Beta
NVIDIA logo

H100

by NVIDIA · Released Oct 2022 · Hopper

MainstreamGPUN4 · TSMC
55
Gecko score
1.41 TFLOPs/W
Memory
80 GB
HBM3
Bandwidth
3.35 TB/s
TDP
700 W
Process
N4
TSMC
FP16
989 T
FP8
1,979 T
Perf / Watt
1.41 T/W
Gecko Score
55
Announced
2022-03-22
Available
2022-10-01
List price
$30,000
Cloud $/h
$1.99/h

FP precision matrix · manufacturer datasheet values

FP16 dense TFLOPs989
BF16 dense TFLOPs989
FP8 dense TFLOPs1,979
FP4 dense TFLOPs
INT8 TOPs1,979
Memory bandwidth (TB/s)3.35
TDP (Watts)700
FP16 TFLOPs per Watt1.41

4 cloud providers · prices refreshed from public APIs

Disclosed in SEC filings, earnings calls, and press releases

CompanyDisclosed quantityYear
Meta~350k (2024)2024
Microsoft~485k (2023-24)2024
Models trained or served
Known to ship on this chip
Rack systems
Where this chip ships at scale
Datacenters
Deployed clusters tracked by BenchGecko
Browse energy + datacenters

What came before · what comes next

HQ and fab locations · feeds /countries

Fab locations
Taiwan

Hand-curated · the same feed powering the AI Compute Demand Index

Live supply tightness for H100 is tracked on the Chip Demand Ladder and rolls up into the AI Compute Demand Index. Updated weekly from hyperscaler order books, earnings calls, and secondary-market signals.

Every spec on this page is reproducible

Other chips from the same maker