In AI, computational efficiency is a crucial business advantage. FLOPS/$ measures cost-effectiveness by showing computational power per dollar invested, enabling clear comparisons between GPU providers.

As AI models grow larger, optimizing FLOPS/$ helps organizations stay competitive while managing costs. This metric identifies which platforms deliver the best performance for your investment, affecting training speeds, inference capabilities, and your bottom line. Higher flops/$ is more cost-effective.

GPU Price-Performance Trends

A comprehensive analysis of GPU price-performance by Epoch AI, “Trends in GPU Price-Performance”, examines how the floating-point operations per second per dollar (FLOP/s per $) has evolved over time. Using a dataset of 470 GPU models released between 2006 and 2021, the researchers found that FLOP/s per $ doubles approximately every 2.5 years.

Key Findings on FLOP/s per $

  • Overall market trend: FLOP/s per $ doubles every 2.46 years across all GPUs analyzed

  • ML-specific GPUs: GPUs commonly used in machine learning research show faster improvement, with FLOP/s per $ doubling every 2.07 years

  • Top-performing GPUs: The highest FLOP/s per $ models at any point in time show slower improvement, with price-performance doubling every 2.95 years

The findings contradict some previous estimates and popular claims:

  • The 2.5-year doubling time is slower than Moore's Law (2-year doubling)

  • It's significantly slower than "Huang's Law" (NVIDIA CEO's claim of ~1.1-year doubling)

  • It's faster than previous analyses like Bergal (2019), which estimated a 4.4-year doubling time

Precision Formats and Price-Performance

For GPUs with both FP32 and FP16 performance data, the research found a FP16 price-performance doubling time of around 2.32 years, not significantly different from FP32 trends. This contradicts earlier findings suggesting much faster improvement rates in lower-precision formats.

When examining actual price-performance improvements from using lower precision on the same hardware, the researchers found that tensor-FP16 operations deliver approximately 8x higher computational performance compared to standard FP32 operations, representing a substantial one-time gain rather than a different improvement trajectory.

The table compares BF16 Tensor-Core TFLOPS/$ (Teraflops per dollar) across different providers for the H100 GPU. Given H100 SXM’s 1979 TFLOPS for B16 Tensor-Core, here’s what the TFLOPS/$ comparison shows.

TFLOPS/$ Comparison Chart (FP16 or BF16 Tensor Core H100 SXM)

All providers use the same H100 GPU with identical raw performance (1,979 BF16 TFLOPS), but Hyperbolic Labs delivers nearly 4× better cost efficiency than AWS.

Conclusion

As compute needs escalate with growing model sizes, companies must maximize their AI budgets by optimizing FLOPS/$. Hyperbolic Labs delivers industry-leading computational efficiency, providing significantly more compute power per dollar compared to competitors, ensuring organizations receive the greatest value for their investment while maintaining the processing capacity needed for diverse AI use cases.

Source: Hobbhahn, M., & Besiroglu, T. (2022). Trends in GPU Price-Performance. Epoch AI. https://epoch.ai/blog/trends-in-gpu-price-performance

About Hyperbolic

Hyperbolic is the on-demand AI cloud made for developers. We provide fast, affordable access to compute, inference, and AI services. Over 195,000 developers use Hyperbolic to train, fine-tune, and deploy models at scale.

Our platform has quickly become a favorite among AI researchers, including those like Andrej Karpathy. We collaborate with teams at Hugging Face, Vercel, Quora, Chatbot Arena, LMSYS, OpenRouter, Black Forest Labs, Stanford, Berkeley, and beyond.

Founded by AI researchers from UC Berkeley and the University of Washington, Hyperbolic is built for the next wave of AI innovation—open, accessible, and developer-first.

Website | X | Discord | LinkedIn | YouTube | GitHub | Documentation