TX · Utils

Vast.ai GPU Accelerator Reference

Vast.ai GPUs with current on-demand marketplace pricing, release year, and peak BF16 FLOPs. Click Year, Memory, TFLOPS, or Vast $/hr to sort. Includes my RTX 3070 as a local baseline. $50 and $100 columns show hours:minutes from that spend at the listed $/hr.

vast.gpu.ref()

GPU	Year	Memory	TFLOPS	Vast $/hr	$50 time	$100 time	vs. RTX 3070
Loading…

Sorted by peak BF16 FLOPs (descending). Click Year, Memory, TFLOPS, or Vast $/hr column headers to re-sort (click again to reverse). TFLOPS are peak dense BF16 tensor (equals FP16 on Ampere+), same methodology as the Colab accelerator reference — no sparsity boost. Year is the product’s public launch / first availability year. Architecture labels derive from NVIDIA compute capability (or AMD generation). Vast $/hr is the lowest verified, rentable, on-demand, single-GPU offer on the marketplace as of Jul 28, 2026 ^* — storage, egress, and instance overhead are extra. Only GPUs with both a curated FLOPs figure and a live on-demand offer are listed (52 models on Vast’s 151-type catalog). RTX 3070 is highlighted as my local training baseline; the listed Vast rate is for price comparison only. Interruptible (spot) instances are often 50%+ cheaper but may be preempted.

* Vast marketplace rates gathered manually by cmc on July 28, 2026.

References