CDNA 3Live Marketplace Asks Available

AMD Instinct MI300X

MI300X is AMD's 192 GB HBM3 accelerator at about 5.3 TB/s. One GPU can hold a 70B model in BF16 that would need two H100s or FP8 on an H200. The catalog host is 8 vCPUs and 241 GB of RAM. The software stack is ROCm, not CUDA.

Browse Live Marketplace →All Hardware Fleet

Flexible Compute Modes

Deploy AMD Instinct MI300X Your Way: Docker Pods or Root KVM VPS

NovaPod lets you decide how you want to harness GPU power. Whether you need an instant pre-cached container stack, a full root virtual machine with dedicated public IP, or a multi-node NVLink cluster, pay only for what you run with second-by-second crypto billing.

Docker GPU PodCold Boot < 15s

Fast, Lightweight AI Containers

Pure GPU compute without hypervisor overhead. Boot straight into PyTorch, vLLM, ComfyUI, or JupyterLab.

1-second precision metering
37+ Pre-cached ML framework templates
Zero idle fee when instance is stopped

Deploy as Docker Pod

Root KVM VPSPersistent NVMe

Full Root Virtual Machine

Dedicated operating system (Ubuntu 24.04 LTS / Windows GPU Server) with complete sudo root SSH access from minute one.

Dedicated public IPv4 (all ports open)
Pause anytime; pay only $0.00014/GB/hr for disk
Run Docker-in-Docker, k3s, custom drivers

Deploy as Root VPS

Multi-GPU Cluster900 GB/s NVLink

Supercomputing & Distributed Training

Pool up to 8× AMD Instinct MI300X on a unified node fabric with Quantum-2 InfiniBand RDMA for multi-node scale.

Zero-loss all-reduce gradient exchange
PyTorch FSDP, DeepSpeed Zero-3 ready
Native GPUDirect Storage & RDMA fabric

Configure Cluster

Hardware Specifications

Architecture: AMD CDNA 3
VRAM: 192 GB HBM3
Bandwidth: 5.3 TB/s
FP16 dense: 1,307 TFLOPS
FP8 dense: 2,615 TFLOPS
FP32: 163.4 TFLOPS
Stream processors: 19,456
Process: TSMC 5 nm / 6 nm
Form factor: OAM
TDP: 750 W
Software: ROCm
Reference host: 8 vCPUs, 241 GB RAM
Two-GPU host: More vCPUs and about twice the RAM, when listed
FP4: No
CUDA images: Will not boot unchanged

192 GB, on ROCm

MI300X is AMD's CDNA 3 accelerator: 192 GB of HBM3 at about 5.3 TB/s. The menu host is 8 vCPUs and 241 GB of RAM, which is a lot of CPU memory next to one GPU. The software stack is ROCm. PyTorch, vLLM, and SGLang have ROCm builds. A CUDA image will not boot.

BF16 weights for a 70B model are about 140 GB. One MI300X holds them with roughly 30 GB to spare. On NVIDIA that footprint is two H100s, or FP8 on a single H200. A pair of MI300X cards is the next step for mixture-of-experts models that still do not fit.

When the memory is worth the stack change

Pick it when the requirement is one 192 GB GPU and the training or serving path is already ROCm. Bandwidth at 5.3 TB/s sits between H200 and B200. Peak tokens per second on CUDA-tuned kernels are often still higher on Hopper or Blackwell. If you are not memory-bound, the extra capacity does not pay for a port.

Do not pick it for TensorRT-LLM, NVIDIA NIM, FlashAttention kernels you have not rebuilt, or a custom CUDA op you have not run through HIP. Test one GPU. Then decide whether a second GPU or a move back to B200 makes sense.

How this marketplace treats it

There is no assumed discount tier. If a host lists a reclaimable MI300X, the offer will say so. Otherwise treat it as a card you keep until you stop it, billed per second at the ask.

B200 is the CUDA card in the same 192 GB class, with HBM3e, 8 TB/s, and FP4. H200 is the CUDA card when 141 GB is enough. MI300X is the card when the model needs more than that on one GPU and the image is already ROCm.

Architectural Generation Comparison

Compare memory first, then the software stack. MI300X is not a CUDA card.

	MI300X	H200	B200
Vendor	AMD CDNA 3	NVIDIA Hopper	NVIDIA Blackwell
VRAM	192 GB HBM3	141 GB HBM3e	192 GB HBM3e
Bandwidth	5.3 TB/s	4.8 TB/s	8 TB/s
Stack	ROCm	CUDA	CUDA
FP4	No	No	Yes
Menu host	8 vCPU · 241 GB	44 vCPU · 182 GB	30 vCPU · 184 GB

Storage & Tooling Infrastructure

High-Speed Storage, Native Ports & Crypto Rails

Every AMD Instinct MI300X instance on NovaPod connects directly to high-throughput infrastructure designed to eliminate I/O bottlenecks and onboarding delays.

25 Gb/s Shared Network Storage

Mount large model checkpoints and training sets across one or multiple instances simultaneously. Data outlives single pods and stays accessible when instances stop.

JupyterLab, VS Code & SSH

Connect from your local terminal with standard SSH keys or open web-based IDEs directly in browser tabs without managing complex reverse proxy configs.

Non-Custodial Multi-Chain Rails

Deposit USDT, USDC, SOL, ETH, or BNB to your dedicated personal deposit address. Confirmed on-chain in seconds with zero mandatory identity verification.

Target Workloads for AMD Instinct MI300X

70B in BF16 on one GPU

Weights are about 140 GB, so one MI300X holds them with roughly 30 GB left for cache. On NVIDIA that footprint is two H100s, or FP8 on a single H200.

Llama-class 70B BF16vLLM on a ROCm imageLong context without a second GPU90% memory target, then measure cache

Long context and concurrency

192 GB is the point. Peak tokens per second on a CUDA-tuned kernel are often still higher on Hopper or Blackwell. If you are not memory-bound, the capacity does not pay for a port.

Large KV cacheMixtral 8x7B in FP16, about 93 GBHigher batch once weights fitMeasure before you move traffic

Fine-tuning with headroom

ROCm PyTorch can run the same recipe you used on NVIDIA, after the image changes. A batch that fit on H100 may need a smaller micro-batch and more gradient accumulation.

LoRA on an 8B as a smoke test70B only after the 8B run is cleanHIP for custom kernelsCopy checkpoints off the machine yourself

A ROCm trial before a larger order

Start on one GPU. A pair is 384 GB and is the next step for a Mixtral 8x22B in BF16. An 8-GPU node is a quote, not the self-serve button.

1× first2× when one GPU OOMs8× by requestSame model on H200 if the port fails

When to Choose AMD Instinct MI300X

Pick MI300X for more than 141 GB on one GPU

192 GB of HBM3 at 5.3 TB/s, and only if the training or serving path is already ROCm.

Pick H200 or H100 if you need CUDA

TensorRT-LLM, NIM, FlashAttention kernels you have not rebuilt, and any custom CUDA op belong on NVIDIA. H200 is the memory upgrade inside CUDA. H100 is the default when 80 GB is enough.

Pick B200 for 192 GB and FP4

Same capacity class, HBM3e at 8 TB/s, CUDA, and Blackwell numerics.

Stay on CUDA if the port is the project

A week spent rebuilding kernels is not a cheaper GPU. Test one MI300X. If the image does not run, go back.

One GPU, BF16 70B

BF16 weights for 70B are about 140 GB. One MI300X holds them with roughly 30 GB left. On NVIDIA that job is two H100s or FP8 on an H200.

What will not boot

A CUDA-only image, TensorRT-LLM, or a custom kernel you have not built with HIP. Test the ROCm image on one GPU before you move traffic.

Frequently Asked Questions

How much does an MI300X cost here?

The live ask, per second. Other clouds publish MI300X rates that are sometimes lower and sometimes only sold as an 8-GPU node. A listed rate you cannot launch is not a price. Deploy the offer that is actually up.

Is there a cheaper tier?

Only if the marketplace marks the offer as reclaimable. Do not assume a spot discount. Keep the bill down by using one GPU, starting the job as soon as it boots, and destroying the machine when the run ends.

Is there a minimum runtime?

Not on this site. Billing starts with the running machine and stops when you stop it. There is no 20-minute or 60-minute floor written into the product.

Can I pause it?

Stopping the rental is how you stop the charge. Copy checkpoints to storage you control before you destroy the machine. Do not assume a paused GPU keeps its disk.

Which image?

A ROCm image. PyTorch, vLLM, and SGLang have ROCm builds. Check the ROCm version on the machine and pick a container tag that matches it. A CUDA-only image will not run.

Will my CUDA code run?

Most PyTorch code will, because ROCm builds still speak the torch.cuda API. TensorRT-LLM, NVIDIA NIM, FlashAttention's CUDA kernels, and a hand-written CUDA op will not, until you port the kernel.

Does vLLM work?

Yes, on a ROCm build. AMD publishes ROCm vLLM images for Instinct cards. Test the exact model. A batch size that was safe on H100 can OOM here and need a smaller micro-batch.

What fits in 192 GB?

A 70B model in BF16, with roughly 30 GB left. Mixtral 8x7B in FP16 fits with room for batch. A pair of GPUs is 384 GB, enough for larger mixture-of-experts models in BF16. A 405B model does not fit on two GPUs at 8-bit.

MI300X or H200?

MI300X when the model needs more than 141 GB on one GPU and you can run ROCm. H200 when the model fits in 141 GB and you want CUDA, TensorRT-LLM, and FP8.

MI300X or H100?

H100 when the job fits in 80 GB. It is often faster per GPU on current CUDA libraries. MI300X when one 192 GB card replaces a tensor-parallel pair of H100s.

Where does it run?

In the region printed on the offer. This page does not pin a single city. Plan latency from that region if you serve users on another continent.

Can I get eight of them?

Ask. Self-serve is the offer you see, often one GPU or a pair. An 8-GPU node is arranged, not assumed.

Alternative Hardware Options

Ready to Deploy AMD Instinct MI300X?

Launch an instance on the live marketplace in under 15 seconds. Fund your balance with cryptocurrency and pay per second with zero commitments.

Launch on Marketplace →