Can I deploy NVIDIA GH200 or AMD MI250X on Runcrate today?

NVIDIA GH200 is not in stock on Runcrate right now; leave your email on its pricing page to register interest. AMD MI250X is not in stock on Runcrate right now; leave your email on its pricing page to register interest.

How much VRAM does each GPU have?

NVIDIA GH200 has 96GB HBM3. AMD MI250X has 128GB HBM2e.

GH200 or MI250X: which holds a 70B model in 8-bit weights?

It needs 70 GB of 8-bit weights. On the GH200: 1 GPU, 26 GB spare. On the MI250X: 1 GPU, 58 GB spare. Pick the GH200: same GPU count, up to 22% more memory bandwidth.

GPU comparison

GH200 vs MI250X.

Pick the MI250X to fit bigger models on one GPU (128 GB vs 96 GB). Pick the GH200 for speed: 2.6x the FP16 throughput and up to 22% more memory bandwidth.

Rent either on Runcrate

GH200Not in stock right now
Register interest GH200 pricing
MI250XNot in stock right now
Register interest MI250X pricing

Which is better for

Which to pick for each job. Memory fit and throughput, from the datasheets.

JobGH200MI250XPick

Hold a 70B model in 8-bit weights
70 GB of 8-bit weights
GH200: 1 GPU, 26 GB spare
MI250X: 1 GPU, 58 GB spare
GH200. Same GPU count, up to 22% more memory bandwidth.
Serve an 8B model in 16-bit weights
16 GB of 16-bit weights
GH200: 1 GPU, 80 GB spare
MI250X: 1 GPU, 112 GB spare
GH200. Same GPU count, up to 22% more memory bandwidth.
Fine-tune a 70B model with QLoRA
About 46 GB with 4-bit weights
GH200: 1 GPU, 50 GB spare
MI250X: 1 GPU, 82 GB spare
GH200. Same GPU count, 2.6x the FP16 throughput.
Train on a full node
FP16 tensor throughput per node
GH200: 1x, 990 TFLOPS
MI250X: 8x, 3,064 TFLOPS
MI250X. 8 GPUs per node, not 1.
Throughput per watt
FP16 TFLOPS per 100 W of TDP
GH200: Superchip power, not a GPU TDP
MI250X: 68 TFLOPS per 100 W
Either. Not comparable.

Compare another pair

Hopper

NVIDIA GH200

Not in stock right now

Memory: 96 GB HBM3
FP16 tensor: 990 TFLOPS
Bandwidth: up to 4.0 TB/s

Full pricing details

CDNA 2

AMD MI250X

Not in stock right now

Memory: 128 GB HBM2e
FP16 tensor: 383 TFLOPS
Bandwidth: 3.3 TB/s

Full pricing details

At a glance

Spec by spec. The longer bar is the better number.

Memory

NVIDIA GH20096 GB

AMD MI250X128 GB

Memory bandwidth

NVIDIA GH200up to 4.0 TB/s

AMD MI250X3.3 TB/s

FP16 tensor

NVIDIA GH200990 TFLOPS

AMD MI250X383 TFLOPS

FP32

NVIDIA GH20067 TFLOPS

AMD MI250X47.9 TFLOPS

Full specs

The complete spec sheet.

Spec	NVIDIA GH200	AMD MI250X
Architecture	Hopper	CDNA 2
Memory	96 GB	128 GB
Memory type	HBM3	HBM2e
Memory bandwidth	up to 4.0 TB/s	3.3 TB/s
FP32	67 TFLOPS	47.9 TFLOPS
FP16 tensor	990 TFLOPS	383 TFLOPS
INT8	1979 TOPS	383 TOPS
Power	up to 1000 W per superchip (CPU, GPU and memory)	560 W per GPU (TDP)
Form factor	Superchip	OAM
NVLink	One GPU	No
GPUs per node	1	8
Hourly price (live)	Not in stock right now	Not in stock right now

Questions

GH200 or MI250X. What buyers ask first.

Other comparisons