p6-b300.48xlarge

Analysis

The p6-b300.48xlarge is a gpu instance instance in the p6-b300 series with 192 vCPUs and 4096 GiB of memory on Intel Xeon Scalable (Emerald Rapids) (up to 3.2 GHz), which makes it very memory-heavy at 21.33 GiB of memory per vCPU.

Running it in N. Virginia instead of São Paulo saves $101,261 a month per instance: $142/hour against $281/hour, a 97% spread. It is not offered in Ohio, Canada Central, Ireland, Frankfurt, Mumbai, Singapore, Sydney and Tokyo, so plan around that if you need a presence there.

Within the same generation and size, the p6-b200.48xlarge on Intel Xeon Scalable (Emerald Rapids) costs $114/hour, 20% less ($20,793 a month).

It carries 8 GPUs with 2144 GB HBM3e of GPU memory, so most of the price is the accelerator rather than the CPUs. It has local instance storage (8 x 3840 GB NVMe SSD), which is fast but lost when the instance stops. Network: 6400 Gigabit.

Per vCPU it is one of the dearer gpu instance types: 99% of the 78 current types in the family cost less than its $0.7418 per vCPU-hour. The premium mostly pays for the accelerators. On demand it is $142/hour in US East (N. Virginia), or $103,964 for a 730-hour month.