damn easy for him.
Which is annoying as the US is going to face some serious consequences from Trumps idiocy, phrasing it the way you did, is handing the future to Trump and his kind.
For SIMD integer to scalar integer, it has to move into separate register, so there is some short penalty(3 cycles iir).
Personally I'd say the only good reason to use assembly over intrinsics is having control over calling convention, for example the windows CC is absolute trash and wastes many SIMD registers.
On a memory safety scale I'd put C++ about 80% of the way from C to Rust.
Eventually I decided fuck this, and starting lifting weights even though it hurt, then I'd massage it which somehow reduced the pain. Repeat that, and very quickly the wrist fixed itself, zero issues now.
When I broke my collarbone years ago, the docs said just let it heal on its own, which it did just fine(I was hiking/running about 2 hours every day, maybe that helped).
I can only hope the dems are smart enough to drop DEI for good, I don't think Trump is healthy for the US.
You know the rest of them are not benefiting from that in the slightest? Actually they are probably hurt by it under the current DIE platforms.
I agree that it would be great if hiring wasn't so focused on a small group of schools(ie excuses to hire people just like yourself).
If DIE is replaced it needs to be with a system based on family income not skin color etc.
Branches fall off vs cmov/select as your lane count increases, they are still useful but only when you are certain the probability all lanes agreeing is reasonably high.
Also they are difficult to switch to, so I would expect very few established projects to bother.
Anyway auto-vectorization will never compete with intrinsics for performance, so this all seems rather silly. Try expressing pshub in auto-vectorizer. Or a bunch of optimal load/swizzles to out perform gather.
C functions can't be vectorized? WTF are you talking about? You can certainly pass vector registers to functions.
Exp can also be vectorized, AVX512 even includes specific instructions to make it easier.( there is no direct exp instruction on most hardware,it is generally a sequence of instructions)
GPU: nope, that would take much longer, Nvidia/ASML/TSMC is too far ahead