Compare · GPUs & AI chips

Blackwell Ultra B300 SXM vs Trainium2

RANKED BY DENSE FP16/BF16 TENSOR TFLOPS PER CHIP #4 VS #12 CHECKED 06 SEPT 2026

Side by side

Verdict

Blackwell Ultra B300 SXM ranks #4 of 20 on Dense FP16/BF16 tensor TFLOPS per chip (2.25 PFLOPS); Trainium2 ranks #12 (667 TFLOPS).

Rank 04 of 20 · leads

Blackwell Ultra B300 SXM

Dense FP16/BF16 tensor TFLOPS per chip

2.25 PFLOPS

MakerNVIDIA

Date2025

Verified04 Sept 2026

Evidence

NVIDIA HGX Platform page, HGX B300 row ("8x NVIDIA Blackwell Ultra SXM", shipping now). FP16/BF16 Tensor Core is the same 36 PFLOPS sparse / 18 PFLOPS dense board total as B200, so 2.25 PFLOPS dense per chip. B300's advertised uplift versus B200 is at FP4 and HBM capacity, not at dense FP16/BF16 — so on this metric B200 and B300 are tied.

Rank 12 of 20

Trainium2

Dense FP16/BF16 tensor TFLOPS per chip

667 TFLOPS

MakerAWS

Date2024

Verified04 Sept 2026

Evidence

AWS Neuron Trainium2 Architecture page, compute table: each Trainium2 chip delivers "667 BF16/FP16/TF32 TFLOPS" dense; sparse counterpart is 2,563 TFLOPS and was not used. 96 GiB device memory. Trn2 instances are generally available (AWS announcement 3 Dec 2024).

Try another pair

More Blackwell Ultra B300 SXM matchups

Source Vendor product pages and official spec documents (AMD Instinct MI355X / MI350X / MI325X / MI300X / MI300A / MI250X / MI250 / MI210; NVIDIA… Last checked 06 Sept 2026

Full GPUs & AI chips ranking Spot an error? Send a tip