Flopper.io Flopper.io
vs

AMD Instinct MI300X 192GB vs NVIDIA H100 SXM5 80GB

The Instinct MI300X delivers 32% more FP8 throughput than the H100 (2,615 vs 1,979 TFLOPS dense). The Instinct MI300X also carries 112 GB more memory (192 GB vs 80 GB).

Instinct MI300X: CDNA 3, 2023 H100: Hopper, 2022
FP8 dense lead
+32%
Instinct MI300X: 2,615 vs 1,979 TFLOPS
Memory
192 vs 80 GB
112 GB more for the Instinct MI300X
Bandwidth
5.3 TB/s vs 3.4 TB/s
Instinct MI300X moves data faster
TDP
750 W vs 700 W
H100 draws 50 W less
See rental pricing from $0.50/hr Download comparison sheet

Share this page

AMD Instinct MI300X 192GB vs NVIDIA H100 SXM5 80GB

A post with this link shows the card on the Image tab as its preview.

Post on X Share on LinkedIn
 Printable version →

AMD Instinct MI300X 192GB

Leads on FP8 Full specs →

What is the AMD Instinct MI300X?

The Instinct MI300X is a GPU from AMD on the CDNA 3 architecture, an accelerator for AI training, inference and high-performance computing. It is an OAM module, the open accelerator format used in multi-GPU servers. Instinct MI300X full specifications →

Can I rent the Instinct MI300X in the cloud?

Yes. Cloud providers tracked by Flopper rent the Instinct MI300X by the hour, and the pricing page compares their current rates and the price history side by side. Instinct MI300X rental prices →

How does the Instinct MI300X compare with other accelerators?

This page sets the Instinct MI300X against the H100. The Instinct MI300X is the newer generation of the two. This is a cross-vendor matchup, AMD against NVIDIA; the Instinct MI300X is an OAM module while the H100 is an SXM module. Flopper also compares it head to head with the H200 SXM 141GB, across every published precision, memory and power. Instinct MI300X vs H200 SXM 141GB →

Where do the Instinct MI300X figures come from?

Every Instinct MI300X figure on this page traces to AMD's own datasheet or specification table. How Flopper verifies specifications →

NVIDIA H100 SXM5 80GB

Full specs →

What is the NVIDIA H100?

The H100 is a GPU from NVIDIA on the Hopper architecture, an accelerator for AI training, inference and high-performance computing. It is an SXM module, the mezzanine format used in dense multi-GPU servers. H100 full specifications →

Can I rent the H100 in the cloud?

Yes. Cloud providers tracked by Flopper rent the H100 by the hour, and the pricing page compares their current rates and the price history side by side. H100 rental prices →

How does the H100 compare with other accelerators?

This page sets the H100 against the Instinct MI300X. The H100 is the older generation of the two. This is a cross-vendor matchup, NVIDIA against AMD; the H100 is an SXM module while the Instinct MI300X is an OAM module. Flopper also compares it head to head with the A100 SXM4 80GB, H200 SXM 141GB and B300 SXM 262GB, across every published precision, memory and power. H100 vs A100 SXM4 80GB →

Where do the H100 figures come from?

Every H100 figure on this page traces to NVIDIA's own datasheet or specification table. How Flopper verifies specifications →

Specifications

Instinct MI300X
H100
Architecture
CDNA 3Hopper
Launch Year
20232022
Form Factor
OAMSXM
Memory
192 GB80 GB
Memory Bandwidth
5.3 TB/s3.4 TB/s
TDP
750 W700 W
Process Node
5nm FinFET4nm

Performance (TFLOPS)

Instinct MI300X
H100
82 TFLOPS 34 TFLOPS
163 TFLOPS 67 TFLOPS
654 TFLOPS
1,307 TFLOPS sparse
494 TFLOPS
989 TFLOPS sparse
1,307 TFLOPS
2,615 TFLOPS sparse
990 TFLOPS
1,979 TFLOPS sparse
1,307 TFLOPS
2,615 TFLOPS sparse
990 TFLOPS
1,979 TFLOPS sparse
2,615 TFLOPS
5,230 TFLOPS sparse
1,979 TFLOPS
3,958 TFLOPS sparse
FP6
No verified data available No verified data available
No verified data available No verified data available
2,615 TOPS
5,230 TOPS sparse
1,979 TOPS
3,958 TOPS sparse
Tensor Core
No verified data available67 TFLOPS

Figures are dense. Where a vendor publishes a figure it sits under the dense one and never decides a winner. Hover any precision for its definition, or read the GPU glossary.

FLOPS by Precision

What actually differs

Price the MI300X per model, not per GPU-hour. Its 192 GB holds a 70B model in BF16 with room for KV cache; the H100’s 80 GB does not, so the same deployment takes two H100s and a tensor-parallel split. On-demand, one MI300X from $2.39 per GPU-hour replaces two H100s from $1.79 each, and it reads weights at 5.3 TB/s against 3.35 TB/s, the figure that bounds decode.

The trade reverses when your model fits in 80 GB. Then the H100 is the cheaper hour, your CUDA build deploys unchanged, and AMD’s higher dense FP8 rating only counts once your kernels reach it on ROCm.

Unlike newer Instinct parts, the MI300X rents from seven providers against 34 for the H100, so supply no longer blocks a trial. Qualify your serving stack on a rented MI300X before moving the fleet, or weigh NVIDIA’s memory answer in MI300X vs H200.

New to the terms? The glossary explains FP64, FP32, TF32, BF16, FP16, FP8, INT8, Sparsity and TFLOPS. For how the generations relate, see NVIDIA architectures explained.

Frequently asked questions

Is the MI300X faster than the H100?

On vendor peak figures, dense to dense, yes. AMD rates the MI300X at 2,614.9 TFLOPS FP8 and 1,307.4 TFLOPS BF16; NVIDIA rates the H100 SXM at 1,979 TFLOPS FP8 and 990 TFLOPS BF16. The MI300X also reads memory at 5.3 TB/s against 3.35 TB/s. Delivered speed depends on how close your ROCm kernels get to those peaks.

Is the MI300X cheaper to rent than the H100?

Per model, often; per GPU-hour, not always. Flopper tracks the MI300X from $0.50 per GPU-hour on community capacity and $2.39 on-demand, against $1.41 community and $1.79 on-demand for the H100 SXM. The saving lands when a model that needs two H100s for memory runs on one MI300X.

How much more memory does the MI300X have than the H100?

The MI300X carries 192 GB of HBM3 against 80 GB on the H100 SXM, 2.4 times as much. A 70B model in BF16 fits on one MI300X with room for KV cache, while the H100 needs two GPUs and a tensor-parallel split to serve the same model.

Can I move an H100 workload to the MI300X?

Usually, after qualification. PyTorch and vLLM both run on ROCm, so standard training and serving code ports, but custom CUDA kernels, quantisation recipes and NVIDIA-only libraries need replacing or retesting. Flopper tracks 16 MI300X listings across 7 providers, against 67 H100 listings across 34, so a pilot is easy to rent.

Where to Rent

AMD Instinct MI300X 192GB

Rent from

$0.50 / GPU-hr

Lowest of 16 live listings across 7 providers · median $3.22

See all Instinct MI300X rental prices →

Price history and every provider on the pricing page. Some links are affiliate links.

NVIDIA H100 SXM5 80GB

Rent from

$1.60 / GPU-hr

Lowest of 66 live listings across 34 providers · median $3.85

See all H100 rental prices →

Price history and every provider on the pricing page. Some links are affiliate links.

Get Comparison Updates

New GPUs added weekly. Be the first to see how they compare.

Open in Full Compare Tool → Instinct MI300X full specs H100 full specs Estimate cluster cost in the calculator →

© 2026 Flopper.io - Compare the GPUs Powering AI