multi
8×H100 SXM (NVSwitch, NVLink4 · ~900 GB/s/GPU)● graded on busbw (NVLink bus-bandwidth efficiency)
The multi-GPU sibling of hard. A coding agent takes a PyTorch + NCCL reference for a distributed op and rewrites it as a fast, fine-grained NVLink kernel (CUDA / Triton / NVSHMEM / CUDA symmetric memory / ParallelKittens) that beats the NCCL baseline. The graded quantity is busbw — achieved NVLink bus bandwidth ÷ NVLink peak, never TFLOPS — so every problem is deliberately communication-dominated. Six-problem deck, all busbw-graded, each forbidding its bare collective. No runs yet.
showing 6 of 6 rows
| files | conversation | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| — | — | AllReduce + Residual | - | no run | no run | clean | - | - | referencesolution | trace0/0 awaiting first run | |
| — | — | ReduceScatter + RMSNorm | - | no run | no run | clean | - | - | referencesolution | trace0/0 awaiting first run | |
| — | — | AllGather + fp8 Dequant | - | no run | no run | clean | - | - | referencesolution | trace0/0 awaiting first run | |
| — | — | MoE All-to-All | - | no run | no run | clean | - | - | referencesolution | trace0/0 awaiting first run | |
| — | — | Ulysses All-to-All | - | no run | no run | clean | - | - | referencesolution | trace0/0 awaiting first run | |
| — | — | fp8 ReduceScatter Grad | - | no run | no run | clean | - | - | referencesolution | trace0/0 awaiting first run |
Methodology and the full problem deck live in the spec. Results will populate here as sweeps complete on the 8×H100 node.