p3dn.24xlarge

Analysis

The p3dn.24xlarge is a gpu instance instance in the p3dn series with 96 vCPUs and 768 GiB of memory on Intel Xeon Platinum 8175 (Skylake) (up to 2.5 GHz), which makes it memory-heavy at 8 GiB of memory per vCPU.

Running it in N. Virginia instead of Tokyo saves $8,447 a month per instance: $31.21/hour against $42.78/hour, a 37% spread. It is not offered in Ohio, Canada Central, São Paulo, London, Frankfurt, Mumbai, Singapore and Sydney, so plan around that if you need a presence there.

It carries 8 GPUs with 256 GiB of GPU memory, so most of the price is the accelerator rather than the CPUs. It has local instance storage (2 x 900 GB NVMe SSD), which is fast but lost when the instance stops. Network: 100 Gigabit.

Per vCPU it is one of the dearer gpu instance types: 87% of the 78 current types in the family cost less than its $0.3251 per vCPU-hour. The premium mostly pays for the accelerators. On demand it is $31.21/hour in US East (N. Virginia), or $22,785 for a 730-hour month.

Looking beyond its own line, 2 of 3 current types with 96 vCPUs and similar memory are cheaper. The cheapest of those is the dl2q.24xlarge on Intel Xeon Platinum 8259 (Cascade Lake) at $8.92/hour.