Trainium3 specs
AWS · Shipping now · first available Dec 2025. Dense compute, memory, bandwidth and power from the vendor's documents; each precision stays separate.
All accelerators · Compare with Trainium2
Specifications
| Dense BF16 · FP32 accumulate | 671 TFLOPS |
|---|---|
| Dense FP8 · FP32 accumulate | 2,517 TFLOPS |
| Dense FP4 · FP32 accumulate | 2,517 TFLOPS |
| Memory | 154.62 GB · HBM144 GiB as published |
| Memory bandwidth | 4.9 TB/s |
| Maximum chip power | Not published |
| Interconnect | NeuronLink v4 · 2.56 TB/s |
| Process node | 3nm |
| First availability | Dec 2025 |
| Availability | Shipping |
Per-chip NeuronCore-v4 specifications: MXFP8 and MXFP4 share the published throughput; dense BF16 is lower and must not be inferred by halving FP8.
Dense tensor rates with FP32 accumulation, never sparse or FP16-accumulate headline figures. Memory is in decimal GB. The est marker flags a value its vendor-document review did not confirm.
Rent it
No tracked rental quote. The site tracks providers' published list prices; this part may still be rentable under terms it does not record.
Value ratios
| $ per dense BF16 PFLOP-hour | No on-demand quote |
|---|---|
| Dense BF16 TFLOPS per watt | No power figure |
| Bandwidth bytes per 1,000 BF16 FLOP | 7.30 |
| $ per dense FP8 PFLOP-hour | No on-demand quote |
| Dense FP8 TFLOPS per watt | No power figure |
| Bandwidth bytes per 1,000 FP8 FLOP | 1.95 |
| $ per GB of memory per hour | No on-demand quote |
Prices divide the cheapest on-demand list price; theoretical ratios do not measure application speed.
Sources
Specifications reviewed . 671 TFLOPS dense BF16; 2,517 TFLOPS MXFP8 and MXFP4. Dense BF16 cannot be inferred by halving FP8; AWS publishes no per-chip power.
First availability: Amazon: Trainium3 UltraServers now available ↗ (2025-12-02): “general availability of Trainium3 UltraServers”