Solutions · Research & HPC
GPU servers for research and HPC.
Simulation, scientific computing and AI research on GPUs with strong FP64 and large HBM, by the month.
- 4× MI355Xbest fit
- 1,152 GBGPU memory
- $5,636per month
- From $5,636cheapest pick
What research and HPC needs from a GPU.
- 1
Double-precision throughput
Many scientific codes need FP64. Hopper (H100, H200) and the AMD MI355X keep strong FP64; the Blackwell B300 trades most of it for AI throughput.
- 2
Large, fast memory
Big grids and meshes live in GPU memory: 288 GB of HBM3E at 8 TB/s on the MI355X, 141 GB at 4.8 TB/s on the H200.
- 3
Predictable budgets
Grants and projects are budgeted by month. A flat price and no metering make the cost of a campaign known in advance.
Our picks for research and HPC.
The first is the best fit; the other two trade speed, memory or ecosystem for price.
-
Best fit
4× AMD MI355X
1,152 GB HBM3E · 8 TB/s per GPU
- Interconnect
- Infinity Fabric · 1.07 TB/s
- vCPU · RAM
- 128 · 1,536 GB
- NVMe
- 15.36 TB
$5,636/mo
$8,088 -
Alternative
4× NVIDIA H200
564 GB HBM3e · 4.8 TB/s per GPU
- Interconnect
- NVLink 4 · 900 GB/s
- vCPU · RAM
- 96 · 1,024 GB
- NVMe
- 15.36 TB
$9,116/mo
$13,036 -
Alternative
4× NVIDIA H100
320 GB HBM3 · 3.35 TB/s per GPU
- Interconnect
- NVLink 4 · 900 GB/s
- vCPU · RAM
- 80 · 800 GB
- NVMe
- 8 TB
$7,116/mo
$10,192
Tools people use, and what we recommend.
- CUDA / HIPNative GPU programming on NVIDIA and AMD.
- OpenMP and OpenACC offloadDirective-based GPU offload for C, C++ and Fortran.
- JAX / PyTorchScientific machine learning and differentiable simulation.
- GROMACS, LAMMPS, OpenFOAMExamples of widely used GPU-accelerated scientific codes.
Tips
- Confirm FP64 needs before choosing between H200, MI355X and Blackwell.
- Pin the driver and toolkit versions your code was validated with.
- Copy results off the server continuously; do not keep the only copy on a rented disk.
Good to know
Check which GPU architectures your code supports: CUDA-only codes need NVIDIA GPUs, HIP/ROCm and OpenMP-offload codes can use the MI355X.
Related guides
- NVIDIA drivers, CUDA and Fabric Manager on UbuntuInstall the driver from NVIDIA’s repository, add the CUDA toolkit only if you compile code, and bring up NVSwitch on 8-GPU HGX servers. Commands assume Ubuntu 22.04 or 24.04 LTS.
- ROCm on AMD Instinct MI355X: install and verifyMI355X (gfx950) needs ROCm 7.0 or later. Install the kernel driver and ROCm from AMD’s repositories, or keep ROCm inside containers and put only the driver on the host.
- Multi-GPU on one server: topology, NCCL and torchrunHow the GPUs in a server are wired decides how well a job scales. Check the topology, launch with torchrun and measure the interconnect before a long run.
Research & HPC questions.
Other workloads: LLM training, Fine-tuning, Inference & serving, Image & video, 3D rendering & VFX.
Which GPU for FP64 work?
The MI355X and the H100/H200 offer strong FP64. Blackwell B300 has little FP64 and suits AI rather than double-precision simulation.
Can I install my own compilers and libraries?
Yes. You have root access and install any toolchain: CUDA, ROCm, compilers, MPI, containers.
Can I pay a research project up front?
Each month is paid in advance. You can renew month after month for the length of your project.
Start with 4× MI355X.
AMD MI355X, 1,152 GB of GPU memory, for $5,636 a month.