1. Chart data
  2. DeepSeek 670B BF16 Training Performance - GB200 NVL72 vs H10
SemiAnalysisTable chart · 120 points · SemiAnalysis

DeepSeek 670B BF16 Training Performance - GB200 NVL72 vs H100

The numbers behind a chart SemiAnalysis published, redrawn natively. Check the original, study the data, then make your own.

Estimated from the image 41%All chart data
Chart by SemiAnalysis. Source: x.com

Simplified redraw

Drawn by us from the numbers below. Hover, tap or use the arrow keys.

A simplified redraw: it shows the numbers in our own look, not the look of the original. One accent marks what the chart is about. Values marked with ~ in the table are estimates.

The numbers

First 10 of 120 points are free
GroupSeriesValueSourceConf.
GPU SKUUnits~0Estimate0.80
GPU SKUH100~0Estimate0.30
GPU SKUGB200 NVL72~0Estimate0.30
GPU SKUGB200 NVL72~0Estimate0.30
GPU SKUGB200 NVL72~0Estimate0.30
GPU SKUH100~0Estimate0.30
GPU SKUGB200 NVL72~0Estimate0.30
GPU SKUGB200 NVL72~0Estimate0.30
Global Batch SizeUnits~0Estimate0.30
Global Batch SizeH100~8192Estimate0.30

Max data, free with your email for now

See all 120 points and download the files

The full table, the JSON, a CSV and the agent skill file for this chart. These are Max in the plan; while this is a prototype your email unlocks them.

Unlocks all 35 giants. One email a week, leave any time. What you get · Log in

Global Batch SizeGB200 NVL72~1024Estimate0.30
Global Batch SizeGB200 NVL72~2048Estimate0.30
Global Batch SizeGB200 NVL72~2048Estimate0.30
Global Batch SizeH100~8192Estimate0.30
Global Batch SizeGB200 NVL72~4096Estimate0.30
Global Batch SizeGB200 NVL72~8192Estimate0.30
Token/s/GPUUnits~0Estimate0.30
Token/s/GPUH100~630Estimate0.30
Token/s/GPUGB200 NVL72~956Estimate0.30
Token/s/GPUGB200 NVL72~958Estimate0.30
Token/s/GPUGB200 NVL72~959Estimate0.30
Token/s/GPUH100~842Estimate0.30
Token/s/GPUGB200 NVL72~2138Estimate0.30
Token/s/GPUGB200 NVL72~3957Estimate0.30
BF16 Model Flops Utilization (MFU)Units~0Estimate0.30
BF16 Model Flops Utilization (MFU)H100~16.6Estimate0.30
BF16 Model Flops Utilization (MFU)GB200 NVL72~10.1Estimate0.30
BF16 Model Flops Utilization (MFU)GB200 NVL72~10.17Estimate0.30
BF16 Model Flops Utilization (MFU)GB200 NVL72~10.18Estimate0.30
BF16 Model Flops Utilization (MFU)H100~22Estimate0.30
BF16 Model Flops Utilization (MFU)GB200 NVL72~22.7Estimate0.30
BF16 Model Flops Utilization (MFU)GB200 NVL72~42Estimate0.30
BF16 Cost Per Million TokensUnits~0Estimate0.30
BF16 Cost Per Million TokensH100~0.468Estimate0.30
BF16 Cost Per Million TokensGB200 NVL72~0.166Estimate0.30
BF16 Cost Per Million TokensGB200 NVL72~0.307Estimate0.30
BF16 Cost Per Million TokensGB200 NVL72~0.684Estimate0.30
BF16 Cost Per Million TokensH100~0.626Estimate0.30
BF16 Cost Per Million TokensGB200 NVL72~0.684Estimate0.30
BF16 Cost Per Million TokensGB200 NVL72~0.686Estimate0.30
BF16 All-In Utility Joules per TokenUnits~0Estimate0.30
BF16 All-In Utility Joules per TokenH100~2.12Estimate0.30
BF16 All-In Utility Joules per TokenGB200 NVL72~0.53Estimate0.30
BF16 All-In Utility Joules per TokenGB200 NVL72~0.97Estimate0.30
BF16 All-In Utility Joules per TokenGB200 NVL72~2.17Estimate0.30
BF16 All-In Utility Joules per TokenH100~2.84Estimate0.30
BF16 All-In Utility Joules per TokenGB200 NVL72~2.17Estimate0.30
BF16 All-In Utility Joules per TokenGB200 NVL72~2.17Estimate0.30
BF16 Tokens per annual US HH energy useUnits~0Estimate0.30
BF16 Tokens per annual US HH energy use¹H100~13.68Estimate0.30
BF16 Tokens per annual US HH energy use¹GB200 NVL72~17.88Estimate0.30
BF16 Tokens per annual US HH energy use¹GB200 NVL72~17.92Estimate0.30
BF16 Tokens per annual US HH energy use¹GB200 NVL72~17.93Estimate0.30
BF16 Tokens per annual US HH energy use¹H100~18.3Estimate0.30
BF16 Tokens per annual US HH energy use¹GB200 NVL72~39.97Estimate0.30
BF16 Tokens per annual US HH energy use¹GB200 NVL72~73.99Estimate0.30
GPU SKUUnits~0Estimate0.50
GPU SKUH100~0Estimate0.50
GPU SKUGB200 NVL72~0Estimate0.50
GPU SKUGB200 NVL72~0Estimate0.50
GPU SKUGB200 NVL72~0Estimate0.50
GPU SKUH100~0Estimate0.50
GPU SKUGB200 NVL72~0Estimate0.50
GPU SKUGB200 NVL72~0Estimate0.50
Global Batch SizeUnits~0Estimate0.50
Global Batch SizeH100~0Estimate0.50
Global Batch SizeGB200 NVL72~0Estimate0.50
Global Batch SizeGB200 NVL72~0Estimate0.50
Global Batch SizeGB200 NVL72~0Estimate0.50
Global Batch SizeH100~0Estimate0.50
Global Batch SizeGB200 NVL72~0Estimate0.50
Global Batch SizeGB200 NVL72~0Estimate0.50
Token/s/GPUUnits~0Estimate0.50
Token/s/GPUH100~0Estimate0.50
Token/s/GPUGB200 NVL72~0Estimate0.50
Token/s/GPUGB200 NVL72~0Estimate0.50
Token/s/GPUGB200 NVL72~0Estimate0.50
Token/s/GPUH100~0Estimate0.50
Token/s/GPUGB200 NVL72~0Estimate0.50
Token/s/GPUGB200 NVL72~0Estimate0.50
BF16 Model Flops Utilization (MFU)Units~0Estimate0.50
BF16 Model Flops Utilization (MFU)H100~0Estimate0.50
BF16 Model Flops Utilization (MFU)GB200 NVL72~0Estimate0.50
BF16 Model Flops Utilization (MFU)GB200 NVL72~0Estimate0.50
BF16 Model Flops Utilization (MFU)GB200 NVL72~0Estimate0.50
BF16 Model Flops Utilization (MFU)H100~0Estimate0.50
BF16 Model Flops Utilization (MFU)GB200 NVL72~0Estimate0.50
BF16 Model Flops Utilization (MFU)GB200 NVL72~0Estimate0.50
BF16 Cost Per Million TokensUnits~0Estimate0.50
BF16 Cost Per Million TokensH100~0Estimate0.50
BF16 Cost Per Million TokensGB200 NVL72~0Estimate0.50
BF16 Cost Per Million TokensGB200 NVL72~0Estimate0.50
BF16 Cost Per Million TokensGB200 NVL72~0Estimate0.50
BF16 Cost Per Million TokensH100~0Estimate0.50
BF16 Cost Per Million TokensGB200 NVL72~0Estimate0.50
BF16 Cost Per Million TokensGB200 NVL72~0Estimate0.50
BF16 All-In Utility Joules per TokenUnits~0Estimate0.50
BF16 All-In Utility Joules per TokenH100~0Estimate0.50
BF16 All-In Utility Joules per TokenGB200 NVL72~0Estimate0.50
BF16 All-In Utility Joules per TokenGB200 NVL72~0Estimate0.50
BF16 All-In Utility Joules per TokenGB200 NVL72~0Estimate0.50
BF16 All-In Utility Joules per TokenH100~0Estimate0.50
BF16 All-In Utility Joules per TokenGB200 NVL72~0Estimate0.50
BF16 All-In Utility Joules per TokenGB200 NVL72~0Estimate0.50
BF16 Tokens per annual US HH energy useUnits~0Estimate0.50
BF16 Tokens per annual US HH energy useH100~0Estimate0.50
BF16 Tokens per annual US HH energy useGB200 NVL72~0Estimate0.50
BF16 Tokens per annual US HH energy useGB200 NVL72~0Estimate0.50
BF16 Tokens per annual US HH energy useGB200 NVL72~0Estimate0.50
BF16 Tokens per annual US HH energy useH100~0Estimate0.50
BF16 Tokens per annual US HH energy useGB200 NVL72~0Estimate0.50
BF16 Tokens per annual US HH energy useGB200 NVL72~0Estimate0.50
GPU SKU#~0Estimate0.50
Number of GPUsH100~1024Estimate0.50
Number of GPUsGB200 NVL72~128Estimate0.50
Number of GPUsGB200 NVL72~256Estimate0.50
Number of GPUsGB200 NVL72~512Estimate0.50
Number of GPUsH100~1024Estimate0.50
Number of GPUsGB200 NVL72~256Estimate0.50
Number of GPUsGB200 NVL72~1024Estimate0.50
Download this chart's data

Max subscribers get the same data by API. API docs

JSONAgent skill file

API · Max

Use this chart via API

On the Max plan you can pull this chart's numbers with the provenance of each one, as JSON or CSV, with its redraw spec, or call it from an agent over MCP. Chart id 65d06ddd5bcd.

Approximate values. Our two readings of this chart agreed only in part (confidence 0.41 of 1), so check the original before you quote a number. The API marks it lowConfidence, and ?min_confidence=0.5 leaves it out.

See the Max planAPI docs
PathReturnsPlan
GET/api/v1/charts/65d06ddd5bcdTitle, publisher, date, form, point count; with a Max key, the data linkOpen
GET/api/v1/charts/65d06ddd5bcd/dataEvery point with its provenance and confidence, plus the redraw spec. Add ?format=csv for CSVMax
GET/api/v1/charts/65d06ddd5bcd/render?view=1The draw-only view this page usesOpen
POST/mcpTool get_chart_data with this id, for Claude, Cursor and other agentsMax
curlMax
curl -s https://viral-chart.com/api/v1/charts/65d06ddd5bcd/data \
  -H "Authorization: Bearer $VIRALCHART_KEY"
Claude CodeMax
claude mcp add --transport http viralchart https://viral-chart.com/mcp \
  --header "Authorization: Bearer $VIRALCHART_KEY"

60 requests a minute per key; over that you get 429 with Retry-After. Up to 5 keys per account, made on your account page and shown once. Without Max, the data route answers 402 with a link to the plans. The numbers describe a chart its publisher owns, and every file carries a license note.

Make one like this

This chart's shape, your numbers.

Live preview. Everything stays on your device.

Pick the form

A simplified form: our own look, not the look of the original. Your numbers start from this chart's data.

More data from SemiAnalysis

All chart data
Estimated from the imageINTC Royalty Payments as % of NVDA ProfitsSemiAnalysis · Grouped bar chart · 21 pointsEstimated from the imageTCO $/GPU/Hr: Total Tokens (Owning - Hyperscaler) Over Time at 100 tok/s/userSemiAnalysis · Line chart · 14 pointsEstimated from the imageNVIDIA Rubin GPU ArchitectureSemiAnalysis · Table chart · 10 pointsEstimated from the imageRubin CPX: DeepSeek v3 using Pipeline ParallelismSemiAnalysis · Table chart · 9 pointsEstimated from the imageThe same $200 Claude plan is worth $2.5K to $12.5K in API valueSemiAnalysis · Grouped bar chart · 6 pointsEstimated from the imageGLM5.2/GLM5.3 744B Agentic Total Tokens per $1 TCO vs. P90 InteractivitySemiAnalysis · Scatter plot · 18 points

Questions people ask

How accurate are these numbers?

They are estimated from the image by two independent passes, with confidence 0.41 on a 0 to 1 scale. Printed numbers read well; values read off an axis can be a few percent off.

Can I reuse this data?

The numbers describe a chart that SemiAnalysis published. The image, layout and styling belong to them. Use the numbers with a credit to the publisher, and make your picture from your own data.

How do I get the file?

The JSON, CSV and agent skill file are under the email gate on this page. In this prototype they are static files at /api/charts/65d06ddd5bcd.json and /api/charts/65d06ddd5bcd.skill.md.