Executive Summary
Our analysis of 470 GPU models released between 2006 and 2021 reveals key insights into price-performance trends:
- Overall Trend: Floating-point operations per second per dollar (FLOP/s per $) doubles every 2.46 years (95% CI: 2.24–2.72 years).
- Top GPUs: High-end models show slower improvement (2.95-year doubling time).
- ML GPUs: Models used in machine learning research improve faster (2.07-year doubling time).
These trends are:
- Slightly slower than Moore's Law (2-year doubling)
- Much slower than Huang's Law (1.08-year doubling)
- Faster than most prior estimates
Introduction
Graphics Processing Units (GPUs) have become the cornerstone of modern machine learning acceleration. This report examines historical price-performance trends to better understand:
- Hardware evolution driving AI progress
- How current trends compare to established laws (Moore's, Huang's)
- Differences between general-purpose and ML-optimized GPUs
Key Findings at a Glance
| Metric | Doubling Time | 10x Improvement Time |
|---|---|---|
| All GPUs | 2.46 years | 8.17 years |
| ML GPUs | 2.07 years | 6.86 years |
| Top GPUs | 2.95 years | 9.81 years |
👉 Explore GPU performance benchmarks
Methodology
Dataset Composition
- Combined data from two major sources (Median Group and Sun et al.)
- 470 GPU models from AMD, Intel, and NVIDIA (2006–2021)
Filtered for:
- Unique models
- Verified specifications
- Post-2006 releases (pre-CUDA era excluded)
Analysis Approach
- Calculated FLOP/s per dollar for each model
- Applied linear regression to identify trends
Compared subgroups:
- General population
- ML research GPUs (26 models)
- Monthly top performers (57 models)
Key Trends
Overall Price-Performance
The foundational trend shows steady improvement at a rate slightly slower than Moore's Law but significantly faster than many previous estimates:
Hypothetical illustration of performance trends
Machine Learning GPUs
GPUs favored by ML researchers demonstrate accelerated improvement, likely due to:
- Targeted architecture optimizations
- Increased R&D focus from manufacturers
- Labs prioritizing cutting-edge hardware
Top-Performing Models
While representing peak capabilities, these models show slower improvement (2.95-year doubling), suggesting:
- Diminishing returns at the high end
- Different optimization priorities for premium products
Comparative Analysis
Against Established Laws
| Benchmark | Doubling Time | Relation to Our Findings |
|---|---|---|
| Moore's Law | 2 years | Faster than observed |
| Huang's Law | 1.08 years | Much faster than observed |
| CPU Historical | 2.32 years | Similar to overall trend |
Precision Comparisons
Analysis of FP16 performance shows:
- No significant difference from FP32 trends
- Contrary to some previous studies suggesting faster FP16 gains
Practical Implications
For ML practitioners and hardware developers:
- Investment planning: Expect ~2.5-year refresh cycles for optimal price-performance
- Architecture choices: FP32 remains reliable benchmark despite FP16 adoption
- Roadmapping: High-end GPUs may not reflect typical improvement rates
FAQ Section
Q: How does this compare to CPU performance trends?
A: GPU improvements are slightly faster than historical CPU trends (2.32-year doubling for CPUs vs. 2.46-year for GPUs).
Q: Why do ML GPUs show faster improvement?
A: Likely due to concentrated R&D efforts and ML-specific optimizations in newer architectures.
Q: Should I wait for next-gen GPUs?
A: Based on these trends, waiting 2–3 years typically yields 2x price-performance improvement.
Q: How reliable are these estimates?
A: With 470 data points and 95% confidence intervals, we consider these among the most robust GPU performance estimates available.
Conclusion
Our analysis suggests GPU price-performance improvements are:
- Consistent (~2.5-year doubling)
- Between Moore's and Huang's Law projections
- Faster for ML-optimized hardware
- More modest for premium models
These findings provide valuable benchmarks for:
- AI research planning
- Hardware investment decisions
- Performance expectation setting
For detailed methodology and complete dataset access, refer to the original publication.
This optimized version:
1. Maintains all key findings while improving readability
2. Incorporates SEO-friendly headings and structure
3. Uses engaging formatting (tables, anchor texts)
4. Adds value through FAQs and practical implications
5. Removes all promotional/duplicate content