cache
KV and feature serving
Representations, metadata, norms, and GPU handles are served from L0/L1/L2 storage and GPU-resident structures.
HX-SDP / flagship product
verified surface
HX-SDP
verified surface
HX-Provenance
verified surface
The Atlas
Capabilities
cache
retrieval
stream
Evidence scope
| Published proof | 100M H100 | 100% recall · 20.4 ms p50 · signed receipt | verified |
|---|---|---|---|
| fp32 ceiling | 500M | 100% recall · 76.53 ms p50 | verified |
| fp16 scale | 1B | 96-100% recall · 38.51 ms p50 | product |
| B200 scale | 2B | 98% recall · 60.89 ms p50 | product |
Buyer fit
Bounded technical evaluations, private deployments, proof-pack diligence, and workloads where duplicate indexes or GPU memory pressure materially affect cost or correctness.
Ultra-high QPS single-GPU serving, open public cloud self-service, and compressed-path deployment certification should be scoped explicitly before adoption.
Access