Compute Resources

Compute-Resource Application

For registered teams only. Resources are provisioned on a rolling basis — access details are emailed within 3–5 working days, so apply early. Each resource may only be used within its stated window.

Available resources

Google Cloud — Nvidia GPUs Open for applications

Apply your credit toward any GPU in Google Cloud's GPU regions & zones list. Each team has a US$1,000 maximum — estimate your cost with the Google Cloud Pricing Calculator before committing. Usage window Jul 15–20; apply as early as possible — grants are issued right before the window opens.

Which GPUs can I use?

Ready out of the box — no quota request needed:

  • V100 — up to 8 (any listed region)
  • T4 — up to 8
  • T4 VWS — up to 4
  • L4 — up to 16
  • L4 VWS — up to 16
  • A100 40GB — up to 16 (confirmed)

Tested, but blocked by default quota — you'd need a quota increase that may not clear within the 5-day window (not recommended):

  • A100 80GB — request the per-SKU quota NVIDIA-A100-80GB-GPUS-per-project-region
  • RTX PRO 6000 (compute) — request the family quota gpu_family=NVIDIA_RTX_PRO_6000
  • RTX PRO 6000 VWS — request the family quota gpu_family=NVIDIA_RTX_PRO_6000_VWS_GPUS
  • H100 (on-demand), H100 MEGA, B200, H200 — request the matching gpu_family quota; very likely needs Google Cloud account-team involvement

Any other GPU on the list may require a quota application that might not be approved — with only 5 days of usage, we don't recommend pursuing those.

Note for mainland-China-based teams: Google Cloud is generally unreachable from mainland China without a VPN (the Great Firewall blocks Google sign-in and API domains), so mainland-China-based teams should rely on the mainland-China-accessible options below.

Computational Freedom — RTX 4090 / L40S Open for applications

Choose one: 40 hrs Nvidia RTX 4090, or 25 hrs Nvidia L40S, with a ready-to-use environment. Mainland-China-based teams only. Usable during the simulation phase (see the schedule on the competition page).
Access requirements: the RTX 4090 servers are reachable only from a mainland-China IP address, and registration for both the 4090 and the L40S requires a mainland-China phone number.

AMD Instinct MI300X Open for applications

A shared 8× MI300X server, hosted in Taipei and accessible from mainland China. Each GPU carries 192 GB of HBM3 — its standout strength for memory-bound work, letting you fit models and batches that won't fit on a smaller card. As a rough, workload-dependent reference, an MI300X is broadly comparable to an Nvidia H200 on many tasks — stronger on memory-bound and FP32/FP64 work, behind on CUDA-optimized inference throughput and latency (MI300X vs H200 comparison). General training/dev compute on a ROCm environment. All teams. Usage Jul 1–31.
Rules: one GPU per job, a 12-hour maximum per job, one running job per team at a time (you may submit only when you have no queued or unfinished job), scheduled first-come, first-served.

Select the resource(s) you're applying for below. Computational Freedom is limited to mainland-China-based teams; eligibility is verified during provisioning.

Which resource(s) are you applying for?
Computational Freedom: choose one of the two (RTX 4090 or L40S), and only mainland-China-based teams are eligible.

We use your information only to organize the EBiM Benchmark. We don't sell or share it, and won't add you to any mailing list. Submissions are processed by Web3Forms.