Bluesky Thread

Google’s TPU v7: Ironwood

April 09, 2025 View original thread

Google’s TPU v7: Ironwood

A massive hardware leap

- 4,614 TFLOPS per chip
- 256 or 9,216 chips per pod
- 192 GB HBM(emory) per chip @ 7.2 Tbps
- ICI (inter-chip interconnect) @ 1.2 Tbps
- SparseCore for ultra large embeddings

blog.google/products/goo...

The image is a bar graph titled “TPUv7 Peak FLOPS Performance”, comparing the peak floating-point operations per second (FLOPS) of various Tensor Processing Unit (TPU) versions. The x-axis represents different TPU versions, while the y-axis indicates peak FLOPS performance, measured in teraflops (TFLOPS). Each TPU version is represented by a vertical bar, with the height corresponding to its performance.

The bars show a clear progression in performance across TPU generations:
• TPUv4: Approximately 275 TFLOPS.
• TPUv5e: Around 197 TFLOPS, indicating a slight decrease from v4.
• Trillium (TPUv6): Significant increase to approximately 926 TFLOPS.
• TPUv7: Further improvement, reaching around 1,000 TFLOPS.

The graph highlights the substantial advancements in TPU performance, especially with the introduction of Trillium (TPUv6) and TPUv7. These newer versions demonstrate Google’s commitment to enhancing AI processing capabilities, with TPUv7 achieving a notable increase in peak FLOPS performance.

21 8

Google is on a steep upward trajectory

Gemini 2.5 is now the favored coding model, above sonnet-3.7, and now with TPU v7, they feel like the only provider that can really hold up under expanding compute demands

More like this