# GPU servers for research and HPC, on fixed monthly budgets

> GPU servers for research and HPC: H100, H200 and MI355X with vendor FP64 figures, CUDA or ROCm, on fixed monthly budgets. Match the GPU to your code.

Source: https://overgpu.com/solutions/research-hpc  
Updated: 2026-10-23  
Publisher: OverGPU (https://overgpu.com/)

## Key points

- **FP64 where it matters:** H100 and H200 for CUDA codes, MI355X for ROCm, chosen by vendor-published double-precision figures.
- **High memory bandwidth:** Up to 8 TB/s per GPU on the MI355X and 4.8 TB/s on the H200, per vendor specifications.
- **Bare-metal 8-GPU nodes:** 2 TB of DDR5 on HGX H100 and H200 nodes, 3 TB on the MI355X node, plus eight NVMe drives each.
- **Fixed monthly budget:** One flat price per server, billed in advance, with no egress fees within the included bandwidth.
- **CUDA or ROCm images:** Ubuntu with NVIDIA drivers and CUDA, or ROCm for AMD, with Docker preinstalled. Debian or Rocky Linux on request.
- **Private networking:** Link several of your servers on a private network for multi-node jobs, on request.

## Recommended configurations

- [8× NVIDIA HGX H100](https://overgpu.com/gpus/nvidia-h100): $16,900 per month — 8× H100 on NVLink for MPI and multi-GPU simulations
- [8× AMD Instinct MI355X](https://overgpu.com/gpus/amd-mi355x): $19,900 per month — 288 GB per GPU and 78.6 TFLOPS FP64 vector, AMD-published
- [NVIDIA H100 SXM](https://overgpu.com/gpus/nvidia-h100): $2,190 per month — One H100 for porting, prototyping and single-GPU codes

## Overview

GPU servers for research and HPC have to match the numerics of your code before anything else. A molecular dynamics package that runs in mixed precision, a CFD solver that needs FP64 throughout and a protein model that runs in BF16 want different hardware. We rent NVIDIA H100 and H200 servers with CUDA and AMD Instinct MI355X servers with ROCm by the month, so a project gets dedicated GPUs at a price fixed before it starts.

## FP64: where it matters and which GPU to pick

Double precision matters for codes where rounding errors accumulate: electronic-structure and quantum chemistry packages, many CFD and finite-element solvers, and some climate and astrophysics models. Many molecular dynamics engines, such as GROMACS and AMBER, run their GPU paths in mixed precision, and AI-for-science models run in BF16 or FP8. For those, FP64 throughput is not the deciding figure.

| GPU | FP64, vendor-published | Memory | Bandwidth | Stack |
| --- | --- | --- | --- | --- |
| H100 SXM | 34 TFLOPS, 67 TFLOPS on FP64 Tensor Cores | 80 GB HBM3 | 3.35 TB/s | CUDA |
| H200 | 34 TFLOPS, 67 TFLOPS on FP64 Tensor Cores | 141 GB HBM3e | 4.8 TB/s | CUDA |
| MI355X | 78.6 TFLOPS vector | 288 GB HBM3e | 8 TB/s | ROCm |

NVIDIA lists the H100 and H200 figures, AMD the MI355X figure. FP64 Tensor Core throughput applies to matrix operations, such as DGEMM-heavy dense linear algebra through cuBLAS; the vector figure applies to everything else. The RTX PRO 6000, RTX 5090 and L40S are designed for graphics, AI and single-precision work, so we do not recommend them for FP64-bound codes. For Blackwell data-center GPUs, check NVIDIA's current FP64 specifications for your exact part before choosing one for double-precision work.

## Choosing GPU servers for research and HPC

Many HPC codes, including stencils, sparse solvers and lattice methods, are limited by memory bandwidth rather than peak FLOPS, so the 4.8 TB/s of the H200 or the 8 TB/s of the MI355X can matter more than any compute figure. Memory per GPU sets the largest domain you can hold without decomposing it further. A double-precision 3D grid of 1,000 × 1,000 × 1,000 cells with 10 variables per cell takes about 80 GB: that fills an H100 but fits on an H200 or an MI355X with room for solver workspace.

- **One H100:** port, profile and validate a code, or run single-GPU jobs.
- **HGX H100 or HGX H200:** eight GPUs on NVLink at 900 GB/s per GPU for halo exchanges and collective operations, eight 400G ConnectX-7 adapters for multi-node runs, and 2 TB of DDR5.
- **8× MI355X:** 2.3 TB of HBM3e in one node, 3 TB of DDR5 and eight 400G RoCE adapters, for large memory-bound problems on ROCm. See the [AMD Instinct MI355X page](https://overgpu.com/gpus/amd-mi355x).

For MPI across nodes, ask for [private networking](https://overgpu.com/products/private-networking) between your servers in the order notes.

### Profile before you scale

Before booking an 8-GPU node for months, profile on one GPU. NVIDIA Nsight Systems and Nsight Compute, or rocprof on AMD, show whether a kernel is bound by memory bandwidth or by compute, which tells you whether the H200's extra bandwidth or the MI355X's FP64 figure will pay off. Then measure strong scaling from one to two, four and eight GPUs: if efficiency drops sharply past four, run two four-GPU jobs side by side on the node, or rent single GPUs instead. Plan storage the same way. Checkpoint-restart files and output fields add up fast; the HGX and MI355X nodes each carry eight 3.84 TB NVMe drives, about 30 TB raw, for scratch space, with 50 TB of outbound traffic a month to bring results home.

## CUDA, ROCm and your software stack

NVIDIA servers start from Ubuntu 22.04 or 24.04 with drivers, CUDA, Docker and the NVIDIA Container Toolkit. Add the NVIDIA HPC SDK, a GPU-aware MPI such as Open MPI with UCX, or your group's containers. If your home cluster uses Apptainer, install it so the same images run in both places. Setup notes are in our [NVIDIA drivers and CUDA guide](https://overgpu.com/docs/nvidia-drivers-cuda).

AMD servers use ROCm images. CUDA code needs a HIP port, which the hipify tools help automate, or a portable backend such as Kokkos, SYCL or OpenMP offload. PyTorch and JAX have ROCm builds. Check that your code and its version support ROCm before choosing the MI355X, and start from our [ROCm guide](https://overgpu.com/docs/amd-rocm). Debian or Rocky Linux images are available on request.

## Fixed monthly budgets for research projects

Each server has one flat price, billed monthly in advance and fixed for each paid month, and we give at least 30 days' notice before any price change for future months. A three-month simulation campaign on an HGX H100 node costs 3 × $16,900 = $50,700, with 50 TB of outbound traffic per month included and inbound traffic not counted. A single H100 is $2,190 a month, and the 8× MI355X node is $19,900.

A dedicated node removes queue waits when shared clusters fill up before a deadline, and gives you root access for custom stacks. For short bursts of a few days a month, hourly capacity elsewhere will cost less. Once you have an active server, you can ask by ticket for term pricing for 6–12 months or several servers, usually from three. Labs that work with grants and procurement offices should also read our page on [GPU servers for universities and research labs](https://overgpu.com/solutions/universities).

## Getting started

[Deploy a server](https://overgpu.com/deploy) from one form: choose the configuration, set up the OS image, optional preinstalled software and your login (root password or SSH key), create a free account or log in, and pay the first month from your prepaid balance, topped up in crypto. Your code, precision, problem size and node count can go in the optional notes for our engineers. The server is delivered in under 10 minutes after payment, with the IP and SSH details in your dashboard. Our engineers provide 24/7 support for incidents through tickets in your dashboard.

## Frequently asked questions

### How do the H100, H200 and MI355X compare on FP64?

AMD lists 78.6 TFLOPS FP64 vector for the MI355X. NVIDIA lists 34 TFLOPS FP64, and 67 TFLOPS on FP64 Tensor Cores, for both the H100 SXM and the H200; the H200 adds memory and bandwidth, not FP64 compute.

### Does my CUDA code run on the MI355X?

Not directly: it needs a HIP port or a portable backend such as Kokkos, SYCL or OpenMP offload. Frameworks such as PyTorch and JAX already have ROCm builds, so check your exact versions before ordering.

### Can I run MPI across several servers?

Yes. Ask for private networking between your servers in the order notes when you deploy them.

### Can I use Rocky Linux instead of Ubuntu?

Yes, Debian or Rocky Linux images are available on request. Our standard images are Ubuntu 22.04 and 24.04 with NVIDIA drivers and CUDA, and ROCm images for AMD servers.

## Related pages

- [GPU servers for universities and research labs, on fixed budgets](https://overgpu.com/solutions/universities): Fixed monthly prices for grants and semesters, invoices for procurement and term pricing. Plan a semester.
- [Rent an 8-GPU AMD MI355X server for $19,900 per month](https://overgpu.com/gpus/amd-mi355x): A bare-metal 8x AMD Instinct MI355X server (288 GB HBM3e per GPU) for $19,900/month, about $3.41 per GPU-hour. ROCm ready, paid in crypto.
- [Rent NVIDIA H100 SXM servers from $2,190 per month](https://overgpu.com/gpus/nvidia-h100): Rent a dedicated NVIDIA H100 SXM for $2,190/month (about $3.00 per GPU-hour at 24/7) or an 8-GPU HGX H100 for $16,900. Flat price, delivered in under 10 minutes.
- [MI355X vs H200: double the memory or the proven stack?](https://overgpu.com/compare/mi355x-vs-h200): Twice the memory per GPU and native FP4 against the mature Hopper and CUDA stack. Compare what fits, interconnect and monthly cost.
- [ROCm setup for AMD Instinct MI355X](https://overgpu.com/docs/amd-rocm): Check GPUs with rocm-smi and amd-smi, set user groups, run PyTorch and vLLM in ROCm containers and test RCCL.
- [NCCL tests on a multi-GPU server: all_reduce_perf step by step](https://overgpu.com/docs/multi-gpu-nccl): Check NVLink topology, build nccl-tests, read busbw from all_reduce_perf and launch torchrun jobs across 8 GPUs.