A.I.
Hardware · Chips
Trainium3
Trainium3 is the AI chip that follows Trainium2, designed by Annapurna Labs, Amazon's in-house chip team, for AWS, Amazon's cloud business. AWS calls it its fourth-generation AI chip and its first built on a 3-nanometer process. AWS announced Trn3 UltraServers, the rentable systems built on it, in December 2025 (AWS, What's New).
The chip
| Figure | One Trainium3 chip | Trainium2 |
|---|---|---|
| Compute, MXFP8 or MXFP4 | 2.52 PFLOPS | 1.3 PFLOPS FP8 |
| Compute, BF16 | 0.67 PFLOPS | 0.67 PFLOPS |
| Memory | 144 GiB HBM3e | 96 GiB |
| Memory bandwidth | 4.9 TB/s | 2.9 TB/s |
| Chip-to-chip link | 2.56 TB/s (NeuronLink-v4) | 1.28 TB/s |
Figures are from AWS's Neuron documentation (Trainium3 architecture). BF16 is a 16-bit number format used heavily in training. NeuronLink is AWS's own link between chips. AWS's product pages print the memory as 144 GB, while the Neuron docs say 144 GiB.
The UltraServer
| Figure | One Trn3 UltraServer |
|---|---|
| Chips | Up to 144 |
| Compute | Up to 362 PFLOPS MXFP8 |
| Memory | Up to 20.7 TB HBM3e |
| Memory bandwidth | 706 TB/s combined |
An UltraServer is a group of servers wired as one machine. UltraServers then join into EC2 UltraClusters 3.0, which AWS says scale to hundreds of thousands of chips (AWS, Trn3).
What AWS claims
- Up to 4.4 times the performance, 3.9 times the memory bandwidth, and 4 times the performance per watt of Trn2 UltraServers. These are AWS's own figures.
Units on this page: a PFLOPS is a thousand trillion math operations per second; an EFLOPS is a thousand PFLOPS. FP4 and FP8 store each number in 4 or 8 bits, and MX in front (MXFP4, MXFP8) means a shared-scale version of that format. GB is gigabytes, GiB is about 7% more than a GB, and TB is a thousand GB. TB/s and PB/s are terabytes and petabytes per second, how fast data moves. W is watts and kW is a thousand watts. TDP is the most power a chip is built to draw.