
Google unveils two new AI chips to make it less dependent on other companies
Google has introduced TPU 8tx and TPU 8i Max, two additions to its eighth-generation TPU line. TPU 8tx was designed with AMD, is based on TPU 8t, and is in
Tag archive

Google has introduced TPU 8tx and TPU 8i Max, two additions to its eighth-generation TPU line. TPU 8tx was designed with AMD, is based on TPU 8t, and is in
This is the log of a few days spent running Captum's attribution methods on Cloud TPUs through...

A Rust crate reads the checkpoint, runs all 35 layers of Gemma 4 E2B and drives the decode loop on one TPU v5e chip in Colab, with every operation a TorchTPU kernel and the same tokens as compiled Python.

Gemma 4 E2B served three ways on one Cloud TPU v5e chip: a pure-JAX engine in Python, the same graph written in Rust and compiled by the libtpu that JAX loads, and vLLM, all timed by one client. bf16 beats 4-bit in the JAX engine, the Rust engine matches the JAX engine's tokens, and over HTTP it decodes at vLLM's rate with the fastest first token.
A PyTorch model can be portable in the way a recipe is portable: the ingredients and steps look...

vLLM, SGLang-JAX, a pure-JAX engine, a Rust model over PJRT, and TorchTPU in Python and Rust serve Gemma 4 E2B on one flex-start v5e-1, with the same prompts, the same 1-to-64-request sweep and the same GSM8K questions. vLLM is the engine that serves out of the box; SGLang-JAX with an E2B patch is the fastest at every load from 8 requests up.

SGLang-JAX serves the Gemma 4 31B and 26B A4B on TPU and fails to load E2B. A two-file patch adds per-layer embeddings, shared K/V and the double-wide MLP; on one flex-start v5e-1 it then serves 159.9 tokens per second for one request and 4,249.4 at 64, 2.52x vLLM, with GSM8K level with the bf16 JAX engine.

Gemma 4 E2B with the model in Rust, served from a TPU v5e-1 VM through PyO3 and a C++ shim over ATen, with no tch and no pinned PyTorch, an OpenAI endpoint, an MCP server and a local chat UI.

Google Cloud is addressing the increasing demands for memory, security, and performance by adding two...

Originally published on The Daily Flare. Key takeaways Google introduced TorchTPU to make its AI...

A space data center can't use air or water: heat must radiate away. Google's first Suncatcher satellite runs its four TPUs about 15 minutes, then cools down.

Five Colab notebooks take Gemma 4 E2B on one v5e chip from 5 tokens per second under Hugging Face generate to 50 with one compiled decode step, and to the full 131,072-token context with a ring cache, against a JAX engine on the same chip.