Compute Comparison
NVIDIAHopper2023

GH200 Grace Hopper 96GB

Grace Hopper superchip — ARM Grace CPU + H100 GPU connected via NVLink-C2C at 900 GB/s. 96GB HBM3 + 480GB LPDDR5X system memory accessible as unified pool.

VRAM
96GB
HBM3
FP16
2.0k
TFLOPS
Bandwidth
4.0k
GB/s
TDP
900W
power
Best for:Memory-bound LLM inferenceLarge context windowsUnified CPU+GPU workloads

GH200 Grace Hopper 96GB Overview

The NVIDIA GH200 combines a Hopper GH100 GPU with an Arm Grace CPU in one superchip, linking them through 900 GB/s NVLink-C2C. The GPU contributes 96GB of HBM3, 1,979 TFLOPS of FP16/BF16 performance, and FP8 Transformer Engine support, while the Grace side adds large LPDDR5X system memory to the addressable platform.

The HBM3 subsystem reaches 4,000 GB/s, making GH200 especially capable for memory-bound transformer inference. Unified CPU-GPU memory access can simplify workloads whose data or context exceeds GPU memory, but CPU-attached memory is not a substitute for HBM bandwidth; performance still depends on where active tensors reside.

GH200 is tailored to large-context LLM inference, data-intensive HPC, and applications that benefit from tightly coupled CPU and GPU memory. Its specialized form factor and high platform power make it unsuitable for ordinary PCIe GPU servers or workloads that only need standard H100 compute.

Memory

VRAM96 GB
Memory TypeHBM3
Bandwidth4000 GB/s
NVLink BW900 GB/s

Compute Performance

FP3267 TFLOPS
FP161979 TFLOPS
BF161979 TFLOPS
FP83958 TFLOPS
INT83958 TOPS

Hardware

ArchitectureGH100 Grace Hopper
GenerationHopper
Process NodeTSMC 4N
Transistors80B
TDP900 W
InterconnectNVLink-C2C / NVLink 4.0
Release Year2023

Relative Performance

FP16 Compute26%
VRAM Capacity33%
Mem Bandwidth25%

Relative to highest-spec GPU in database

Limitations

Very limited cloud availability — specialized superchip form factor
Requires Grace CPU — cannot use in standard GPU servers
High cost vs standard H100 for compute-bound workloads

Live Cloud PricingOn-demand hourly rates

Loading live prices…

Compare GH200 Grace Hopper 96GB vs…

Use Case Guidance

Memory-bound LLM inference
Large context windows
Unified CPU+GPU workloads

LLM Model Size Guidance

Max model (FP16)~48Bparameters at FP16 precision
Max model (INT8)~96Bparameters at INT8 precision
Max model (INT4)~192Bparameters at INT4/GGUF

Estimates only. Actual capacity depends on context length, KV cache, and framework overhead.

Related Guides

LLM APIs Running on This GPU Class

Providers that serve frontier LLM inference on Hopper-class hardware.

Browse all 42 LLM models

Related GPUs

Frequently Asked Questions

How much VRAM does the GH200 Grace Hopper 96GB have?

The GH200 Grace Hopper 96GB has 96GB of HBM3 memory with 4000 GB/s bandwidth. This enables running models up to approximately 192B parameters at INT4 precision, 96B at INT8, or 48B at FP16.

What is the FP16 performance of the GH200 Grace Hopper 96GB?

The GH200 Grace Hopper 96GB delivers 1979 TFLOPS of FP16 performance and 1979 TFLOPS BF16, and 3958 TFLOPS FP8. INT8 throughput is 3958 TOPS. For transformer inference, memory bandwidth (4000 GB/s) is often the binding constraint rather than raw TFLOPS.

What is the GH200 Grace Hopper 96GB best used for?

The GH200 Grace Hopper 96GB is best suited for: Memory-bound LLM inference, Large context windows, Unified CPU+GPU workloads. Grace Hopper superchip — ARM Grace CPU + H100 GPU connected via NVLink-C2C at 900 GB/s. 96GB HBM3 + 480GB LPDDR5X system memory accessible as unified pool.

What interconnect does the GH200 Grace Hopper 96GB use?

The GH200 Grace Hopper 96GB uses NVLink-C2C / NVLink 4.0 with 900 GB/s NVLink bandwidth for multi-GPU configurations. NVLink enables near-linear tensor-parallel scaling across multiple cards for models that exceed single-card VRAM.

What LLM model sizes can the GH200 Grace Hopper 96GB run?

With 96GB of HBM3, the GH200 Grace Hopper 96GB can run models up to approximately 48B parameters at FP16 (2 bytes/param), 96B at INT8 (1 byte/param), or 192B at INT4/GGUF (0.5 bytes/param). These are estimates — actual capacity depends on context length, KV cache size, and framework overhead. Longer context windows require more KV cache memory, reducing the effective model size that fits.

How does the GH200 Grace Hopper 96GB compare to the A100 for LLM inference?

The GH200 Grace Hopper 96GB has 1979 TFLOPS FP16 vs the A100 80GB's 312 TFLOPS, and 4000 GB/s memory bandwidth vs the A100's 2,039 GB/s. For memory-bound autoregressive LLM inference, bandwidth is the primary determinant of tokens-per-second. The GH200 Grace Hopper 96GB's higher bandwidth gives it a throughput advantage for large model inference.

What is the power consumption of the GH200 Grace Hopper 96GB?

The GH200 Grace Hopper 96GB has a TDP (Thermal Design Power) of 900W. This is the maximum sustained power draw under full load. For data center deployments, total rack power consumption is typically 1.2–1.5× the GPU TDP when accounting for CPU, memory, networking, and cooling overhead. At 900W, the GH200 Grace Hopper 96GB is in the high-power tier — requires specialized data center infrastructure with high-density power delivery.

Ready to rent?

Compare GH200 Grace Hopper 96GB prices across 97+ providers

Live on-demand & spot rates · monthly cost estimates · availability status

Compare rental prices