DEPLOYDatabase

Chip comparison

AWS Trainium 2 vs NVIDIA GB300 (Blackwell Ultra)

Side-by-side specs

Straight from each vendor's datasheet. Winner in bold: higher throughput, memory, or transistor count; lower power draw or price. Sparse (2:4) throughput roughly doubles the FP8/INT8 numbers shown.

FieldAWS Trainium 2NVIDIA GB300 (Blackwell Ultra)
Process nodeTSMC N5TSMC 4NP
TDP500 W1400 W
Memory96 GB HBM3288 GB HBM3e
Memory bandwidth2,900 GB/s8,000 GB/s
BF16 (dense)667 TFLOPS
FP8 (dense)1,300 TFLOPS5,000 TFLOPS
INT8 (dense)1,300 TOPS
Form factorAWS Neuron 2 (Trn2 UltraServer with 64 chips per unit)Superchip (Blackwell Ultra + Grace)
Announced2024-12-032025-03-18
Released2025-01-012025-11-01

Robots running AWS Trainium 2

No robots publicly running it yet.

Robots running NVIDIA GB300 (Blackwell Ultra)

No robots publicly running it yet.

Data centers with NVIDIA GB300 (Blackwell Ultra)

Common questions

AWS Trainium 2 vs NVIDIA GB300 (Blackwell Ultra): which is faster for training?

NVIDIA GB300 (Blackwell Ultra) has 3.85x the dense FP8 throughput of the other (AWS Trainium 2: 1,300 TFLOPS; NVIDIA GB300 (Blackwell Ultra): 5,000 TFLOPS). Real-world training speed also depends on memory bandwidth, interconnect, and software stack: see the spec diff for those inputs.

AWS Trainium 2 vs NVIDIA GB300 (Blackwell Ultra): which has more memory?

NVIDIA GB300 (Blackwell Ultra) carries more HBM (AWS Trainium 2: 96 GB HBM3; NVIDIA GB300 (Blackwell Ultra): 288 GB HBM3e). Memory capacity sets the largest model that fits on one chip without partitioning; memory bandwidth sets inference throughput.

AWS Trainium 2 vs NVIDIA GB300 (Blackwell Ultra): which is more power-efficient?

Dense FP8 TFLOPS per watt: AWS Trainium 2 2.60 (1,300 ÷ 500 W); NVIDIA GB300 (Blackwell Ultra) 3.57 (5,000 ÷ 1400 W). NVIDIA GB300 (Blackwell Ultra) wins at the die level; system-level efficiency also depends on cooling and interconnect.

AWS Trainium 2 vs NVIDIA GB300 (Blackwell Ultra): which is more widely used?

AWS Trainium 2: 0 robots and 3 data centers publicly running it. NVIDIA GB300 (Blackwell Ultra): 0 robots and 1 data center. Empty here means no operator has said publicly, not zero adoption in the market.

See also: every chip comparison · AWS Trainium 2 full page · NVIDIA GB300 (Blackwell Ultra) full page.