AMD · CDNA 3 · Data centre
AMD Instinct MI300X
AMD's CDNA 3 accelerator in the OAM form factor, with a large HBM3 memory pool per GPU, programmed with ROCm.
01 / SPECIFICATIONS
What the hardware holds
- Memory
- 192 GB
- Memory bandwidth
- 5.3 TB/s
- FP8 dense
- 2,614.9 TFLOPS
- FP16 / BF16 dense
- 1,307.4 TFLOPS
- Variant
- Instinct MI300X OAM
- Memory type
- HBM3
- Form factor
- OAM
- Interconnect
- Infinity Fabric 7x 128 GB/s; PCIe Gen5 x16 128 GB/s
- Max power
- 750 W
Source notes for Instinct MI300X OAM
AMD datasheet lists FP16 and BF16 1307.4 TFLOPS dense (2614.9 with sparsity), FP8 2614.9 dense (5229.8 with sparsity); bandwidth 5.3 TB/s max peak theoretical (product page footnote MI300-05A: 5.325 TB/s); maximum TBP 750 W. Datasheet lists 7x 128 GB/s scale-up Infinity Fabric links; the product page (https://www.amd.com/en/products/accelerators/instinct/mi300/mi300x.html) lists 8 Infinity Fabric links at 128 GB/s peak link bandwidth. Product page gives launch date 12/06/2023.
Compute figures are dense, without sparsity. AMD source for Instinct MI300X OAM Checked 8 Nov 2026.
02 / MEMORY FIT
How many do you need to hold a model?
| Model size | FP16 / BF16 | FP8 | INT4 |
|---|---|---|---|
| 8B parameters | 1 GPU16 GB of weights | 1 GPU8 GB of weights | 1 GPU4 GB of weights |
| 70B parameters | 1 GPU140 GB of weights | 1 GPU70 GB of weights | 1 GPU35 GB of weights |
| 405B parameters | 6 GPUs810 GB of weights | 3 GPUs405 GB of weights | 2 GPUs203 GB of weights |
GPUs = parameters × bytes per parameter × 1.2 headroom ÷ 192 GB, rounded up. This covers weights only. Long contexts, large batches, and training need much more memory.
03 / OFFERS
Reviewed on-demand rates
No reviewed rate yet. The sellers below list the MI300X; check their pricing pages for current terms.
Per GPU-hour divides a machine's hourly price by its GPU count, so single GPUs and 8-GPU nodes compare fairly. The billed price is what one machine costs per hour. Storage, network transfer, and minimum terms can add to the bill.
04 / SELLERS
Who lists the MI300X
Crusoe, DigitalOcean list the MI300X without a rate we have reviewed. Ask them for a quote, especially for clusters of 8 or more GPUs.
05 / COMPARE WITH