Skip to content

GPU comparison

GH200 vs RTX 3090.

The GH200 is ahead on memory (96 GB vs 24 GB) and speed (13.9x the FP16 throughput and up to 4.3x the memory bandwidth). The RTX 3090 makes sense when your model fits in 24 GB and its hourly price is lower.

Rent either on Runcrate

Which is better for

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

  • Hold a 70B model in 8-bit weights

    70 GB of 8-bit weights

    GH200: 1 GPU, 26 GB spare

    RTX 3090: 4 GPUs, more than one 2x node

    GH200. Fits on one GPU.

  • Serve an 8B model in 16-bit weights

    16 GB of 16-bit weights

    GH200: 1 GPU, 80 GB spare

    RTX 3090: 1 GPU, 8 GB spare

    GH200. Same GPU count, up to 4.3x the memory bandwidth.

  • Fine-tune a 70B model with QLoRA

    About 46 GB with 4-bit weights

    GH200: 1 GPU, 50 GB spare

    RTX 3090: 2 GPUs (48 GB)

    GH200. Fits on one GPU.

  • Train on a full node

    FP16 tensor throughput per node

    GH200: 1x, 990 TFLOPS

    RTX 3090: 2x, 142 TFLOPS, NVLink

    GH200. 13.9x the FP16 throughput per GPU.

  • Throughput per watt

    FP16 TFLOPS per 100 W of TDP

    GH200: Superchip power, not a GPU TDP

    RTX 3090: 20 TFLOPS per 100 W

    Either. Not comparable.

Compare another pair

vs

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

Ampere

NVIDIA RTX 3090

Not in stock right now
Memory
24 GB GDDR6X
FP16 tensor
71 TFLOPS
Bandwidth
936 GB/s
Full pricing details

At a glance

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

Memory

NVIDIA GH20096 GB
NVIDIA RTX 309024 GB

Memory bandwidth

NVIDIA GH200up to 4.0 TB/s
NVIDIA RTX 3090936 GB/s

FP16 tensor

NVIDIA GH200990 TFLOPS
NVIDIA RTX 309071 TFLOPS

FP32

NVIDIA GH20067 TFLOPS
NVIDIA RTX 309035.6 TFLOPS

Full specs

The complete spec sheet.

SpecNVIDIA GH200NVIDIA RTX 3090
ArchitectureHopperAmpere
Memory96 GB24 GB
Memory typeHBM3GDDR6X
Memory bandwidthup to 4.0 TB/s936 GB/s
FP3267 TFLOPS35.6 TFLOPS
FP16 tensor990 TFLOPS71 TFLOPS
INT81979 TOPS284 TOPS
Powerup to 1000 W per superchip (CPU, GPU and memory)350 W per GPU (TDP)
Form factorSuperchipConsumer
NVLinkOne GPUYes
GPUs per node12
Hourly price (live)Not in stock right nowNot in stock right now

Questions

GH200 or RTX 3090. What buyers ask first.

Other comparisons

More pairs with the GH200 or the RTX 3090.