Drag to turn the package
Obsidian G3
Sampling
A reticle-limit training accelerator with stacked HBM sitting directly on the interposer beside the compute die.
- Process node
- 5nm · Dense
- Category
- AI Training Accelerator
- Cores
- 128 tensor cores
- Package
- 2.5D w/ HBM stack
- Power
- 350W TDP
- Die size
- 700 mm² + HBM stack
Datasheets are shared under NDA. Applications engineering typically responds within one business day.
About this part
Obsidian is our largest die, full stop — close to the reticle limit, with high-bandwidth memory stacks placed directly on the same interposer rather than routed out to separate packages. That proximity is most of the performance story: training workloads are as often memory-bound as compute-bound, and the interposer route is dramatically shorter than going through a PCB.
The harder problem was thermal, not electrical. A die this size under a 350W load needs a vapor-chamber lid custom-fitted to the package outline, and two of our five sampling rounds were spent purely on lid contact-pressure tuning, not silicon.
Sampling now with select training-cluster customers; general sampling opens once the current thermal validation round closes.