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.<p>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.<p>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.
Interesting example! I've been learning AVX512 by using it to optimize Huffman coding. I found _mm512_permutexvar_epi8 and used it to do byte-indexed lookups, but _mm512_permutex2var_epi8 means I can get by with 2 shuffles and 1 comparison instead of 4 shuffles and 3 comparisons. In the end, on my CPU (AMD 9950x), changing to _mm512_permutex2var_epi8 only sped up compression by ~2%.<p>Compared to Huff0[1] (used by Zstd), my AVX512 code is currently ~40% faster at both compression and decompression. This requires using 32 datastreams instead of 4 used by Huff0.<p>[1] <a href="https://github.com/Cyan4973/FiniteStateEntropy" rel="nofollow">https://github.com/Cyan4973/FiniteStateEntropy</a>
Oh, this is cool, I wanted to look into using SIMD for huffman as well.<p>For decode, do you use AVX512 to speedup the decode via caching the decode of small codewords?<p>Do you decode serially or use the self syncronizing nature of huffman codes to decode the stream from multiple offsets in parallel? I haven't seen the later done in SIMD before.<p>Are there any new SIMD instructions you'd like to see in future ISA extensions?<p>OpenPower has proposed a scalar instruction to speedup prefix-code decoding: <a href="https://libre-soc.org/openpower/prefix_codes/" rel="nofollow">https://libre-soc.org/openpower/prefix_codes/</a>
Could you please elaborate on your example? Thanks.
Sure.. in detail and abstracted slightly, the byte table problem:<p>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.<p>It looks like this in scalar code:<p>transform_lut(byte* dest, const byte* src, int size, const byte* lut) {
for (int i = 0; i < size; i++) {
dest[i] = lut[src[i]];
}
}<p>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)<p>Here's a snippet of the AVX512 version.<p>You load the lookup table into 4 registers outside the loop:<p><pre><code> __m512i p0, p1, p2, p3;
p0 = _mm512_load_epi8(lut);
p1 = _mm512_load_epi8(lut + 64);
p2 = _mm512_load_epi8(lut + 128);
p3 = _mm512_load_epi8(lut + 192);
</code></pre>
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.<p><pre><code> auto tLow = _mm512_permutex2var_epi8(p0, x, p1);
auto tHigh = _mm512_permutex2var_epi8(p2, x, p3);
</code></pre>
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.<p>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.<p>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.
Here's a non-parallel and unoptimized implementation of that operation in Go:<p><pre><code> func _mm512_permutex2var_epi8(a, idx, b [64]uint8) [64]uint8 {
var dst [64]uint8
for j := 0; j < 64; j++ {
i := idx[j]
src := a
if i&0b0100_0000 != 0 {
src = b
}
dst[j] = src[i&0b0011_1111]
}
return dst
}
</code></pre>
Basically, for a lookup table of 8-bit values, you need only 1 instruction to perform up to 64 lookups simultaneously, for each 128 bytes of table.
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.<p>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: <a href="https://www.jasonthorsness.com/2" rel="nofollow">https://www.jasonthorsness.com/2</a>)
> a small-multiple boost<p>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.
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.
Aren't they loosely modeled on riscV? RiscV has an arbitrary length vector extension. So did cray. They seem like a good idea. In practice, I usually structure intermediate data products in memory with the largest simd size I expect the vectorizer will ever optimize for, so compiling it for something longer won't get me anything.
You mean the gains from SIMD on the CPU, because all GPU programs use even wider SIMD than the CPUs, with width of 1024 bits or 2048 bits, for even greater die area and power consumption savings in comparison with non-SIMD.<p>The width of the SIMD instructions is not visible when programming with NVIDIA CUDA or with the similar compilers for Intel CPUs (ispc or oneAPI with SYCL targeting CPUs or OpenMP with appropriate pragmas), but only because the compiler takes care of that.
I merged a few PRs to SIMD optimize Wasm WASI libc, but it all got stalled in str(c)spn (which is slightly more sophisticated than the rest).<p>There wasn't much appetite for any of it on Emscripten.<p><a href="https://github.com/WebAssembly/wasi-libc/pulls?q=is%3Apr+optional+simd" rel="nofollow">https://github.com/WebAssembly/wasi-libc/pulls?q=is%3Apr+opt...</a>
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).
It's SIMD-based at the lowest level, but there's also the use of very high hardware multithreading (the threads are called, AIUI, "wavefronts" or "warps") on each compute unit/stream processor to hide memory access latency. Recent SPARC CPU's have 8-way hardware multithreading on the individual CPU core, GPU's can easily go even higher than that.
Yep, this also reflects the design target of GPUs targeting much larger working sets, so have higher main memory bandwidth and rely less on caches. CPUs rather have few fast threads of execution working on hot cached data than many slow ones talking to main memory (because N-way thread level parallelism often splits your cache N ways, to N working sets)
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).<p>PSHUFB wins in case of unpredictable access patterns. Though I don't remember how much it typically wins.<p>PMOVMSKB can replace several conditionals (up to 16 in SSE2 for byte operands) with only one, winning in terms of branch prediction.<p>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.<p>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.
The author has neglected the 3DNow! SIMD instructions from AMD.<p>They were notable for several reasons, although they are no longer included in modern silicon.<p><a href="https://en.wikipedia.org/wiki/3DNow" rel="nofollow">https://en.wikipedia.org/wiki/3DNow</a>!
Related: Go is looking to add SIMD intrinsics, which should provide a more elegant way to use SIMD instructions from Go code: <a href="https://go.dev/issue/73787" rel="nofollow">https://go.dev/issue/73787</a>
Wider SIMD would be useful, especially with AVX-512 style improvements. 1024 or even 2048 bits wide operations.<p>Of course memory bandwidth should increase proportionally otherwise the cores might have no data to process.
I wouldn't mind, but might need to increase cache line size on x86, as avx512 has reached the current size.
I would love to be able to fit small matrices (4x4 or 16x16 depending on precision) in SIMD registers together with intrinsics for matrix arithmetic.
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.<p>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!
This would start looking a lot like a GPU.
AMX registers are 1024 *bytes*
Recent and related:<p><i>Why do we even need SIMD instructions?</i> - <a href="https://news.ycombinator.com/item?id=44850991">https://news.ycombinator.com/item?id=44850991</a> - Aug 2025 (8 comments)
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.
While the blog post clearly highlights the core benefit of SIMD.<p>>Yet this new family of instructions was able to describe a great deal more work per instruction – the key benefit of SIMD.<p>It doesn't really explain why it is a benefit. From a ALU perspective, each SIMD lane could in theory execute a different operation. So why would simultaneous instructions as in SIMD ever be a thing? You could just keep cranking up the ILP (instruction level parallelism) by adding more ALUs, naturally extending to MIMD. Each SIMD lane would be equivalent to a full CPU core, so why is everyone stopping short of unlocking even more performance?<p>Because instruction memory (SRAM) is incredibly expensive.<p>By using SIMD you are reducing the number of instructions needed to describe repeating calculations. If you work with 32-bit floats and 512-bit instructions representing 16 lanes, you've made a 16-fold reduction to the required quantity of instruction memory and subsequent fetches from that memory to express the same parallel computation compared to using ILP.
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.<p>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.
And this is one of the biggest reasons why AVX-512 is so important. It doesn't just double the register width, it adds masking, and crucially suppresses faults from memory accesses when the corresponding element is masked. This makes it much more practical for compilers to automatically vectorize generated code.
ISPC is a great way to target this. C-like, generates the SIMD code, can be called from C/C++/etc.
Why does such an abbreviation still exist in 2025?<p>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.