Amazon EC2 Inf1 Instances - first-generation Inferentia chip specs and instance table

READ 2026-10-05 for the first-generation Inferentia record, which did not previously exist. VERBATIM: 'Each AWS Inferentia chip has four first-generation NeuronCores and provides up to 128 tera operations per second (TOPS) of performance and support for FP16, BF16, and INT8 data types.' and 'They feature up to 16 AWS Inferentia chips'. Instance table (Product details) verbatim columns: Instance size / vCPUs / Memory (GiB) / Instance Storage / Inferentia chips / Inferentia chip-to-chip interconnect / Network bandwidth (Gbps) / EBS bandwidth / On-demand price / 1yr RI / 3yr RI. Rows: inf1.xlarge 4 / 8 / EBS only / 1 / N/A / up to 25 / up to 4.75 / $0.228 / $0.137 / $0.101; inf1.2xlarge 8 / 16 / EBS only / 1 / N/A / up to 25 / up to 4.75 / $0.362 / $0.217 / $0.161; inf1.6xlarge 24 / 48 / EBS only / 4 / Yes / 25 / 4.75 / $1.180 / $0.709 / $0.525; inf1.24xlarge 96 / 192 / EBS only / 16 / Yes / 100 / 19 / $4.721 / $2.835 / $2.099. WHAT THIS PAGE DOES NOT PUBLISH, and it matters: there is NO memory-bandwidth figure anywhere on it, no accelerator-memory capacity column at all (the Memory (GiB) column is HOST memory: inf1.xlarge has 8 GiB of host memory AND one Inferentia chip, so the host-memory column must not be read as the chip's memory), no per-chip HBM capacity, and no bus width. It also publishes no FP16 TFLOPS figure - only the 128 TOPS INT8-class figure. NEGATIVE FINDING, recorded so a future pass does not go looking again: 'N/A' in the chip-to-chip interconnect column for inf1.xlarge and inf1.2xlarge means those single-chip shapes have no chip-to-chip link; it is not a missing value.