Mistral Small 4 119B-A6.5B on Dell Precision 7960 Rack (2x RTX PRO 6000, 192 GB)

1. Multi-GPU: 1792 GB/s per card, with each additional card contributing 40% of its bandwidth — 2509 GB/s effective, not the 3584 GB/s aggregate. Cross-GPU collectives use PCIe rather than a dedicated GPU fabric.
2. Decode: reading 6.5B active parameters at 8.50 bits/weight takes 3.82 ms at 2509 GB/s x 72% achieved efficiency.
3. MoE routing penalty of 15% applied: expert gathers are less bandwidth-efficient than a dense sweep.
4. Per-token overhead of 1.4 ms (kernel launches, attention bookkeeping, sampling) is significant here — this model is not purely bandwidth-bound on this hardware.
5. Prefill: 380 TFLOPS (FP16) x 2 for native FP8 tensor cores x 26% assumed model-FLOPs utilisation, divided by 2 x 27.9B parameters per token.
6. Prefill uses 27.9B effective parameters, not the 6.5B active in decode: a batch of hundreds of tokens routes across most of the expert pool.

• MoE decode depends on how well the engine batches expert gathers; real results vary more than for dense models.
• With few active parameters on fast memory, fixed per-token overhead rather than bandwidth sets the ceiling. Real engines vary widely in how well they hide it.
• Prefill throughput is highly engine-dependent. Flash attention, batch size and quantized KV all move this number substantially, and for sparse mixture-of-experts models it is the least reliable figure we produce.
• This is a calculated estimate, not a measurement. It assumes a single request, a short prompt, no speculative decoding and a warm model already resident in memory.

Estimator version estimator_v1. Stored with every estimated row so old estimates can be regenerated when the model improves.