Skip to main content

GPU Product

Rent AMD Instinct MI350X GPUs

Datacenter AI accelerator with 288 GB HBM3e for large-scale training, LLM inference, and HPC workloads

Technical Specifications

ArchitectureAMD CDNA 4
Memory Size288 GB HBM3e
Memory Bandwidth8 TB/s
FP8 Performance~9.2 PFLOPS
TDP1 000 W
AMD Instinct MI350X

Rental Options

MI350X Rental Options

Rent AMD Instinct MI350X GPUs with reserved pricing. Get datacenter-class AI compute without upfront hardware costs, and scale as your training needs grow.

Start rental in consoleContact Us

Comparison

MI350X vs B200

MI350XB200% Diff
ArchitectureCDNA 4BlackwellN/A
Process Tech3 nm4 nmN/A
Memory TypeHBM3eHBM3eN/A
VRAM288 GB192 GB+50%
Memory Bandwidth8 TB/s~8 TB/s~0%
FP8 Performance~9.2 PFLOPS~9 PFLOPS~+2%
FP16/BF16 Performance~4.6 PFLOPS~4.5 PFLOPS~+2%
TDP1 000 W~1 000 W~0%
InterfaceOAM / PCIe 5NVLink / PCIe 5N/A
Multi-GPU InterconnectInfinity FabricNVLink 5N/A
Max GPUs per Node880%
Vendor EcosystemROCmCUDAN/A

Performance

Key performance metrics

Massive AI Compute

With ~9.2 PFLOPS of FP8 matrix performance on CDNA 4, the MI350X delivers datacenter-class AI throughput for large-scale training and inference workloads.

Industry-Leading Memory

288 GB of HBM3e memory with 8 TB/s bandwidth eliminates memory bottlenecks for the largest foundation models, enabling full-model inference without tensor parallelism.

Open Ecosystem

Built on AMD ROCm, the MI350X supports PyTorch, JAX, and major ML frameworks out of the box — no vendor lock-in, with broad compatibility across AI software stacks.

Use Cases

Use Cases

Large Language Models

Train and serve 100B+ parameter LLMs with 288 GB HBM3e — fit entire models on a single accelerator.

Generative AI

Accelerate diffusion models, video generation, and multi-modal AI pipelines at scale.

Scientific Computing

Run HPC simulations, molecular dynamics, and climate modeling with massive memory and bandwidth.

Multi-Tenant Inference

Serve multiple models or users concurrently with abundant VRAM and high throughput.

MI350X FAQ

Common Questions About the MI350X

The AMD Instinct MI350X is a datacenter AI accelerator built on the CDNA 4 architecture. It features 288 GB of HBM3e memory, 8 TB/s bandwidth, and ~9.2 PFLOPS of FP8 matrix performance, designed for large-scale AI training and inference.
The MI350X offers 50% more memory (288 GB vs 192 GB) with comparable bandwidth (~8 TB/s each) and FP8 compute (~9.2 vs ~9 PFLOPS). Both draw around 1000W TDP. The MI350X's primary advantage is its larger memory capacity for fitting bigger models.
MI350X GPU rental pricing starts at approximately $2.85/hr for 3-month reserved instances and $3.00/hr for 1-month reserved. Contact us for custom pricing and on-demand availability.
The MI350X runs on AMD ROCm and supports PyTorch, JAX, TensorFlow, and other major ML frameworks out of the box. The open ecosystem means no vendor lock-in.
Yes. MI350X nodes support up to 8 GPUs connected via AMD Infinity Fabric for high-bandwidth multi-GPU training and inference workloads.
Absolutely. With 288 GB of HBM3e, the MI350X can serve 100B+ parameter models on a single accelerator. The 8 TB/s bandwidth delivers high tokens-per-second throughput for production inference.
Contact us via our contact form or reach out on Discord with your requirements. We'll work with you to set up MI350X instances tailored to your workload.
Get in touch

Ready to get started?

Get in touch with our team to discuss your requirements and find the right solution for your infrastructure.

Contact Us