Alongside our own AI work, we offer engineering services for NVIDIA DGX Spark and Grace-Blackwell (GB10) hardware — the platform where numbers from H100 and B200 simply don't transfer. We reproduce your exact problem on real GB10 silicon, find the root cause, and hand back a measured fix.
Every number below is labeled with what it actually measures — decode is what your users feel, prefill is prompt ingestion, aggregate is many streams at once. Vendors mix these on purpose. We never do: all figures are medians or n≥15 benchmarks at temperature 0, of tokens actually delivered to the client.
| workload | hardware | metric | result |
|---|---|---|---|
| ~1-trillion-parameter MoE (4-bit), 256k context | 8 nodes over an ordinary 100 Gbps switch | decode | 26–28 tok/s delivered, single stream (median 27.8) |
| 35B-A3B MoE, tuned from a 25.5 tok/s out-of-box baseline | one GB10 node | decode | 45.2 tok/s code · 41.2 dialogue |
| large sparse-MoE, after one collective-protocol fix | 8-node cluster | prefill | 699 → 1,520 tok/s · TTFT halved |
| same class, 16 concurrent streams | same cluster | aggregate | 170 tok/s · prefix caching up to 67× on repeated prompts |
For scale: the best-known public 8-Spark cluster result runs a 397B unquantized model at 24 tok/s on 200G-class networking. The first row above is a model roughly 2.5× that size, faster, on half the network — because the bottlenecks on this platform are fixable, and we've fixed them.
Narrow on purpose: the problems we solve on our own cluster, applied to yours.
vLLM, SGLang, TRT-LLM. The NCCL, interconnect, and topology issues that only appear at world size > 1 — and never show up in the single-GPU tutorials.
CUDA-graph capture, speculative decoding, quantization (nvfp4 / fp8 / w4a16), KV-cache and long-context tuning. Measured tokens/sec, with confidence intervals.
Weight-load stalls, memory leaks, silent misconfig, aarch64 kernel fallbacks, and benchmark artifacts that make a fast system look slow.
Almost every published GPU-inference playbook targets x86 + datacenter GPUs at single-node scale. Grace-Blackwell is ARM64, unified-memory, sm_121 — the numbers do not transfer, and we've measured exactly where they break. Our own multi-node GB10 cluster doubles as a reproduction lab, so we can stand up your configuration and test a fix the same day.
Low-risk to start. You never float our costs; we never waste your time.
One problem, one or two nodes. We reproduce it on real GB10 hardware and deliver written root-cause findings plus a recommended fix. Paid up front — and if we find nothing actionable, it's free.
The full workup for 4–8 node and production deployments: fabric, loading, memory, serving stack, and benchmark forensics — every layer measured, every finding written up with the recoverable headroom ranked. Same guarantee.
Hourly with an honest estimate, or a fixed price for a well-scoped target ("get model X to N tok/s"). Billed in advance, milestone by milestone. Ongoing retainer available for production teams.
Production cluster down or degraded and burning money? Immediate triage and priority scheduling — a direct line to the engineer, response measured in hours, not days. Minimum 4 hours.
Tell us the model, the config, and the number you're stuck at. If it's a fit, we start with the 1-day diagnostic.
Start here →Every engagement ships reproducible measurements — tokens/sec with confidence intervals and negative controls, always.