AMD MI350X

AMD MI350X

Air-cooled CDNA 4 accelerator for standard data center racks.

Last update 1 hour ago
Compare vs other GPUs →
Jump to: Specs Pricing AI models FAQ Alternatives

Key Specifications

Architecture
CDNA 4
Memory per GPU
288 GB HBM3e
Memory bandwidth
8,000 GB/s
GPU interconnect
Infinity Fabric 1,075 GB/s
Release date
Q2 2025

MI350X Pricing and Availability

Of 2 listings across 1 cloud provider, none are verified in stock. Spot is listed from $4.00 per GPU per hour and reservations from $4.76 (12 mo).

1 providers

How this list works

The two views

By provider shows one card per company, placed where its best offer ranked. By configuration lists every offer.

Order

What you can rent comes first: in stock, then waitlist, then not reported, then out of stock. Priced offers rank ahead of quote-only, and on-demand ahead of other billing types.

Within a group, five factors set the order:

  1. Location: datacenter proximity, blended with provider HQ.
  2. Price: hourly price, per GPU and in total.
  3. Billing type: reserved ahead of spot, spot ahead of quote-only.
  4. Specs: more VRAM, vCPUs and RAM.
  5. Provider diversity: a provider's repeat rows rank slightly lower, so one company can't take all the top spots.

Search

We match the provider, its country, GPU model, form factor, billing type, availability and instance name. Matching is partial and case-insensitive. Everyday words work too, so "interruptible" finds spot and "sold out" finds out of stock.

Transparency and funding

  • Ads and sponsors: Paid placements sit at the top of the list and are always labeled as sponsored content. Sponsorship never influences the organic ranking itself.
  • Affiliates: Affiliate links are marked. We may earn a commission if you click them, but commissions never affect the order.
  • Prices: Shown in USD, converted at daily reference rates where a provider publishes in another currency. A month means 720 hours. How we estimate costs.
DigitalOcean logo

DigitalOcean

United States of America flag USA 2 configs 1x-8x

Reserved from $4.76 Spot from $4.00
From $4.76 / GPU / hr Reserved (12mo) Visit website

Heads up: Prices are estimates from published rates for common setups and vary by region and usage; verify with the provider before provisioning.

What size AI models can the MI350X run?

One MI350X has 288 GB of HBM3e. On one card it fits models up to about 453B parameters at 4-bit, and 125B at 16-bit, assuming one request at 32K context. The table below shows the smallest node that fits.

Model 4-bit8-bit16-bit
Alibaba Cloud logo Qwen3.8-27B
1 GPU 25 GB
1 GPU 37 GB
1 GPU 64 GB
Google Cloud logo Gemma 4 31B
1 GPU 27 GB
1 GPU 41 GB
1 GPU 71 GB
OpenAI logo GPT-OSS-120B
1 GPU 74 GB
1 GPU 127 GB
1 GPU 244 GB
DeepSeek logo DeepSeek V4 Flash
1 GPU 166 GB
Needs 2 GPUs 294 GB
Needs 3 GPUs 578 GB
Z.AI logo GLM-5.3-Flash
1 GPU 186 GB
Needs 2 GPUs 330 GB
Needs 3 GPUs 650 GB
MiniMax logo MiniMax-M3
1 GPU 245 GB
Needs 2 GPUs 438 GB
Needs 4 GPUs 866 GB
Moonshot AI logo Kimi K3
Needs 6 GPUs 1,550 GB
Needs 11 GPUs 2,810 GB
Needs 22 GPUs 5,610 GB

Estimates, not quotes. Memory is weights plus KV cache at 32K context per request. How we estimate costs.

Vast.ai Logo Vast.ai
Our sponsor
Per-second billing. Pay only for what you use.
50+ GPU models from RTX 3060 to B200, from $0.04/hr
No minimums, no long-term contracts.
SOC2 certified · PyTorch, CUDA & TensorFlow templates ready
Learn more

Frequently Asked Questions

Why choose the MI350X?

288GB HBM3e with MXFP6 and MXFP4 support on CDNA 4, in a passive OAM module at 1000W that drops into air-cooled racks. The same memory and bandwidth as the MI355X without a liquid loop.

When is the MI350X not a good fit?

Clocked lower and held to a smaller power envelope than the MI355X, which returns more throughput per GPU where direct liquid cooling is available. The ROCm software stack, while rapidly improving, is narrower than CUDA, so check framework support first.

Which cloud providers offer the MI350X?

The MI350X is listed by 1 provider.

What data types does the MI350X support?

The MI350X supports 8 precision formats. Training: BF16, FP16, FP32. Inference: FP4, FP6, FP8, INT8. Scientific: FP64.

Does the MI350X support Infinity Fabric?

Yes. 1,075 GB/s bidirectional.

How much does the MI350X cost per hour?

As of August 28, 2026, we track 2 listings from 1 provider, none of them on demand. Prices are per GPU per hour.

Billing type Listings Cheapest in stock
Reserved 1 $4.76 (12 mo)
Spot 1 $4.00

Technical Specifications
FP16 Vector 144.2 TFLOPS
FP16 Matrix 2.3 PFLOPS (w/ sparsity: 4.6 PFLOPS)
BFLOAT16 2.3 PFLOPS (w/ sparsity: 4.6 PFLOPS)
INT8 Matrix 4.6 POPS (w/ sparsity: 9.2 POPS)
MXFP8 4.6 PFLOPS
OCP-FP8 4.6 PFLOPS (w/ sparsity: 9.2 PFLOPS)
MXFP6 9.2 PFLOPS
MXFP4 9.2 PFLOPS
FP64 Vector 72.1 TFLOPS
FP32 Vector 144.2 TFLOPS
FP64 Matrix 72.1 TFLOPS
FP32 Matrix 144.2 TFLOPS
Memory Capacity 288 GB HBM3E
Memory Bandwidth 8 TB/s max. peak theoretical
Memory Interface 8192 bits
Memory Clock 8 GHz
AMD Infinity Cache™ (Last Level) 256 MB
Form Factor OAM module
Cooling Passive OAM (air-cooled)
Lithography TSMC 3nm/6nm FinFET
Transistor Count 185 billion
GPU Compute Units 256
Matrix Cores 1024
Stream Processors 16,384
Peak Engine Clock 2.2 GHz
Scale-up AMD Infinity Fabric™ Links 7x 153 GB/s
Scale-out AMD Infinity Fabric™ Links 1x 128 GB/s
I/O Interconnect 1 PCIe® Gen 5 x16
RAS Features Full-chip ECC memory
Maximum TBP 1000W

Source: official AMD MI350X datasheet.

Alternatives to AMD MI350X

Last updated

Back to top ↑