Mouchak TensorHive engine
Benchmarks · 2× NVIDIA DGX Spark · ConnectX-7
Every number has a receipt.
What TensorHive measured on two DGX Sparks, how we measured it, and what is still only a target. No marketing math.
Head to head
Where TensorHive stands against the engine we measure ourselves against, on the same two machines.
Rank-to-rank exchange latency
Median time for one 16 KiB exchange between the two Sparks, GPU in the loop. Lower is better.
TensorHiveBaseline engine
Show as table
| Engine | p50, 16 KiB | Transport |
|---|---|---|
| TensorHive | 6.0–6.5 µs | RDMA verbs |
| Baseline engine | 18.1–19.1 µs | RoCE all-gather |
Range covers both ranks and every rail policy tested. Microbenchmark, 22 Oct 2026.
Single-stream decode step
Time to produce one token for one user. Lower is better. The floor is the time just to read the active weights.
TensorHive targetBaseline, measuredWeight-read floor
Show as table
| Row | ms / step | Status |
|---|---|---|
| Weight-read floor | ≈21 | Computed |
| TensorHive | ≤24 | Target |
| Baseline engine | 25.44 | Measured, 16 Oct 2026 |
Hatched bar is a target, not a result. We'll publish the measured number when the benchmark run shows it.
Receipts
Every claim on this page comes from a run that wrote its own evidence file. Here is what each one measured.
Parity
Measured
16 / 16
replies match reference
GLM-5.3-Flash, 321B MoE, TP=2
2,032 decode steps · both ranks agree on every token · 30 Oct 2026
Network
Measured
~3×
lower exchange latency
16 KiB rank-to-rank exchange, p50
6.0–6.5 µs vs 18.1–19.1 µs · ConnectX-7 · 22 Oct 2026
Kernels
Measured
89%
of 273 GB/s peak
Rust GPU kernels on GB10
240–245 GB/s · bit-exact against a CPU reference · built with cuda-oxide
Decode
Next
≤24 ms
per token, one user
Beat the baseline's 25.44 ms
Within 1.14× of the ~21 ms weight-read floor
Questions about a number?
Ask us for the run behind it.