Google Colab Accelerator Reference

Order: local RTX baseline, then Colab TPUs by peak FLOPs (fastest → slowest), then Colab GPUs the same way. FLOPs are dense BF16 tensor peak (equal to FP16 on Ampere+ GPUs; native BF16 on TPU) — no sparsity boost. TPU rows are Colab v*-1 single-chip slices (not a full pod). Colab cost is approximate compute units (credits) per hour; ~$0.10 per CU via Pay As You Go / Pro packs. $10 time is hours:minutes from a monthly $10 pack (100 CU) at that burn rate. Rates can vary by region and demand — check Colab’s Resources panel for the live burn rate. ^* A100 and T4 are listed separately for standard and High-RAM tiers (same peak FLOPs, different credit rates). Colab memory can be slightly lower than datasheet capacity (e.g. T4 ~15 GB usable). Real training throughput also depends on memory bandwidth, batch size, framework, and interconnect.

* Colab compute-unit rates were gathered manually by cmc on July 27, 2026.