I'm just happy that finally, with the popularity of zen4 and 5 chips, AVX512 is around ~20% of the running hardware in the steam hardware survey. It's going to be a long while before it gets to a majority - Intel still isn't shipping its own instruction set in consumer CPUs - but its going the right direction.
Compared to the weird, lumpy lego set of avx1/2, avx512 is quite enjoyable to write with, and still has some fun instructions that deliver more than just twice the width.
Personal example: The double width byte shuffles (_mm512_permutex2var_epi8) that takes 128 bytes as input in two registers. I had a critical inner loop that uses a 256 byte lookup table; running an upper/lower double-shuffle and blending them essentially pops out 64 answers a cycle from the lookup table on zen5 (which has two shuffle units), which is pretty incredible, and on its own produced a global 4x speedup for the kernel as a whole.
Sure.. in detail and abstracted slightly, the byte table problem:
Maybe you're remapping RGB values [0..255] with a tone curve in graphics, or doing a mapping lookup of IDs to indexes in a set, or a permutation table, or .. well, there's a lot of use cases, right? This is essentially an arbitrary function lookup where the domain and range is on bytes.
It looks like this in scalar code:
transform_lut(byte* dest, const byte* src, int size, const byte* lut) {
for (int i = 0; i < size; i++) {
dest[i] = lut[src[i]];
}
}
The function above is basically load/store limited - it's doing negligible arithmetic, just loading a byte from the source, using that to index a load into the table, and then storing the result to the destination. So two loads and a store per element. Zen5 has 4 load pipes and 2 store pipes, so our CPU can do two elements per cycle in scalar code. (Zen4 has only 1 store pipe, so 1 per cycle there)
Here's a snippet of the AVX512 version.
You load the lookup table into 4 registers outside the loop:
Then, for each SIMD vector of 64 elements, use each lane's value as an index into the lookup table, just like the scalar version. Since we only can use 128 bytes, we DO have to do it twice, once for the lower and again for the upper half, and use a mask to choose between them appropriately on a per-element basis.
auto tLow = _mm512_permutex2var_epi8(p0, x, p1);
auto tHigh = _mm512_permutex2var_epi8(p2, x, p3);
You can use _mm512_movepi8_mask to load the mask register. That instruction sets each lane is active if its high bit of the byte is set, which perfectly sets up our table. You could use the mask register directly on the second shuffle instruction or a later blend instruction, it doesn't really matter.
For every 64 bytes, the avx512 version has one load&store and does two permutes, which Zen5 can do at 2 a cycle. So 64 elements per cycle.
So our theoretical speedup here is ~32x over the scalar code! You could pull tricks like this with SSE and pshufb, but the size of the lookup table is too small to really be useful. Being able to do an arbitrary super-fast byte-byte transform is incredibly useful.
The main problem with SIMD instructions is that regular code doesn't use them. Almost always someone need to write SIMD code manually to achieve good performance, which is rarely done and if so, only in some tight loops and niche cases. Like cryptography-related code in a browser may be SIMD-based, but other code uses almost no SIMD.
Modern compilers are able sometimes to vectorize regular code, but this is done only occasionally, since compilers can't often prove that read/write operations will access valid memory regions. So one still needs to write his code in such a way that compiler can vectorize it, but such approach isn't reliable and it's better to use SIMD instruction directly to be sure.
When I optimize stuff, I just think of the SIMD instructions as a long sandwich toaster. You can have a normal toaster that makes one sandwich, or you can have a 4x toaster that makes 4 sandwiches as once. If you have a bunch of sandwiches to make, obviously you want to align your work so that you can do 4 at a time.
If you want to make 4 at a time though, you have to keep the thing fed. You need your ingredients in the cache, or you are just going to waste time finding them.
Compared to GPU programming the gains from SIMD are limited but it's a small-multiple boost and available pretty much everywhere. C# makes it easy to use through Vector classes. WASM SIMD still has a way to go but even with the current 128-bit you can see dramatic improvements in some buffer-processing cases (I did a little comparison demo here showing a 20x improvement in bitwise complement of a large buffer: https://www.jasonthorsness.com/2)
Quick reminder that a 20x boost is better than going from O(n) to O(log n) for up to a million items. And, that log n algorithms often are simply not possible for many problems.
Am I getting the math wrong here? Going from O(n) to O(log n) (with no change in constant factor) for a million items would be going from ~1000000c ops to ~20c ops, which would be a 50000x improvement?
The high arithmetic bandwidth on GPUs is of course SIMD based as well. They just tend to have a ISPC style compilation model that doesn't expose the SIMD lanes in the source code. (Whereas SIMD even after decades is very lightly utilized by compilers on the CPU side).
The WASM folks should just include an arbitrary-length vector compute extension. We should also explore automatically compiling WASM to GPU compute as appropriate, the hardware independence makes it a rather natural fit for that.
4 lanes of SIMD (like in say SSE) is not necessarily 4x faster because of the memory access, sometimes it's better than that (and often it's less).
PSHUFB wins in case of unpredictable access patterns. Though I don't remember how much it typically wins.
PMOVMSKB can replace several conditionals (up to 16 in SSE2 for byte operands) with only one, winning in terms of branch prediction.
PMADDWD is in SSE2, and does 8 byte multiplies not 4.
SSE4.1 FP rounding that doesn't require changing the rounding mode, etc.
The weird string functions in SSE4.2. Non-temporal moves and prefetching in some cases.
The cool thing with SIMD is that it's a lot less stress for the CPU access prediction and branch prediction, not only ALU. So when you optimize it will help unrelated parts of your code to go faster.
Related: Go is looking to add SIMD intrinsics, which should provide a more elegant way to use SIMD instructions from Go code: https://go.dev/issue/73787
Much better to burn the area for multiple smaller units, its a bit more area for frontend handling, but worth it for the flexibility (see Apple's M-series chips vs intel avx*).
Yes and no. I think neon is undersized for today at 128bit registers -- if you're working with doubles for example, that's only two values per register, which is pretty anemic. Things like shuffles and other tricky bitops benefit from wider widths as well (see my other reply)
Agreed that 128 bit is undersized, but 512 feels pretty good for the time being. We're unlikely to see further size increases since going to 1024 would require doubling the cache line, register file, and ram bandwidth, while just adding an extra fma port is far less hardware.
totally - especially given how bandwidth constrained CPUs still are, going wider than 512 doesn't make much sense. 512 itself was a stretch for quite a long time (and all the negative press on the original implementations was a consequence of being not-quite-ready for primetime), but for current hardware I think it's perfect.
But 128bit is just ancient. If you're going to go to significant trouble to rewrite your code in SIMD, you want to at least get a decent perf return on investment!
It's less a difference in instruction set capability and more a difference in mentality.
Like, for SIMD, you have to say "ok, we're working in vector land now" and start doing vector loads into vector registers to do vector ops on them. Otherwise, the standard variables your program uses are scalars and you get less parallelism. On a GPU this is flipped: the regular registers are vector, and the scalar ones (if you have any) are the weird ones. And because of this the code you write is (more or less) scalar code where everything so happens to magically get done sixteen times at once.
As you can imagine, this isn't foolproof and there's a lot of other things that have to change on GPU programming in order to be viable. Like, conditional branching has to be scalar, since the instruction pointer register is still scalar. But you can have vectors of condition flags (aka "predicates"), and make all the operations take a predicate register to tell which specific lanes should and shouldn't execute. Any scalar conditional can be compiled into predicates, so long as you're OK with having to chew through all instructions on both branches[0].
[0] A sufficiently smart shader compiler could check if the predicate is all-false or all-true and do a scalar jump over the instructions that won't execute. Whether or not that's a good idea is another question.
One way to think of SIMT is that instead of vector instructions you have a 'fork' instruction which turns on vector mode where the scalar instructions execute on all the vector lanes. Your SIMT code must then include a 'lane index' variable somewhere (of course e.g. in CUDA it's more complicated with blocks, warps etc etc., but in principle it's just a more detailed way of doing lane indexing) so that all the threads work on different data. There is traditionally a shared program counter (PC) (in reality on GPU's, something like per-warp PC's so you still have multiple PC's), where in case of divergent control flow lanes are masked off (though post-Volta Nvidia HW has per-lane PC's). Then finally when you're done with your parallel algorithm you execute a 'join' instruction which blocks until all the lanes have reached that point, and then turns off all the lanes except one so you're now in scalar mode again.
Now whether this is actually how the hardware operates or whether the compiler in the GPU driver turns the SIMT code into something like SIMD code for the actual HW is another question.
I think what you're describing is SPMD, which is a compilation strategy, not a hardware architecture. I am not sure but I think SIMT is SIMD but with multiple program counters (1 per N lanes) to enable some limited control flow divergence between lane groups.
The PC is shared in traditional SIMT, but diverging branches are masked out until they execute. Nvidia introduced per-thread PCs with Volta. I think AMD still uses a shared PC across each wavefront?
No mention of branches, which is a complementary concept. If you unwind your loop, you can get part of the way to SIMD performance by keeping the CPU pipeline filled.
Why does such an abbreviation still exist in 2025?
They have been in the CPUs for so long that I expected them to be inseparable to the degree that people wouldn't even remember they were a separate thing in the past.
Compared to the weird, lumpy lego set of avx1/2, avx512 is quite enjoyable to write with, and still has some fun instructions that deliver more than just twice the width.
Personal example: The double width byte shuffles (_mm512_permutex2var_epi8) that takes 128 bytes as input in two registers. I had a critical inner loop that uses a 256 byte lookup table; running an upper/lower double-shuffle and blending them essentially pops out 64 answers a cycle from the lookup table on zen5 (which has two shuffle units), which is pretty incredible, and on its own produced a global 4x speedup for the kernel as a whole.
Maybe you're remapping RGB values [0..255] with a tone curve in graphics, or doing a mapping lookup of IDs to indexes in a set, or a permutation table, or .. well, there's a lot of use cases, right? This is essentially an arbitrary function lookup where the domain and range is on bytes.
It looks like this in scalar code:
transform_lut(byte* dest, const byte* src, int size, const byte* lut) { for (int i = 0; i < size; i++) { dest[i] = lut[src[i]]; } }
The function above is basically load/store limited - it's doing negligible arithmetic, just loading a byte from the source, using that to index a load into the table, and then storing the result to the destination. So two loads and a store per element. Zen5 has 4 load pipes and 2 store pipes, so our CPU can do two elements per cycle in scalar code. (Zen4 has only 1 store pipe, so 1 per cycle there)
Here's a snippet of the AVX512 version.
You load the lookup table into 4 registers outside the loop:
Then, for each SIMD vector of 64 elements, use each lane's value as an index into the lookup table, just like the scalar version. Since we only can use 128 bytes, we DO have to do it twice, once for the lower and again for the upper half, and use a mask to choose between them appropriately on a per-element basis. You can use _mm512_movepi8_mask to load the mask register. That instruction sets each lane is active if its high bit of the byte is set, which perfectly sets up our table. You could use the mask register directly on the second shuffle instruction or a later blend instruction, it doesn't really matter.For every 64 bytes, the avx512 version has one load&store and does two permutes, which Zen5 can do at 2 a cycle. So 64 elements per cycle.
So our theoretical speedup here is ~32x over the scalar code! You could pull tricks like this with SSE and pshufb, but the size of the lookup table is too small to really be useful. Being able to do an arbitrary super-fast byte-byte transform is incredibly useful.
Modern compilers are able sometimes to vectorize regular code, but this is done only occasionally, since compilers can't often prove that read/write operations will access valid memory regions. So one still needs to write his code in such a way that compiler can vectorize it, but such approach isn't reliable and it's better to use SIMD instruction directly to be sure.
If you want to make 4 at a time though, you have to keep the thing fed. You need your ingredients in the cache, or you are just going to waste time finding them.
Quick reminder that a 20x boost is better than going from O(n) to O(log n) for up to a million items. And, that log n algorithms often are simply not possible for many problems.
There wasn't much appetite for any of it on Emscripten.
https://github.com/WebAssembly/wasi-libc/pulls?q=is%3Apr+opt...
PSHUFB wins in case of unpredictable access patterns. Though I don't remember how much it typically wins.
PMOVMSKB can replace several conditionals (up to 16 in SSE2 for byte operands) with only one, winning in terms of branch prediction.
PMADDWD is in SSE2, and does 8 byte multiplies not 4. SSE4.1 FP rounding that doesn't require changing the rounding mode, etc. The weird string functions in SSE4.2. Non-temporal moves and prefetching in some cases.
The cool thing with SIMD is that it's a lot less stress for the CPU access prediction and branch prediction, not only ALU. So when you optimize it will help unrelated parts of your code to go faster.
They were notable for several reasons, although they are no longer included in modern silicon.
https://en.wikipedia.org/wiki/3DNow!
Of course memory bandwidth should increase proportionally otherwise the cores might have no data to process.
But 128bit is just ancient. If you're going to go to significant trouble to rewrite your code in SIMD, you want to at least get a decent perf return on investment!
It's less a difference in instruction set capability and more a difference in mentality.
Like, for SIMD, you have to say "ok, we're working in vector land now" and start doing vector loads into vector registers to do vector ops on them. Otherwise, the standard variables your program uses are scalars and you get less parallelism. On a GPU this is flipped: the regular registers are vector, and the scalar ones (if you have any) are the weird ones. And because of this the code you write is (more or less) scalar code where everything so happens to magically get done sixteen times at once.
As you can imagine, this isn't foolproof and there's a lot of other things that have to change on GPU programming in order to be viable. Like, conditional branching has to be scalar, since the instruction pointer register is still scalar. But you can have vectors of condition flags (aka "predicates"), and make all the operations take a predicate register to tell which specific lanes should and shouldn't execute. Any scalar conditional can be compiled into predicates, so long as you're OK with having to chew through all instructions on both branches[0].
[0] A sufficiently smart shader compiler could check if the predicate is all-false or all-true and do a scalar jump over the instructions that won't execute. Whether or not that's a good idea is another question.
Now whether this is actually how the hardware operates or whether the compiler in the GPU driver turns the SIMT code into something like SIMD code for the actual HW is another question.
Why do we even need SIMD instructions? - https://news.ycombinator.com/item?id=44850991 - Aug 2025 (8 comments)
They have been in the CPUs for so long that I expected them to be inseparable to the degree that people wouldn't even remember they were a separate thing in the past.