A Bold Claim from Google at Cloud Next 2025

At the recent Google Cloud Next 2025 event, the tech colossus boldly declared that its latest Ironwood TPU v7p pod outshines the world-renowned El Capitan supercomputer by a staggering 24 times in speed!

Expert Skepticism on Google's Comparison

However, industry expert Timothy Prickett Morgan of TheNextPlatform quickly shot down this assertion. He argued that Google's comparison juxtaposes the sustained performance of El Capitan—harnessing 44,544 AMD 'Antares-A' Instinct MI300A chips—against the theoretical peak of the Ironwood pod, equipped with 9,216 TPU v7p engines. He called this a 'silly comparison' and emphasized that executives at Google should know better.

Understanding Different Purposes of AI and HPC Systems

Morgan further clarified that while it's valid to contrast AI systems with high-performance computing (HPC) machines, the two serve very different functions. El Capitan is tailored for high-precision simulations, making it unsuitable for AI training and inference tasks like those for which the Ironwood pod was designed.

Cost Efficiency: A Crucial Factor

The expert also pointed out an important metric: cost. He noted that high performance should come at the lowest possible cost, and when it comes to HPC gear, nobody secures better deals than the U.S. Department of Energy.

The Numbers: Unpacking Performance and Pricing

According to estimates, the Ironwood pod achieves 21.26 exaflops of FP16 and 42.52 exaflops of FP8 performance, with a build cost of $445 million and $1.1 billion to rent over three years, translating to $21 per teraflop (build) or $52 (rental). In contrast, El Capitan delivers a hefty 43.68 FP16 exaflops and 87.36 FP8 exaflops at a build cost of $600 million, resulting in a more budget-friendly $14 per teraflop.

The Verdict: El Capitan's Edge

Morgan concluded that El Capitan provides 2.05 times more performance than the Ironwood pod at peak performance levels. Despite Google's optimistic portrayal, the Ironwood pod does not truly rival El Capitan's capabilities.

Performance Meets Precision: The HPL-MxP Factor

Morgan added a technical note regarding HPL-MxP, which utilizes mixed precision calculations to yield results similar to all-FP64 math, achieving significant performance boosts.

A Comparative Analysis of AI and HPC Systems

For those curious about how these systems stack up against each other, Morgan's analysis includes a detailed table comparing the top AI and HPC systems on performance, memory, storage, and cost-effectiveness. Although Google's TPU pods remain competitive, Morgan reiterates that El Capitan holds the upper hand in terms of cost-to-performance ratio.

Final Thoughts: A Complex Comparison

Morgan acknowledges that while the comparison may not be flawless, it sheds light on crucial distinctions between these two formidable technologies. With question marks on estimates where ambiguity exists, the debate continues.