Google’s new Trillium TPU v5e outperformed Nvidia’s H100 in image generation tests during MLPerf v4.0 benchmarks, claiming 40% better energy efficiency amid intensifying AI chip competition.
Google’s Trillium TPU made its MLPerf debut in June benchmarks, outperforming Nvidia’s H100 in image generation while emphasizing energy efficiency.
Google’s newly launched Trillium TPU v5e processor has entered the competitive AI accelerator arena, making its first appearance in the industry-standard MLPerf Training v4.0 benchmarks released on June 5. The custom-designed chip specifically targeted the image generation benchmark using Stable Diffusion v2, where it demonstrated significant performance gains over both Google’s previous generation and rival offerings.
Performance Breakthrough
According to results published by MLCommons, the nonprofit consortium behind MLPerf, Google’s system completed the Stable Diffusion image generation task in just 1.67 minutes using 256 Trillium chips. This represented a 2.7x performance improvement per chip compared to Google’s previous-generation TPU v4. Notably, Trillium outperformed Nvidia’s flagship H100 processor, which required 2.5 minutes for the same task using equivalent hardware resources.
Efficiency Focus
Beyond raw performance, Google emphasized Trillium’s energy efficiency advantages, claiming up to 40% lower total cost of ownership than competing solutions. This focus comes as data center operators grapple with the soaring electricity demands of generative AI workloads. “Our fifth-generation TPUs deliver exceptional performance while dramatically improving efficiency,” stated Google Cloud executives in a June 5 technical blog post detailing the results.
The MLPerf v4.0 round featured 34 participants submitting over 600 performance results, including new benchmarks for large language model fine-tuning. While Nvidia chips powered approximately 90% of submissions according to industry analysts, Google’s competitive entry alongside Intel and cloud-specific chips signals fragmentation in the rapidly expanding AI hardware market.
The push for energy efficiency in AI hardware reflects growing operational challenges. Data centers dedicated to AI workloads now consume power equivalent to small countries, with projections suggesting they could account for 3.5% of global electricity by 2028 according to Goldman Sachs research. Google’s efficiency claims target this pain point directly, positioning Trillium as a sustainable alternative in an AI chip market projected to reach $76 billion by 2028 according to Gartner.
Historically, Nvidia has dominated the MLPerf benchmarks, with its GPUs powering the majority of submissions since the benchmark series launched in 2018. In the 2023 training benchmarks, Nvidia chips were used in over 90% of competitive entries according to MLCommons data. This dominance stemmed from Nvidia’s early investment in CUDA software ecosystem that remains widely adopted by AI researchers.
The current focus on energy efficiency mirrors earlier semiconductor industry pivots. During the 2010s mobile computing revolution, chipmakers shifted priorities from pure performance to power efficiency to extend battery life. Similarly, Google’s first-generation TPU in 2016 emphasized operations-per-watt metrics for data center applications, establishing efficiency as a continuous design philosophy that culminates in Trillium’s latest claims.
