The main differences in cost between TPU v2 and TPU v3 lie in their hourly pricing and overall performance.
- TPU v2: The cost for a TPU v2 core is approximately $6.50 per hour as of 2019[3]. For larger configurations like a TPU v2 pod, the cost can be significantly higher, with a 512-core pod costing $384 per hour[5].
- TPU v3: The cost for a TPU v3 core is around $8.50 per hour as of 2019[3]. More recent pricing indicates that a full TPU v3 costs about $4.50 per hour[1]. For larger setups, the costs scale accordingly, but specific pod pricing for TPU v3 is slightly higher than the v2 on an hourly basis for smaller configurations[5].
In general, TPU v3 offers more processing power (123 TOPS) compared to TPU v2 (45 TOPS)[2], which can lead to faster computation times and potentially lower overall costs for large-scale machine learning tasks despite the higher hourly rate. However, the choice between TPU v2 and v3 depends on specific project needs and budget constraints.
Citations:[1] https://www.datacamp.com/blog/tpu-vs-gpu-ai
[2] https://community.fs.com/encyclopedia/-tensor-processing-unit-tpu-.html
[3] https://www.dataknobs.com/generativeai/tpu/tpu-gpu-cost.html
[4] https://cloud.google.com/tpu/docs/quota
[5] https://www.reddit.com/r/MachineLearning/comments/epx5vg/d_google_cloud_tpu_pod_pricing_grid_a_512core_tpu/
[6] https://openmetal.io/docs/product-guides/private-cloud/tpu-vs-gpu-pros-and-cons/
[7] https://news.ycombinator.com/item?id=22195516
[8] https://eng.snap.com/training-models-with-tpus