4 ms·
I'm interested to know where people use AVX-512. Off the top of my head, I can think of machine learning and crypto. Machine learning---that is, matrix/tensor s
by 764_OC 7y ago
I'm interested to know where people use AVX-512. Off the top of my head, I can think of machine learning and crypto. Machine learning---that is, matrix/tensor stuff---seems more like a GPU workload. That leaves crypto, but most crypto (eg ChaCha or Keccak) only make use of 128 or 256-bit extensions (in fact, they seem to designed with this constraint in mind.)
- brobinson 7y agoVideo encoding comes to mind.
- arsome 7y agoDepends on your use case but the most recent GPU encoders are pretty solid actually. Used to be that they were awful for quality, but more recently they seem to have vastly improved.
- BubRoss 7y agoThe person above was asking what people use AVX512 for and video encoding is a big use case, better GPU encoders don't change that.
- arsome 7y agoThe person above also pointed out that things like machine learning are often better done on a GPU anyways. Certainly if you offload your video encoding to your GPU it does change the value proposition of that entirely.
- BubRoss 7y agoI'm not sure what you are trying to say exactly but it seems like you are making some sort of prediction of the future. Video encoding, especially newer codecs like HEVC and VP9 are CPU intensive and SIMD optimized tasks that people are doing right now.
- arsome 7y agoI'm not referring to any sort of future prediction, those people can switch to GPU encoding of video right now and enjoy better performance. https://developer.nvidia.com/video-encode-decode-gpu-support-matrix https://developer.nvidia.com/video-encode-decode-gpu-support...
- BubRoss 7y agoThey asked what people are using it for right now and video encoding is one of those things. This is very direct. They didn't ask about solutions or what they should do about a particular problem, just about what people are currently doing.
- unlinked_dll 7y agoThere are diminishing returns when evaluating recursive formulas using SIMD, as the cost to gather/scatter into the vectors, duplication of computation, and transfer overhead to the GPU cost more than doing it in a single core with vector instructions. Real-time audio processing is probably the most common case where you see huge gains in SIMD on the CPU but very minimal gains on the GPU. Other examples are certain polynomial approximations and physics simulations. Also string manipulation, parsing, evaluating a regex. The overhead of getting to the GPU outweighs keeping it on the CPU core.