GPU coverage
Which GPUs are in the estimator, and why others are not
11 GPU entries have a dense BF16 figure printed by their maker and a current price. Each shows the quote and the page. 10 groups are out, each with the reason.
Covered
Dense BF16 printed by the maker
MI355X
2500 TFLOPS dense BF16 · 2 providers · lowest $2.59 per GPU-hour
“Peak bfloat16 (BF16) Matrix Performance 2.5 PFLOPs”
MI300X
1307.4 TFLOPS dense BF16 · 4 providers · lowest $1.85 per GPU-hour
“1307.4 TFLOPS peak theoretical Bfloat16 format precision (BF16)”
MI325X
1307.4 TFLOPS dense BF16 · 2 providers · lowest $2.00 per GPU-hour
“1307.4 TFLOPS peak theoretical Bfloat16 format precision (BF16)”
H100 SXM
1000 TFLOPS dense BF16 · 6 providers · lowest $2.95 per GPU-hourpreliminary
“Peak BF16 Tensor Core1 · 1000 TFLOPS · 2000 TFLOPS2 · 800 TFLOPS · 1600 TFLOPS2”
Printed with: “2. Effective TFLOPS / TOPS using the Sparsity feature”
Printed with: “1. Preliminary performance estimates for H100 based on current expectations and subject to change in the shipping products”
34 further current listings for this GPU do not state a form factor this figure applies to and are not priced here.
GH200
990 TFLOPS dense BF16 · 2 providers · lowest $1.99 per GPU-hour
“BFLOAT16 Tensor Core 1,979 teraFLOPS* · 990 teraFLOPS”
H100 PCIe
800 TFLOPS dense BF16 · 2 providers · lowest $2.50 per GPU-hourpreliminary
“Peak BF16 Tensor Core1 · 1000 TFLOPS · 2000 TFLOPS2 · 800 TFLOPS · 1600 TFLOPS2”
Printed with: “2. Effective TFLOPS / TOPS using the Sparsity feature”
Printed with: “1. Preliminary performance estimates for H100 based on current expectations and subject to change in the shipping products”
42 further current listings for this GPU do not state a form factor this figure applies to and are not priced here.
L40S
362.05 TFLOPS dense BF16 · 14 providers · lowest $0.55 per GPU-hour
“BFLOAT16 Tensor Core TFLOPS · 362.05 I 733* ·”
A100 80GB
312 TFLOPS dense BF16 · 12 providers · lowest $1.09 per GPU-hour
“BFLOAT16 Tensor Core · 312 TFLOPS · 624 TFLOPS* ·”
L40
181.05 TFLOPS dense BF16 · 6 providers · lowest $0.79 per GPU-hour
“BFLOAT16 Tensor Core TFLOPS 181.05 · 362.1**”
A40
149.7 TFLOPS dense BF16 · 3 providers · lowest $0.59 per GPU-hour
“Peak BF16 Tensor TFLOPS with FP32 Accumulate 149.7 · 299.4*”
A10
125 TFLOPS dense BF16 · 4 providers · lowest $1.00 per GPU-hour
“BFLOAT16 Tensor Core · 125 teraFLOPS · 250 teraFLOPS* ·”
Not covered
No dense BF16 figure printed
- H200
- NVIDIA prints BF16 for the H200 as 1,979 TFLOPS (SXM) and 1,671 TFLOPS (NVL) with the footnote "With sparsity". No dense figure is printed.
- B200
- NVIDIA prints FP16/BF16 for Blackwell as a sparse figure and says "Dense is one-half of the sparse spec shown". That is a rule for halving, not a printed dense figure, and we do not halve sparse figures.
- B300
- No per-GPU BF16 figure is printed on the pages read. The HGX B300 table gives an 8-GPU total with the same sparse footnote, and we do not divide totals or halve sparse figures.
- H100 NVL and HGX listings
- The NVIDIA table we read has columns for H100 SXM5 and H100 PCIe only. Listings that do not say SXM or PCIe, or that say NVL or HGX, cannot be matched to a printed column, so they are not priced here.
- A100 40GB
- The A100 40GB PCIe product brief prints no TFLOPS. We do not copy the 80 GB card’s figure.
- L4
- NVIDIA marks the L4 TFLOPS as sparse ("one-half lower without sparsity") and prints no dense figure.
- V100
- The V100 datasheet prints one unlabelled Tensor figure. Volta has no BF16 tensor mode.
- RTX 5090, 4090 and 3090
- The GeForce pages print AI TOPS without a precision or sparsity label and no BF16 TFLOPS.
- RTX PRO 6000
- The page prints "FP16 | BF16 Tensor Core 1 PFLOP" without saying whether it is sparse or dense.
- RTX 6000 Ada and RTX A6000
- The datasheets print one tensor figure with no BF16 label (the Ada figure is footnoted as using sparsity).
We never compute a dense figure from a sparse one, even where the maker says dense is half. If a maker later prints a dense figure we will add the GPU. NVIDIA's H100 figure comes from its Hopper architecture post, which labels the numbers preliminary.
Questions
Coverage: common questions
Why not halve the sparse figure?
Because that would be our computation, not a printed number. Where NVIDIA says dense is one-half of the sparse spec, it still prints only the sparse figure, and a chip that is not run sparse is not guaranteed to hit half.
Why is the H100 figure labelled preliminary?
NVIDIA’s own Hopper post prints 1000 TFLOPS dense BF16 for the SXM5 and 800 for the PCIe, in a table it calls preliminary performance specs, subject to change in shipping products.
Where are the figures stored?
In the site’s data files, each with its quote, source page and read date. Anyone can re-check them against the page.
Disclosure
Fine-Tune Cost Lab is published by the Runcrate team, who also run Runcrate, an AI cloud. Runcrate is listed in the table, marked "ours, live". It is sorted by price like every other row, with no pinned position. Other providers' prices are read from their own public pricing pages. We have no paid placement.