Mouchak TensorHive engine
How do we measure TensorHive?
Every number on mouchak.ai comes from a run on our two DGX Sparks that wrote its own evidence file. We compare outputs against a reference run, report medians with the hardware, link and date, and label anything we have not measured yet as a target.
Last updated · by Mouchak
What we measured
| Measurement | What | How | Date | Result |
|---|---|---|---|---|
| Output parity | GLM-5.3-Flash, TP=2 | 2,032 decode steps; ranks agree on every token; 16 replies compared with a reference capture | 30 Oct 2026 | 16 / 16 match |
| Exchange latency | 16 KiB rank-to-rank exchange, GPU in the loop | Median (p50), both ranks, every rail policy tested; baseline is the other engine’s all-gather on the same machines | 22 Oct 2026 | 6.0–6.5 µs vs 18.1–19.1 µs |
| Kernel bandwidth | Rust (cuda-oxide) memory-bound kernels on GB10 | Effective bandwidth against the 273 GB/s peak; outputs bit-exact against a CPU reference | Oct 2026 | 240–245 GB/s (~89%) |
| Decode step | One user, one token | Target only. Floor computed from bytes read; baseline measured 25.44 ms on 16 Oct 2026 | Not yet | Target ≤ 24 ms |
Our rules
- No number without a receipt: a run that wrote its own evidence file.
- Comparisons use the same two machines and the same link.
- Every number carries a date. Targets are marked as targets, never as results.
- Correctness first: a speed-up only counts if outputs still match the reference.
Hardware
Two NVIDIA DGX Spark systems (GB10 Grace Blackwell, 128 GB unified memory, 273 GB/s each), connected directly by a 200 Gb/s ConnectX-7 cable.