Summary

  • Expanse's published peak, its allocation ceilings and a project's achieved performance answer different questions. None can stand in for the other two.
  • Public benchmarks and allocation rules do exist. Voyager's dated annual-report entry also records a move beyond its testbed phase, despite the older wording still visible in its guide.

A researcher can win an award of computing credits and still have no machine ready to use. ACCESS's getting-started instructions spell out this unglamorous moment: the newly awarded project can show “No active resources.” The next step is to exchange credits for time on a suitable resource, subject to that provider's approval.

That small gap between winning an award and starting work is a better introduction to Expanse than its biggest number. San Diego Supercomputer Center, or SDSC, operates the system at UC San Diego. Its user guide advertises 5.16 peak petaflops. A research team, though, needs to know which hardware it can use, how much time it will receive and how its own software performs there.

A ceiling is not a balance

The guide allows a principal investigator up to 15 million core-hours and 75,000 GPU-hours. A core-hour is one processor core used for an hour; a GPU-hour counts time on a graphics processor. Neither is a measure of how much useful work a program finishes.

Science Gateways applying through the large-project scheme Maximize may request up to 30 million core-hours. The important word is “request.” A team cannot put those hours into its schedule as though they have already been awarded.

The route to an award is set out on the ACCESS portal and in its allocation policies. Smaller and medium-scale projects receive credits, which they exchange for time on a chosen resource. Exchange rates depend on the resource. Larger Maximize requests are reviewed twice a year. Reviewers assess the science first; an award may then be reduced if there is not enough capacity to meet demand.

An exchange of credits is not an exchange of processing power. One CPU hour does not necessarily do the work of one GPU hour, and an award of CPU time does not automatically open the GPU machines. ACCESS advises allowing up to a week for exchange approval. Once approved, a computing job may still have to wait its turn. The guidance does not say how long that wait will be.

The original NSF award link identifies the funding record behind Expanse. Its public NSF data describes a system intended to support diverse, modest-scale research. The point is not simply to build a fast machine, but to make useful portions of it available to many projects.

There are measurements. Read the labels.

Expanse does have a TOP500 entry. For the Dell/AMD hardware it lists, the Linpack computing test achieved 2.48 petaflops against a theoretical peak of 3.30. Here there is a test result, not just a specification.

Why not divide 2.48 by the guide's 5.16 and call the answer efficiency? Because the two records do not establish that they cover the same hardware and task. The calculation would be easy; the comparison would not be sound.

A 2021 paper by the system's operators offers a more revealing detail. Across its CPU application tests, performance per core relative to the predecessor Comet ranged from 0.97 to 1.9 times as fast. The same new machine could be slightly slower on one tested case and markedly faster on another. These are historical operator-reported tests, not a forecast for a new model or today's H100 hardware.

The hardware labels matter too. The guide's technical table lists 728 standard AMD computing nodes and 52 V100 nodes. Its introduction says 54 GPU nodes. That discrepancy remains unresolved; neither count should be passed off as a freshly checked inventory.

The AI resource consists of 34 Dell XE9640 servers with four H100 GPUs each: 136 in all. They share Expanse's infrastructure, but researchers obtain time on them through the NAIRR Pilot. In the scheduling system, these resources appear as nairr-gpu and nairr-gpu-shared. Sharing infrastructure does not mean sharing access rights. Nor is the advertised 5.16 figure an H100 training result.

Voyager has moved on from the old label

SDSC's other AI system illustrates a different trap: a page heading can outlive the phase it describes. The Voyager guide still presents a three-year testbed beginning in early 2022, open to selected projects. It names Kubernetes as the software managing jobs and lists 42 training nodes with 336 Gaudi processors. Two further nodes carry 16 first-generation processors for inference, the work of using a trained model.

But SDSC's 2024–2025 annual report, page 16, dates the start of Voyager's two-year broader allocation phase to June 1, 2025, and directs applicants to ACCESS. It also mentions Gaudi1 and Gaudi2 hardware, so the older guide's count should not be mistaken for a newly verified complete inventory. The report establishes a documented transition, not a guarantee of spare capacity on demand.

For a team trying to finish a piece of research, the encouraging news is that there is more to consult than a peak figure: allocation procedures, named benchmark configurations and application tests. The remaining question is personal to the project, not mysterious about the machine. How much of the right resource will it receive, and what can its software finish with it?

Source clarification: the previously linked TOP500 record 179831 describes an Inspur system, not Expanse. It cannot support an SDSC performance comparison. Sources were checked on October 19, 2026; current queue times and an individual project's outcome were not tested.