# Below CUDA: GPU Kernels, PTX and Streaming Multiprocessors The article's DeepSeek example uses three layers of GPU vocabulary. From the top down: ## CUDA Nvidia's programming platform for its GPUs: a language extension, compiler and a large set of libraries (cuBLAS for maths, cuDNN for neural networks, NCCL for GPU-to-GPU communication). Nearly all AI software runs on CUDA through frameworks like PyTorch. Most labs never need to look underneath it, and CUDA's ecosystem is a large part of Nvidia's moat. ## GPU kernels In GPU work a "kernel" is not the operating-system kernel described in [[Kernels]]. It is a small function that runs in parallel across the GPU's cores: one matrix multiply, one attention step, one all-to-all send. Writing a custom kernel means replacing a generic library routine with one shaped to your exact model and hardware. Much of the efficiency gap between labs lives here. ## PTX "PTX" refers to NVPTX, the backend in the LLVM compiler toolchain that generates code for Nvidia GPUs. Working at this level is closer to assembly than to ordinary CUDA: more control over exactly what runs on the chip, at the cost of far more engineering effort. ## Streaming multiprocessors Nvidia's name for the groups of parallel processing cores on a GPU that actually execute kernels. Dedicating some of them to a single job (here, handling communication) is a hardware-level trade: less capacity for maths, but no waiting on a slow interconnect. > [!warning] Sourcing > The "PTX beneath CUDA" and "20 of 132 streaming multiprocessors per GPU" details come from a Mirae Asset Securities analysis reported by Tom's Hardware in January 2025, not from DeepSeek's own V3 technical report. The report itself describes custom cross-node all-to-all communication kernels over NVLink and InfiniBand, computation–communication overlap, and FP8 training. Worth citing accordingly. ## Why it matters Going below the abstraction layer is how a lab turns constrained hardware ([[Nvidia H800]]) into competitive training runs. It is the clearest example of [[Systems and Compiler Depth]]. Related: [[DeepSeek Technical Differentiators]], [[NVLink]], [[InfiniBand]], [[OpenCL]]