Enjoy the "big ferocious" performance of NVIDIA's Ampere-based RTX 3090 for less than $1 an hour. Supplies are limited so reserve one while you can.
The RTX 3090 delivers over 35 Tera-FLOPs of single precision (fp32) math, making it nearly as fast as the new A100. It also features 24 GB of GPU memory, enabling support for larger transformer models. Up to 4 GPUs can be allocated to a single notebook or worker, giving you up to 96 GB of GPU memory for a single training instance. All RTX 3090 instances have a minimum of 4 CPU cores (8 threads) and 64 GB of CPU memory per GPU. GPUs are connected at a minimum of 8 PCIe 4.0 lanes to ensure high speed data exchange. All NVMe local storage is connected with PCIe 4.0 bandwidth as well.