AlphaGenome pre-training and distillation compute
Training compute of AlphaGenome
Pre-training used TPU v3 with 8-way sequence parallelism (each 1 Mb interval split across 8 TPU cores); each gradient step of 64 samples required 512 TPUv3 cores, and a pre-training run of 15,000 steps typically completed in about 4 hours. Distillation ran without sequence parallelism on 64 NVIDIA H100 GPUs (batch size 64, one sample per GPU) for 250,000 steps, taking about 3 days.
Descriptive fact transcribed from the pinned source.
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.
7 evidence rows matching the loaded filters
| Property and statement | Original source and location | Review and provenance |
|---|---|---|
| Claimed field training_compute Context-only references | AlphaGenome Nature 2026 supplementary methods Supplementary Methods, 'Sequence Parallelism' (page 20), 'Pre-training' (page 21) and 'Distillation' / 'Model Training Parameters' (pages 21-22) Version: Supplement to Nature version of record, 28 January 2026; content hash pinned | not individually reviewed No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of exact publisher PDF bytes Format: original_pdf |
Limitations1 values
| AlphaGenome Nature 2026 supplementary methods Supplementary Methods, 'Sequence Parallelism' (page 20), 'Pre-training' (page 21) and 'Distillation' / 'Model Training Parameters' (pages 21-22) Version: Supplement to Nature version of record, 28 January 2026; content hash pinned | not individually reviewed No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of exact publisher PDF bytes Format: original_pdf |
| Source location Supplementary Methods, 'Sequence Parallelism' (page 20), 'Pre-training' (page 21) and 'Distillation' / 'Model Training Parameters' (pages 21-22) Context-only references | AlphaGenome Nature 2026 supplementary methods Supplementary Methods, 'Sequence Parallelism' (page 20), 'Pre-training' (page 21) and 'Distillation' / 'Model Training Parameters' (pages 21-22) Version: Supplement to Nature version of record, 28 January 2026; content hash pinned | not individually reviewed No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of exact publisher PDF bytes Format: original_pdf |
| Claimed value Pre-training used TPU v3 with 8-way sequence parallelism (each 1 Mb interval split across 8 TPU cores); each gradient step of 64 samples required 512 TPUv3 cores, and a pre-training run of 15,000 steps typically completed in about 4 hours. Distillation ran without sequence parallelism on 64 NVIDIA H100 GPUs (batch size 64, one sample per GPU) for 250,000 steps, taking about 3 days. Context-only references | AlphaGenome Nature 2026 supplementary methods Supplementary Methods, 'Sequence Parallelism' (page 20), 'Pre-training' (page 21) and 'Distillation' / 'Model Training Parameters' (pages 21-22) Version: Supplement to Nature version of record, 28 January 2026; content hash pinned | not individually reviewed No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of exact publisher PDF bytes Format: original_pdf |
| Description Descriptive fact transcribed from the pinned source. Context-only references | AlphaGenome Nature 2026 supplementary methods Supplementary Methods, 'Sequence Parallelism' (page 20), 'Pre-training' (page 21) and 'Distillation' / 'Model Training Parameters' (pages 21-22) Version: Supplement to Nature version of record, 28 January 2026; content hash pinned | not individually reviewed No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of exact publisher PDF bytes Format: original_pdf |
| Relationship: subject catalog-model-alphagenome Context-only references | AlphaGenome Nature 2026 supplementary methods Supplementary Methods, 'Sequence Parallelism' (page 20), 'Pre-training' (page 21) and 'Distillation' / 'Model Training Parameters' (pages 21-22) Version: Supplement to Nature version of record, 28 January 2026; content hash pinned | not individually reviewed No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of exact publisher PDF bytes Format: original_pdf |
| Name AlphaGenome pre-training and distillation compute Context-only references | AlphaGenome Nature 2026 supplementary methods Supplementary Methods, 'Sequence Parallelism' (page 20), 'Pre-training' (page 21) and 'Distillation' / 'Model Training Parameters' (pages 21-22) Version: Supplement to Nature version of record, 28 January 2026; content hash pinned | not individually reviewed No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of exact publisher PDF bytes Format: original_pdf |
Sources and history
Release 2026-11-02-a7f03c9e55a5 · Record review: source checked
1 source record and release history
- AlphaGenome Nature 2026 supplementary methods · Original source · Supplement to Nature version of record, 28 January 2026; content hash pinned
Technical metadata and extraction receipts
Stable ID: gfm-guide-20261014-claim-alphagenome-training-compute
- field
- training_compute
- value
- Pre-training used TPU v3 with 8-way sequence parallelism (each 1 Mb interval split across 8 TPU cores); each gradient step of 64 samples required 512 TPUv3 cores, and a pre-training run of 15,000 steps typically completed in about 4 hours. Distillation ran without sequence parallelism on 64 NVIDIA H100 GPUs (batch size 64, one sample per GPU) for 250,000 steps, taking about 3 days.
- source locator
- Supplementary Methods, 'Sequence Parallelism' (page 20), 'Pre-training' (page 21) and 'Distillation' / 'Model Training Parameters' (pages 21-22)
- review
- method: source-hash-verification; ai-assisted-source-review; method note: Each source sentence the claim rests on asserted verbatim in the reviewer's own text extraction (review/check_gfm_guide.py); the claim text was then read against them.; reviewer: claude; reviewer note: Independent Claude (Opus 5.5) review agent for PR #150; did not extract this batch; no human review claimed; date: 2026-10-14; artifact sha256: 86b2e6a07543e3c201e2e157a3e9c6eb5235ff13eb8a5224f6fcb6fe1808d4c0; retrieval url: https://media.springernature.com/original/springer-static/esm/art%3A10.1038%2Fs41586-025-10014-0/MediaObjects/41586_2025_10014_MOESM1_ESM.pdf; note: Claim matches the source.
- limitations
- Per-run figures; the number of pre-training runs (fold-specific and all-folds teachers, 64 all-folds models used in distillation) is described elsewhere and not totalled here.
Related records
- subject: AlphaGenome