Skip to content

AMD · CDNA 2 · 2021

AMD MI250X. Not in stock right now.

128 GB of HBM2e and 383 TFLOPS FP16 tensor, for training and inference. The A100 80GB is on demand now from $1.73/hr per GPU on an 8-GPU machine.

Not in stock right now

MI250X is not in stock right now.

Leave your work email to register interest. These requests decide which GPUs we bring on demand next.

The card

128 GB of HBM2e. At 3.3 TB/s.

Memory
128 GB
HBM2e
Bandwidth
3.3 TB/s
Memory to compute
FP16 tensor
383
TFLOPS
Per node
8 GPUs
Infinity Fabric, 560 W each

Spec sheet

How the MI250X compares. Against its closest alternatives.

FP16 tensor

AMD MI250X383 TFLOPS
NVIDIA L40S362 TFLOPS
NVIDIA A100 80GB312 TFLOPS
NVIDIA A100 40GB312 TFLOPS

Memory

AMD MI250X128 GB
NVIDIA L40S48 GB
NVIDIA A100 80GB80 GB
NVIDIA A100 40GB40 GB

Bandwidth

AMD MI250X3.3 TB/s
NVIDIA L40S864 GB/s
NVIDIA A100 80GBup to 2.0 TB/s
NVIDIA A100 40GB1.6 TB/s

Alternatives

Alternatives to the MI250X. Same account, different price and performance.

GPUMemory · FP16 tensorStands outPrice gapFrom
AMD MI250XThis GPU128 GB HBM2e · 383 TFLOPSNot in stock right now
NVIDIA L40S48 GB GDDR6 · 362 TFLOPS-80 GB memory$0.55/hr
NVIDIA A100 80GB80 GB HBM2e · 312 TFLOPS-48 GB memory$1.73/hr
NVIDIA A100 40GB40 GB HBM2 · 312 TFLOPS-88 GB memory$1.99/hr
NVIDIA RTX 509032 GB GDDR7-96 GB memory$0.72/hr

Full specs

AMD MI250X specifications.

Memory

Memory
128 GB
Type
HBM2e
Bandwidth
3.3 TB/s

Compute

FP32
47.9 TFLOPS
FP16 tensor
383 TFLOPS
INT8
383 TOPS
Stream processors
14,080

Power and form

TDP
560 W
Form factor
OAM
Architecture
CDNA 2
Released
2021

Node

GPUs per node
8
Interconnect
Infinity Fabric
Maker
AMD

Questions

Renting the MI250X. What buyers ask first.

Other GPUs

Browse every GPU by architecture.