GenomeOcean-4B autoregressive generation throughput, power, energy efficiency and perplexity, BF16 versus FP8 (Yu et al. 2026 Table 4)
Autoregressive generation in vLLM on one H100 with three precision configurations.
Overview
An explanatory profile has not yet been reviewed for this record.
Consult the linked sources for architecture or protocol details. Missing evidence is not evidence of a missing capability.
3 recorded evaluations, 12 metric rows. A comparison chart has not yet been validated for these results. The table retains the individual findings and their sources.
Results
Results are available, but no reviewed comparison panel is linked in this release.
All evaluations
3 evaluations · 12 results. Different protocols are not a single leaderboard.
Protocol: GenomeOcean-4B autoregressive generation throughput, power, energy efficiency and perplexity, BF16 versus FP8 (Yu et al. 2026 Table 4) · Dataset: 96 GTDB Bacteria genome sequences bounded to 102,400 tokens (Yu et al. 2026)
Sorted by Energy efficiency (Autoregressive generation; generated tokens per second per watt) (higher is better). The best value in each column is highlighted. Decimals are rounded for display; each value links to the printed value and its source.
| Tested configuration | Energy efficiency (Autoregressive generation; generated tokens per second per watt) | Inference throughput (Autoregressive generation; generated tokens per second) | Perplexity (Sliding-window perplexity on 96 GTDB Bacteria sequences (10,240-token context, 2,560-token stride)) | Power draw (Autoregressive generation; mean steady-state GPU power from NVML) |
|---|---|---|---|---|
| GenomeOcean-4B, W8A8 plus FP8 KV cache, vLLM generation (Yu et al. 2026)Evaluation | 3.35 ± 0.02 (best in column) | 2169.04 ± 24.30 (best in column) | 128.73 ± 1.04 | 647.01 ± 8.55 (best in column) |
| GenomeOcean-4B, FP8 KV cache, vLLM generation (Yu et al. 2026)Evaluation | 3.14 ± 0.02 | 2085.89 ± 34.49 | 126.63 ± 1.04 | 663.77 ± 8.84 |
| GenomeOcean-4B, BF16, vLLM generation (Yu et al. 2026)Evaluation | 2.63 ± 0.01 | 1818.79 ± 7.95 | 125.97 ± 1.04 (best in column) | 692.40 ± 1.88 |
All 12 result rows with coverage, uncertainty and sources
Filter evaluations
Applied filters: All linked evaluations
Source checking is not independent reproduction. Release 2026-10-22-1b7cf027e4f6.
Methods and evaluation design
Procedure, tasks and evaluated configurations
Recorded evaluations
Each evaluation records what was tested and under which conditions.
Baseline coverage
Reference methods help show what a model adds beyond simple controls. We track a null control and a conventional method for each protocol.
0 of 2 active baseline roles have published Rewire measurements in this release. Measurements on a selected protocol do not establish coverage of an entire suite.
No execution recipe linked to this protocol. Recipe availability does not establish a completed evaluation.
- Author-reported evaluations
- 3
Literature evidence is not a Rewire measurement. Executed but unpublished runs and private review status are not included.
Null control
Proposed control: requires review
Training-set mean where supervised fitting is permitted
Protocol-specific applicability, permitted inputs, access, split, evaluator and execution requirements need review before implementation or execution.
This is a suggested selection rule, not a validated method or a measured score.
Conventional reference
Proposed control: requires review
Simple features with train-only ridge regression
Protocol-specific applicability, permitted inputs, access, split, evaluator and execution requirements need review before implementation or execution.
This is a suggested selection rule, not a validated method or a measured score.
Protocol coverage CSV (gzip) · Model evaluation matrix (gzip) · Source table (gzip) · Release and checksums (gzip)
Coverage is derived from release 2026-10-22-1b7cf027e4f6. Source citations describe the original records; they do not validate an unreviewed baseline proposal. No results have been generated by this audit.
Run instructions
No runnable recipe has been reviewed for this protocol. Dataset access, model requirements, licences and compute requirements must be checked against its sources before execution.
Strengths, limitations and unresolved questions
Evidence
Source checking verifies the cited claim or transcription. It does not establish independent reproduction.
Evidence table
Inspect claims, sources and review details
Trace each statement to its source and review. A context-only reference supports the record generally; it does not verify an individual field. Source checking does not reproduce an experiment.
One row per statement and cited source. Multiple citations are not independent evaluations. Shared locators are labelled explicitly.
0 evidence rows matching the loaded filters
| Property and statement | Original source and location | Review and provenance |
|---|
No evidence rows match these filters. Choose another scope or clear the search.
Sources and history
Release 2026-10-22-1b7cf027e4f6 · Record review: source checked
1 source record and release history
- FP8 Inference in Genomic Foundation Models: Theoretical vs. Realized Speedups on GenomeOcean · Original source · bioRxiv preprint version 1, posted 2026-08-14 (PDF footer); bioRxiv API date 2026-08-18
Technical metadata and extraction receipts
Stable ID: model-execution-20261011-protocol-yu2026-genomeocean-generation-fp8
- areas
- dna-genomes
- contexts
- research
- protocol
- Token-by-token generation with vLLM; throughput in generated tokens per second; GPU power from NVML at 100 ms intervals, efficiency as throughput per mean watt; sliding-window perplexity with 10,240-token context and 2,560-token stride on 96 GTDB sequences.
- version
- Table 4
- metric
- inference-throughput
- limitations
- Author-reported: two authors are at the DOE Joint Genome Institute, which develops GenomeOcean; no competing-interest statement is printed.; Microbial genomes (96 GTDB Bacteria sequences) and a perplexity check, not human diagnostic sequence; the paper says perplexity measures sequence-likelihood preservation, not the biological function of generated sequences (Section 4.4).; Throughput, power and efficiency are mean and standard deviation over five trials; perplexity standard deviation reflects sequence-level heterogeneity across 96 fixed sequences, not trial-to-trial variation.; Generation and embedding use different engines, so their absolute throughputs are not comparable (Section 2.4).; Peak memory for generation is not printed in Table 4.; One GPU type (H100 SXM5); the paper says results for other GPUs and engines need their own evaluation.
- random seeds
- 42-46
- source locator
- Sections 2.3-2.7; Table 4
Related records
- uses data: 96 GTDB Bacteria genome sequences bounded to 102,400 tokens (Yu et al. 2026)
- assessment: GenomeOcean-4B BF16 generation (Yu et al. 2026)
- assessment: GenomeOcean-4B FP8 KV-Cache generation (Yu et al. 2026)
- assessment: GenomeOcean-4B W8A8 + FP8 KV-Cache generation (Yu et al. 2026)
- assessed by: Select an execution workflow for large-scale diagnostic genomics