Google’s TPUv7 Ironwood Achieves 50% Better Performance Per Dollar Than NVIDIA’s Blackwell Ultra, Says Semi Analysis

SemiAnalysis benchmarks show Google's TPUv7 Ironwood delivers up to 50% better performance per dollar than NVIDIA's Blackwell Ultra in inference tests, marking one of the first independent verifications of Google's cost-efficiency advantage. The gains stem from Ironwood's SparseCore, which offloads mixture-of-experts data routing, plus Google's new TorchTPU software and Pallas inference kernels that ease PyTorch-based development on TPUs.

Google could be preparing to make its TPU chips available for external users. Semiconductor research firm SemiAnalysis has published fresh benchmarks showing that Google’s TPUv7, codenamed Ironwood, delivers up to 50% better performance per dollar than NVIDIA’s flagship Blackwell Ultra chip, when tested under similar, apples-to-apples conditions on SemiAnalysis’s own inference benchmarking platform, InferenceX. For anyone who hasn’t been following the AI hardware race closely, here’s the quick version: NVIDIA’s GPUs are the chips that have powered almost the entire generative AI boom, and demand for them has made NVIDIA the world’s most valuable company. Google, meanwhile, has spent over a decade quietly building its own alternative called the Tensor Processing Unit, or TPU, largely for its own internal use. That equation has been changing fast lately, with Google’s TPUs finding growing external interest https://officechai.com/ai/anthropic-google-tpu-deal/ from AI labs looking for alternatives to NVIDIA’s chips. What SemiAnalysis Actually Found Companies serving AI models to users care about two things above all else: how fast the chip can respond, and how much it costs to run. SemiAnalysis measured both by running large language models on Ironwood and comparing the results to Blackwell Ultra chips under matching precision settings, so that the comparison was fair rather than stacking a stronger setting on one side. The result was that Ironwood consistently produced more useful output for every dollar spent than Blackwell Ultra did, in some configurations by as much as half again. The gains weren’t uniform across every scenario tested, but the headline figure is significant because it is one of the first independently verified comparisons showing Google’s chips beating NVIDIA’s newest hardware on a cost-efficiency basis, rather than Google simply claiming an advantage internally. Where the Advantage Comes From A big part of the story here isn’t just the chip itself, but the software running on it. Google has been steadily opening up pieces of its once tightly-guarded TPU software stack, similar to how it recently open-sourced a library called TPU Raiden https://officechai.com/ai/google-has-open-sourced-its-tpu-raiden-inference-library-in-apparent-bid-to-externalize-tpu-stack/ for moving data between chips during inference. The latest piece of that puzzle is something called TorchTPU, which lets developers run TPUs using PyTorch, the same popular AI programming framework most engineers already use for NVIDIA GPUs. Historically, getting the best performance out of TPUs required using Google’s own, less widely adopted tools, which made switching over from NVIDIA a hassle. TorchTPU removes a lot of that friction. Alongside this, Google also released a set of optimized building blocks called Pallas inference kernels, which are essentially pre-written, highly efficient instructions that tell the chip exactly how to carry out common AI tasks as fast as possible. The other piece of the puzzle is a specialized component inside the Ironwood chip itself called the SparseCore. Modern AI models, especially the very large ones, often use a design called “mixture of experts,” where only certain smaller sub-sections, or “experts,” of the model are activated for any given task rather than running the entire model every time. The tricky part is that the data belonging to each expert needs to be gathered and organized before the main computing engine can process it, and this housekeeping work can slow things down if handled poorly. The SparseCore is dedicated hardware built specifically to handle that organizing job, freeing up the chip’s main processing unit, the TensorCore, to focus purely on the heavy math. SemiAnalysis found that offloading this task to the SparseCore alone improved throughput by 12%, and combined with Ironwood’s lower overall total cost of ownership, it adds up to the reported 50% edge. Why This Matters None of this means NVIDIA is suddenly in trouble. Blackwell Ultra remains an extremely capable chip, and NVIDIA’s software ecosystem is still far more mature and widely used across the industry, a point NVIDIA itself has been keen to stress amid the recent buzz around Google’s TPUs https://officechai.com/ai/delighted-by-googles-success-nvidia-a-generation-ahead-of-industry-nvidia-on-google-tpu-buzz/ . But the direction of travel is what makes this notable. Google is no longer treating TPUs as a strictly internal tool. It’s actively working to make them easier for outside companies to adopt, and firms like Anthropic have already committed to using over a million of them, while OpenAI has also begun tapping into Google’s TPUs https://officechai.com/ai/openai-has-begun-using-google-tpus-in-move-which-will-reduce-reliance-on-nvidia-report/ for parts of its operations, despite being NVIDIA’s biggest customer. Google, by some measures, already owns more AI computing capacity than any other company https://officechai.com/ai/google-leads-in-owning-the-most-ai-chips-followed-by-microsoft-and-amazon-epoch-ai-data/ once its TPU fleet is counted alongside its NVIDIA GPUs. If TPUs continue closing the performance gap while undercutting NVIDIA on cost, and the software tools needed to use them keep getting easier to work with, more AI companies may start treating TPUs as a genuine everyday alternative rather than a curiosity, and that would mark a real shift in a market NVIDIA has dominated almost single-handedly for years.