Skip to content

Performance Evidence

Appliance release binding: v0.1.0-enterprise-ready · reference AMI ami-010806d4d3445660e · 2026-05-19
Evidence reconciliation: 2026-09-05, from the April 2026 benchmark records identified below. This documentation update is not a new benchmark run.

Results are specific to the workload, representation, precision, hardware, and test configuration. The synthetic scale campaign and the named embedding-corpus campaign measure different operating points. A capacity ceiling from one cannot be combined with latency or recall from the other.

Measured synthetic scale envelope

These are measured latent-scan runs on generated low-rank data, with D=384, intrinsic dimension/rank 32, and no SQ8 rerank sidecar. They are not named production embedding corpora. B200's two-billion-row result is a measurement, not a sizing extrapolation.

Representation Hardware Rows at this operating point Measured R@10 Query p50 Query p99 Evidence
fp32 latent-only H100 80 GB 500,000,000 1.000 76.53 ms 77.68 ms Master §6.1, §7.2; H100 fp32 ceiling
fp16 latent-only H100 80 GB 1,000,000,000 1.000 38.51 ms 41.02 ms Master §6.1, §7.3; H100 fp16 ceiling
fp16 latent-only B200 192 GB 2,000,000,000 0.980 60.89 ms 62.58 ms Master §6.2, §7.3; B200 2B record

The H100 fp16 ladder includes lower recall at other scales, including 0.96 at 400M rows. The 1.000 value above describes the tested 1B operating point; it is not an exact-recall guarantee for arbitrary data. The B200 record is dated April 5, 2026 and records a 132 GB serving footprint and 142 GB query VRAM peak.

Measured retrieval on real embedding corpora

The following rows are the canonical hardened H100 80 GB HBM3 reference from April 15, 2026. Configuration: DeltaMainIndex, target_quality=0.95, rank_cap=4096, GPU-resident SQ8, 200 held-out queries, cosine ground truth, K=10, and rerank_k=100; PyTorch 2.11.0 and CUDA 12.8. RR denotes the reported result after the retrieval policy's rerank stage. Class-dependent shortcuts can skip a separate rescore.

Compression is the total index footprint relative to dense fp32 for the same corpus, including GPU-resident SQ8. It is not a latent-only or GPU-only ratio. These corpus sizes are the measured row counts, not maximum capacities.

Model / corpus Dimensions Corpus rows Rank Compression vs fp32 RR R@10 RR p50
CLIP B/32 (AG) 512 99,800 275 1.3× 0.982 1.16 ms
CLIP B/32 (CIFAR) 512 119,800 95 2.9× 0.964 1.19 ms
CLIP L/14 (AG) 768 99,800 390 1.3× 0.989 1.16 ms
CLIP L/14 (CIFAR) 768 119,800 153 2.8× 0.976 1.22 ms
Gemini-001 768 99,800 498 1.1× 0.996 1.19 ms
MPNet 768 127,400 310 2.2× 0.968 1.32 ms
Cohere embed-v3 1024 499,800 418 2.2× 0.994 2.10 ms
OpenCLIP ViT-H/14 1024 99,800 401 2.2× 0.995 1.31 ms
Snowflake Arctic-L v2 1024 99,800 337 2.4× 0.993 1.26 ms
ResNet-50 2048 119,800 138 3.5× 0.987 1.22 ms
Gemini-001 3072d 3072 99,800 667 2.8× 0.998 1.42 ms
OpenAI 3-large 3072 99,800 1,261 2.2× 0.966 1.70 ms
E5-Mistral-7B 4096 99,800 1,146 2.5× 0.934 1.63 ms

Source: Master Reference §9.1. The master table takes precedence over older synthesis and rank-cap tables. A separate B200 two-tier result at 2B rows and 4 ms is not established by these records and is not presented as a measurement or extrapolation here. Qualification of another hardware, corpus, or serving policy requires its own result.

Exact retrieval and fidelity

The EXACT API (POST /v1/query/exact) bypasses compressed approximation and evaluates against the registered raw corpus, returning a signed receipt under its stated numerical and corpus contract. A compressed-path test that happens to attain R@10 = 1.000 does not establish that exact-path contract for other workloads.

QTT is an alternate supported representation path, described in Concepts. The scale and embedding tables above do not measure QTT serving.

Operational measurements

  • Cold start: the recorded S-100M run reports 114.37 s total recovery and R@10 = 1.000. This is a historical engineering result, not a startup SLA. Source: Master §10.2 and the Phase 1/3 artifact listed below.
  • Completed 24-hour soak: the S-10M, H100, fp64 campaign recorded 2,934,201 queries, 8,640,686 inserts, three query errors and six insert errors. Five of seven gates passed; flush-boundary errors and 17.22% VRAM drift failed their gates. The master attributes the errors to a transient buffer-swap race and the drift to allocator fragmentation. This records that campaign's outcome, not the state of every later release. Source: Master §11.
  • The earlier synthesis's 0.39 errors per million operations refers to its 10.5-hour interim checkpoint. It must not be presented as the completed soak result or as an availability SLA.

Evidence references

The source documents and underlying run artifacts are provided through the agreed evaluation channel. Request the HX-SDP performance evidence packet. Identify the table row, hardware and intended workload. Access is subject to the applicable evaluation scope and agreement.

Source document Version / declared date Relevant sections
HX_SDP_MASTER_REFERENCE.md Header dated April 20, 2026; snapshot SHA-256 below §1 representation; §6/§7 synthetic scale; §9.1 hardened corpus rows; §10.2 recovery; §11 soak; §18 artifact index
EXECUTIVE_SYNTHESIS.md Version 2.1, April 20, 2026; snapshot SHA-256 below §1 synthetic scale; §5 interim/final soak distinction; §8 two-tier corpus results
HX_SDP_ENTERPRISE_BENCHMARK_VV.md April 2026 campaign record; snapshot SHA-256 below §0.3c B200 measured run; detailed run methodology and validation

The files include later annotations, so the hashes identify the exact reference snapshots used for this page, independently of their header dates. The hashes are document identifiers; they are not substitutes for signed run evidence or an independently trusted issuer key.

Reference snapshot SHA-256
HX_SDP_MASTER_REFERENCE.md c998eca17f5ca11962378f54f90660b401187b4a79eb7f957731f9b5503bc39a
EXECUTIVE_SYNTHESIS.md f15d06aeaac4b34bb4af78ca699e15c3b6e784900bc4b4d062bc770acb2538a5
HX_SDP_ENTERPRISE_BENCHMARK_VV.md c0de1ea6f16276159ff3124664c3bd5fdc688164430c86728ddce82d20424ddc
Run evidence identifier in the master index Supports
artifacts/h100_fp32_ceiling.json H100 300M–500M fp32 synthetic ceiling
artifacts/h100_fp16_ceiling.json H100 400M–1B fp16 synthetic ceiling
artifacts/b200_2b_fp16.json B200 2B fp16 synthetic run
Master §9.1 hardened H100 table and its matching corpus/run records Named real-corpus configuration, retrieval and compression
artifacts/phase1_phase3_20260402/h100_phase1_phase3.json Cold-start campaign
artifacts/remaining_20260402/h100_soak_24h.json Completed soak campaign

These identifiers name the evidence to request; they are not public download links. Release and issuer verification requirements are described in Release Verification.

Standing caveat

These are engineering measurements on the named hardware. The v0.1.0-enterprise-ready appliance release carries no contractual recall, latency, or throughput claim (see Known Boundaries). Customers validate representative workloads; hx-sdp validate and hx-sdp sizing are shipped for that purpose. Current ordering is through Private Appliance or a scoped pilot.

Results outside this publication

Broad max-VRAM latency claims, sustained availability as an SLA, multi-tenant resource quotas, and unmeasured hardware or corpus combinations are excluded. The published B200 synthetic row remains within scope. Hardware-sizing estimates belong in Hardware and Sizing and must retain their extrapolation label.