Jalapeño inference chip emphasizes useful work per watt and latency

What changed. OpenAI reported 1.5-1.9x more AI work per watt at peak throughput and 1.7-3.6x lower end-to-end latency than selected commercial systems on InferenceX public-model workloads.

Technique / discovery. Model-compiler-serving-interconnect-hardware co-design evaluated at complete-workload level.

Why it matters. Sequential agent workloads make energy and latency per completed unit of work more relevant than peak arithmetic alone.

Application. High-volume inference, interactive agents, and long reasoning chains

Evidence maturity. Technical report (source tier A)

Identified bottleneck. Software, batching, quantization, memory placement, model shape, and power accounting can dominate vendor comparisons.

Caveat / evidence note. Vendor-reported, vendor-controlled comparison. Some power normalization uses published ratings rather than matched measured system power.

Review status. Reviewed. User requested: Yes.

• OpenAI (opens in a new tab)

A reproduction tutorial is not yet available for this entry. The closest reviewed material is Infrastructure, efficiency, and open ecosystems and Agent planning and cognitive architectures.