Memory
128 GB
HBM2e
Bandwidth
3.7 TB/s
TDP
900 W
Process
N5
TSMC
FP16
1,835 T
FP8
1,835 T
Perf / Watt
2.04 T/W
Gecko Score
57
Announced
2024-04-09
Available
2024-09-01
List price
$15,625
Cloud $/h
$10.42/h
Fabricated at
TSMC
N5 · die specs TBD
Memory generation
HBM2e · 128 GB
3.7 TB/s · 8 stacks · supplied by Samsung, SK hynix
Performance envelope
FP precision matrix · manufacturer datasheet values
| FP16 dense TFLOPs | 1,835 |
| BF16 dense TFLOPs | 1,835 |
| FP8 dense TFLOPs | 1,835 |
| FP4 dense TFLOPs | — |
| INT8 TOPs | 1,835 |
| Memory bandwidth (TB/s) | 3.7 |
| TDP (Watts) | 900 |
| FP16 TFLOPs per Watt | 2.04 |
Rentable on
2 cloud providers · prices refreshed from public APIs
Known buyers
Disclosed in SEC filings, earnings calls, and press releases
| Company | Disclosed quantity | Year |
|---|---|---|
| IBM | undisclosed · watsonx deployment | 2024 |
Geography
HQ and fab locations · feeds /countries
HQ country
USA · CaliforniaFab locations
Taiwan
Supply signal
Hand-curated · the same feed powering the AI Compute Demand Index
Live supply tightness for Gaudi 3 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.
Sources
Every spec on this page is reproducible