Shareable technical brief · July 2026
H100 prefill throughput and board-energy comparison.
Measured head-to-head on the same hardware, same shape. Version 100 internal measurement.
Summary
On a matched prefill-heavy workload, the LuxiEdge energy/throughput path is both faster and more energy-efficient (lower board joules per token position) than a current open serving stack (vLLM), while holding deterministic dual-run behavior on the Luxi path.
| Question | Answer |
|---|---|
| Faster? | Yes - about 1.17–1.18× higher throughput |
| Better or equal energy per token? | Yes - about 10–14% lower board J per position |
| Determinism held (Luxi TRADE path)? | Yes - dual-run score 1.0 |
Peer stack used greedy sampling (temperature = 0), not a separate
“deterministic product mode.” Token accounting is matched prefill positions.
Protocol (what was compared)
| Setting | Value |
|---|---|
| Model class | Qwen2-7B-Instruct (FP16) |
| Sequence length | 128 |
| Batch sizes | 16 and 32 |
| Workload | Prefill-heavy (one generated token per iteration; positions = iterations × batch × 128) |
| Hardware | Single NVIDIA H100 80GB-class GPU |
| Comparison | Sequential arms (not concurrent on the same GPU) |
| Metrics | Throughput (positions/s), board energy (NVML joules / position), dual-run determinism |
Peer stack: vLLM (current release in test), tensor parallel 1, prefix cache off, same token accounting.
Luxi path: production energy/throughput configuration (Flash-class attention control + device-resident multi-layer stack + FP16 weight residency). Not the high-fidelity AUDIT receipt lane.
Results
| Batch | System | Throughput (pos/s) | Board J/position | Determinism |
|---|---|---|---|---|
| 16 | LuxiEdge | ~41,700 | ~0.0171 | 1.0 |
| 16 | vLLM | ~35,800 | ~0.0190 | 1.0 |
| 32 | LuxiEdge | ~43,900 | ~0.0158 | 1.0 |
| 32 | vLLM | ~37,100 | ~0.0182 | 1.0 |
Ratios (Luxi / vLLM)
| Batch | Throughput ratio | Energy ratio (lower is better) |
|---|---|---|
| 16 | 1.17× | 0.90× |
| 32 | 1.18× | 0.86× |
Board power was measured under load via standard GPU power/energy interfaces (NVML). Energy is board joules, not facility wall-plug AC.
How to read this
- Throughput counts prompt positions under a fixed sequence length and batch (prefill-oriented), not chat tokens/s under a full OpenAI-compatible server.
- J/position is board energy per useful position under that same definition.
- Determinism is dual-run agreement on the measured path (same inputs → same reported behavior).
- This is an absolute, same-day head-to-head on one GPU class - not a claim against every serving recipe, every model, or every sequence length.
Why it matters
For operators, the useful question is rarely “who wins a microkernel.” It is:
- Do we move more useful work per second at commercial batch?
- Do we spend fewer joules per unit of that work?
- Can we still stand behind reproducible behavior?
Non-claims
- Not a full multi-tenant serving stack comparison (scheduling, continuous batching product surface, multi-GPU TP/PP).
- Not wall-plug or PUE.
- Not bit-exact identity between energy path and high-fidelity audit path (those remain separate product lanes).
- Not a claim of higher open-chat quality than the peer model implementation.
Contact
LuxiEdge · e@ewaller.com
For deeper diligence packs (methods, multi-run series, broader shapes), contact for a
controlled evaluation.