trn1.32xlarge

Analysis

The trn1.32xlarge is a machine learning asic instances instance in the trn1 series with 128 vCPUs and 512 GiB of memory on Intel Xeon Scalable (Icelake) (up to Up to 3.5 GHz), which makes it balanced at 4 GiB of memory per vCPU.

Nothing else with 128 vCPUs and similar memory is cheaper: among 2 current same-shape types it has the lowest on-demand price.

It carries 16 GPUs with 512 GB of GPU memory, so most of the price is the accelerator rather than the CPUs. It has local instance storage (4 x 1900 GB NVMe SSD), which is fast but lost when the instance stops. Network: 800 Gigabit.

Its price is almost flat across the 3 regions that offer it ($21.50 to $21.50/hour), so latency and data residency should decide where it runs. It is not offered in Canada Central, São Paulo, Ireland, London, Frankfurt, Mumbai, Singapore, Sydney and Tokyo, so plan around that if you need a presence there.

It is the cheapest of the 2 processor variants AWS sells at this size and generation. The trn1n.32xlarge on Intel Xeon Scalable (Icelake) costs 15% more, and no others also sit above it.

Per vCPU it is one of the dearer machine learning asic instances types: 77% of the 13 current types in the family cost less than its $0.1680 per vCPU-hour. The premium mostly pays for the accelerators. On demand it is $21.50/hour in US East (N. Virginia), or $15,695 for a 730-hour month.