Rent AMD MI355X: 8-GPU AMD Instinct servers by the month

If you want to rent AMD MI355X capacity, we offer it as a bare-metal AMD Instinct server with eight MI355X GPUs for $19,900 per month. That is 2.3 TB of HBM3E in one machine, paid monthly by card or crypto, with a ROCm image ready on delivery. We do not offer a single-GPU MI355X plan: the 8-GPU server is the only configuration, and stock is limited.

What the MI355X is

The MI355X is AMD's Instinct accelerator on the CDNA 4 architecture. Per AMD's specifications, each GPU has 288 GB of HBM3E at 8 TB/s and delivers 5.0 PFLOPS of dense FP8 and 2.5 PFLOPS of dense BF16. CDNA 4 also adds support for the MXFP4 format, which AMD lists at 10.1 PFLOPS. The GPU comes as an OAM module rated at 1,400 W and is liquid-cooled.

The interconnect works differently from NVIDIA's HGX servers. AMD lists seven Infinity Fabric links of 153.6 GB/s per GPU, about 1.08 TB/s in total, and in an eight-GPU platform each GPU links directly to each of the other seven. Collective operations such as all-reduce, which spread traffic across all peers, can use all seven links at once; a transfer between just two GPUs is limited to one link. HGX B200 and B300 servers, by contrast, give every GPU 1.8 TB/s of NVLink bandwidth to any peer. Our article on the AMD MI355X for AI covers the architecture in more depth.

An 8-GPU AMD Instinct server, not a single GPU

If you were looking for MI355X cloud instances sold by the GPU, note that we rent this one only as a whole server. You get the full machine: two AMD EPYC 9575F processors with 128 cores in total, 3 TB of DDR5, eight 3.84 TB NVMe drives (about 30 TB) and eight 400G RoCE network adapters for multi-node work. It is bare metal, so there is no hypervisor between your code and the GPUs.

That shape suits teams with enough work to keep eight GPUs busy: serving several large models, fine-tuning, or research code that already runs on ROCm. If you need only one GPU with 288 GB, the single NVIDIA B300 offers the same memory per GPU on CUDA, at $5,190 per month.

Memory math: 288 GB per GPU, 2.3 TB per server

Weights take roughly 2 bytes per parameter in BF16 and 1 in FP8, plus KV cache and 10–30% headroom. With 288 GB per GPU:

• 70B model in BF16: about 140 GB, so it fits on one GPU with well over 100 GB left for KV cache. One server can run eight independent 70B replicas, with no tensor parallelism and no inter-GPU traffic.
• 405B-class model in FP8: about 405 GB, which fits on two GPUs. You can run up to four copies per server, or one copy spread over all eight GPUs for a much larger KV cache.
• 671B mixture-of-experts model in FP8: about 671 GB, which needs at least three GPUs and leaves roughly 1.6 TB of the server for KV cache and overhead.
• Full fine-tuning of a 70B model: Adam needs roughly 16–20 bytes per parameter before activations, about 1.1–1.4 TB, which fits in the 2.3 TB of HBM3E with roughly 0.9–1.2 TB left for activations.

Our VRAM calculator runs the same math for other models and precisions.

ROCm software: what runs and what to check

The server comes with root access over SSH and a ROCm image; Debian or Rocky Linux are available on request. PyTorch has official ROCm builds that keep the familiar torch.cuda interface, so PyTorch code that sticks to standard operators usually runs without changes, and vLLM supports ROCm for serving. A first check:

rocm-smi                                  # GPUs, memory use, temperature
docker run -it --device=/dev/kfd --device=/dev/dri --group-add video \
  --ipc=host --shm-size 16g rocm/pytorch:latest \
  python -c "import torch; print(torch.cuda.device_count())"

What needs testing: custom CUDA kernels, which must be ported to HIP, and libraries that ship only CUDA binaries, such as some quantization or attention kernels. Multi-GPU jobs use RCCL, AMD's counterpart to NCCL. Our ROCm documentation covers the details. Because the minimum term is one month, you can validate your stack on real hardware before you commit to anything longer.

What it costs to rent AMD MI355X

The 8-GPU MI355X server costs $19,900 per month, about $3.41 per GPU-hour at full use: the monthly price divided by eight GPUs and 730 hours. Within our own catalog, the HGX B300 also carries 2.3 TB of HBM across eight GPUs, for $39,900, and the HGX H200 carries 1.1 TB for $23,900. The MI355X gives the most GPU memory per dollar of our 8-GPU servers; what you trade is the CUDA ecosystem and NVLink's switched fabric. Our comparisons of the MI355X and the B200 and of the MI355X and the H200 weigh those trade-offs.

Against hourly billing, the break-even point is $19,900 divided by the hourly rate you would pay for eight comparable GPUs. If you need the server for most of the month, monthly wins; for a few days of testing, hourly billing costs less where it is available. The price includes 50 TB of outbound traffic per month with no egress fees, a public IPv4 address and an IPv6 /64, DDoS mitigation, 24/7 support from engineers and the 99.9% availability SLA. There is no setup fee.

Ordering an MI355X server

Request the server on the site, and we confirm the configuration, monthly price, location and delivery date in writing. Payment by card or crypto follows, with the crypto exchange rate locked at checkout, and root access details arrive by email when the server is ready. If we miss the confirmed delivery date, you can cancel for a full refund. Most orders need only an email address and a payment; orders for several 8-GPU servers may require verification first. Ask about volume pricing or 6–12 month terms through the contact form.