{"id":"research-compute-resourcing","title":"Compute Resourcing for Atomistic Simulation and ML","subtitle":"Public allocation programs, cloud and spot pricing, and resourcing routes, compiled 2026-07-21.","category":"references","tags":["literature-review","savings-stack","allocations","cloud","funding"],"source":"articles/lit-review/research-compute-resourcing.md","lang":"en","words":1498,"readMinutes":7,"toc":[{"depth":2,"text":"1. Public allocation programs","id":"1-public-allocation-programs"},{"depth":3,"text":"US — NSF","id":"us-nsf"},{"depth":3,"text":"US — DOE (open to any researcher worldwide, no DOE funding required)","id":"us-doe-open-to-any-researcher-worldwide-no-doe-funding-required"},{"depth":3,"text":"Europe","id":"europe"},{"depth":3,"text":"UK — ARCHER2 — [access page](https://www.archer2.ac.uk/support-access/access.html)","id":"uk-archer2-access-page-https-www-archer2-ac-uk-support-access-access-html"},{"depth":3,"text":"China [partially uncertain — thin English-language primary sources]","id":"china-partially-uncertain-thin-english-language-primary-sources"},{"depth":2,"text":"2. Cloud research credit programs","id":"2-cloud-research-credit-programs"},{"depth":3,"text":"GPU-specialty clouds (verified list prices)","id":"gpu-specialty-clouds-verified-list-prices"},{"depth":2,"text":"3. Economics","id":"3-economics"},{"depth":2,"text":"4. Community / sharing models","id":"4-community-sharing-models"},{"depth":2,"text":"Suggested stack for a small atomistic-sim + ML team","id":"suggested-stack-for-a-small-atomistic-sim-ml-team"}],"html":"
\n

Provenance: explore agent agent-12 (director-commissioned deep research, backdrop research, 2026-07-21) — materialized verbatim. Quantitative claims are as reported by the research agent from sources it accessed; see citations inline. Citation-verification pass pending before any external publication.

\n
\n

Compute Resourcing for Atomistic Simulation + ML: Program Digest

Compiled 2026-07-21. Prices are volatile and region-dependent; all figures come from the cited pages — treat them as pointers, not quotes. Items I could not verify against a primary source are marked [uncertain].

\n

1. Public allocation programs

US — NSF

ACCESS (XSEDE successor; NSF awarded $52M over 5 years) — access-ci.org

\n\n

NAIRR Pilot — nairrpilot.org. US-based researchers/educators; 3-page application; distributes HPC time (Frontera, Expanse/Voyager, Anvil, DOE AI testbeds) and cloud credits from AWS/Google/Microsoft/NVIDIA partners. 375+ projects supported. ML-for-materials fits its focus areas.

\n

US — DOE (open to any researcher worldwide, no DOE funding required)

\n

Europe

EuroHPC JU (the PRACE successor for allocation purposes) — access policy

\n\n

UK — ARCHER2 — access page

\n

China [partially uncertain — thin English-language primary sources]

\n

2. Cloud research credit programs

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
ProgramOfferHow a small team qualifiesLink
AWS Cloud Credit for ResearchCredits for on-demand + spot EC2; students ≤$5,000; faculty/staff uncapped; 1-year termRolling application, 90–120 day review; academic affiliationaws.amazon.com
Google Cloud research creditsCredits expire 365 days after redemption; faculty, PhD students, postdocs at accredited institutions in approved countries; application requires a 250-word proposal + pricing-calculator estimate. Award size not published — historically ~$5,000 [uncertain]Online form, rollingedu.google.com
Azure Research Credits"Azure for Academic Research"; contact-form intake. Third-party listings cite up to ~60,000∗∗[uncertain;institutionaldealsvary—e.g.,StanfordHAIoffersupto60,000 **[uncertain; institutional deals vary — e.g., Stanford HAI offers up to60,000∗∗[uncertain;institutionaldealsvary—e.g.,StanfordHAIoffersupto50k]**Rollingmicrosoft.com
Oracle for Research12-month OCI credits (IaaS+PaaS, incl. HPC/GPU); amounts vary by proposal — UCL cites ~£50k average; a Fellows track adds cash awardsApplication form, reviewed on impact/feasibilityprogram PDF, UCL page
CloudBankNSF-funded billing/brokerage so NSF awardees can spend grant funds on AWS/GCP/Azure with negotiated discountsUS institution + NSF awardcloudbank.org
NAIRR PilotDistributes partner cloud credits (see §1)US-based, 3-page appnairrpilot.org

GPU-specialty clouds (verified list prices)

\n

3. Economics

Purchase models. Spot/preemptible: AWS Spot up to 90% off, GCP Spot VMs 60–91% off, Azure Spot up to 90% off (ProsperOps comparison). Verified example: GCP h3-standard-88 4.92–5.97/hron−demandvs4.92–5.97/hr on-demand vs4.92–5.97/hron−demandvs2.45/hr spot (devzero). Reserved/committed: ~37–57% off 1–3 yr (AWS RI), up to 60% off (CoreWeave), up to 50% off (Vast). Batch DFT/MD is checkpointable and embarrassingly parallel → spot-friendly; long MD trajectories need fault-tolerant restart logic.

\n

Cloud HPC instances, verified prices (on-demand, cheapest region):

\n\n

When cloud beats an allocation. Reference point: ARCHER2's notional £0.20/node-hour ≈ £0.0016/core-hour — allocations are 10–20× cheaper per core-hour than any cloud; ACCESS/EuroHPC/DOE are free outright. Cloud wins when: (a) the workload is bursty or small (<100k core-hours), (b) you can't wait 3–9 months of review cycles, (c) you have credits making marginal cost $0, (d) you need elasticity (thousands of short jobs now, not queued), or (e) GPU inference/fine-tuning where big-3 managed tooling matters. Steady, large MPI workloads belong on allocations.

\n

4. Community / sharing models

\n

Suggested stack for a small atomistic-sim + ML team

    \n
  1. Today: ACCESS Explore (400k credits, ~2 days) + AWS/Google/Oracle research credits for GPU work.
    2. \n
  2. Throughput: OSPool for screening ensembles; institutional cluster if available.
    3. \n
  3. Scale-up: ACCESS Discover→Accelerate; EuroHPC Development→Regular if EU-affiliated; DOE Director's Discretionary → ALCC for exascale ambitions.
    4. \n
  4. Burst/ML training: spot GPUs on CoreWeave/RunPod/Vast with checkpointing; reserved only for sustained GPU needs.
    5. \n
\n

Uncertain items to verify before publishing: Google/Azure credit award amounts, Lambda and Vast live prices, Azure HBv5 pricing, AWS hpc7g price, China scnet practical access, WCG submission process, EGI onboarding, AWS p5 current on-demand price.

\n"}