Use case · scientific computing
Scientific computing GPUs: cluster-grade compute, no allocation committee
CFD, molecular dynamics, genomics, quant backtests: rent the same Hopper silicon national labs queue for — H200 with 4.8 TB/s of HBM3e at $3.058/hr, by the second, starting now instead of next quarter.
The HPC cards, ranked
| Tier | GPU | VRAM | On-demand | Interruptible | Why this card | Action |
|---|---|---|---|---|---|---|
| Good | A100 SXM4 | 80 GB | $0.583 | $0.291 | FP64 tensor cores at the lowest HBM price — the budget line for double precision. | Deploy |
| Better | H100 SXM | 80 GB | $1.587 | $0.793 | ~34 TFLOPS FP64, 3.35 TB/s — the general-purpose simulation default. | Deploy |
| Best | H200 | 141 GB | $3.058 | $1.529 | Same compute, 141 GB and 4.8 TB/s — memory-bound codes see the biggest jump. | Deploy |
Legacy budget: Tesla V100 at $0.089/hr still does honest FP64 for coursework and small models.
What runs here today
- Molecular dynamics — GROMACS, AMBER, LAMMPS with CUDA builds in the CUDA-dev template.
- CFD & FEA — AmgX-accelerated solvers, lattice-Boltzmann codes, in full VMs when licences demand them.
- Genomics — Parabricks-style pipelines: a 30× WGS alignment in under an hour on one H100.
- Quant research — Monte-Carlo sweeps and backtests that parallelise embarrassingly well on interruptible.
Licensing note: bring-your-own-licence software runs in VMs with hardware dongle passthrough available on reserved machines.
A simulation week, budgeted
| Parameter sweep | 40 runs × 3 h on A100 SXM4, interruptible | $34.92 |
|---|---|---|
| Hero run | 36 h on 8× H100 SXM, on-demand | $457 |
| Result volume 1 TB | trajectories + fields, one month | $80 |
The sweep on interruptible costs less than the storage — that is the shape per-second pricing gives research budgets.
HPC GPUs: FAQ
Which GPUs have real FP64 performance?
The datacenter line: H100/H200 (~34 TFLOPS FP64 tensor), A100 (~19.5 with tensor cores), V100 (~7.8). Consumer RTX cards throttle FP64 to 1/64th of FP32 — fine for ML, wrong for double-precision simulation. If your solver is FP64-bound, rent Hopper or Ampere datacenter parts.
Can I run MPI jobs across machines?
Yes — clusters connect 8× SXM nodes over InfiniBand with CUDA-aware MPI and NCCL/UCX preinstalled on request. Single-node 8× machines cover a surprising share of mid-size simulations by themselves.
Why H200 for memory-bound codes?
Bandwidth: 4.8 TB/s of HBM3e per GPU — stencil kernels, CFD and lattice codes that starve on GDDR parts often see near-linear speedups just from feeding the cores. 141 GB also keeps bigger domains resident.