5 ms·
Very interesting. It sounds like tuning at the PTX level can increase workload efficiencies, such as quote "Specifically, we employ customized PTX (Parallel Thr
by the_panopticon 1y ago
Very interesting. It sounds like tuning at the PTX level can increase workload efficiencies, such as quote "Specifically, we employ customized PTX (Parallel Thread Execution) instructions" from the DeepSeek folks https://arxiv.org/abs/2412.19437 https://arxiv.org/abs/2412.19437.
- shetaye 1y agoAgreed! The gulf between pure-C++ CUDA and PTX is getting larger with these optimizations. My understanding is that Deepseek used PTX instructions that either had no corresponding C++ implemented (like `wgmma` mentioned in the article) or uncommon permutations of modifiers (`LD.Global.NC.L1::no_allocate.L2::256b`).
- saagarjha 1y agoThey didn’t employ custom PTX instructions; they used existing ones in ways they were not designed to be used.