Amazon EC2 Trn1 Instances - first-generation Trainium chip specs and instance table

READ 2026-10-05 for the first-generation Trainium record, which did not previously exist (the repo referenced trn1 in prices and in Neuron runtime notes but had no accelerator record). VERBATIM: 'Trn1 instances are powered by up to 16 AWS Trainium chips purpose built to accelerate DL training and deliver up to 3 petaflops of FP16/BF16 compute power. Each chip includes two second-generation NeuronCores.' and 'each Trn1 instance has 512 GB of shared accelerator memory (HBM) with 9.8 TB/s of total memory bandwidth' and 'Trn1 instances support up to 768 GB/s of NeuronLink, a high-speed, nonblocking interconnect'. Data types verbatim: 'FP32, TF32, BF16, FP16, UINT8, and the new configurable FP8 (cFP8)'. Product details table verbatim: trn1.2xlarge 1 chip / 32 GB / 8 vCPU / $1.34 on-demand; trn1.32xlarge 16 chips / 512 GB / 128 vCPU / 800 Gbps / $21.50; trn1n.32xlarge 16 chips / 512 GB / 128 vCPU / 1600 Gbps / $24.78. THE FIGURE THAT MATTERS FOR A PER-CHIP RECORD: the 9.8 TB/s and 3 PFLOPS are PER-INSTANCE AGGREGATES, not per-chip. Dividing them by the 16 chips of a .32xlarge gives 32 GB (which agrees exactly with the per-chip capacity the trn1.2xlarge row prints) and 612.5 GB/s and 187.5 TFLOPS. The 612.5 GB/s is BELOW the 0.8 TB/s per-chip figure AWS publishes in the Neuron Trainium2 architecture page's own Trainium-vs-Trainium2 comparison table (source w7z-neuron-trainium2-arch-page), and the 187.5 is below that table's 191. The gap is not an error: an instance aggregate divided by chip count yields an AVERAGE the instance sustains, not the chip's peak. Both numbers are kept in the record and the distinction is named, per the repo's disagreement rule.