TL;DR
This paper analyzes the performance of NVIDIA's Jetson TX2 and AGX Xavier GPUs for automotive applications, identifying opportunities to optimize hardware configurations for better performance or reduced cost. By modeling these GPUs and running benchmarks, the authors demonstrate that specific hardware parameters can be adjusted to achieve significant improvements without sacrificing performance or by reducing hardware costs.
What Problem Does This Paper Address?
The increasing complexity of automotive systems, particularly those related to Advanced Driver Assistance Systems (ADAS) and Autonomous Driving (AD), demands significant computational power. While GPUs have emerged as a solution to meet these demands, it's unclear if current high-performance GPU designs are optimally suited for the automotive domain, which has strict power and cost constraints. This paper addresses the gap in understanding how to best configure automotive GPUs by exploring the impact of various microarchitectural parameters on performance and identifying optimization opportunities.
Methods
The authors employed a cycle-accurate, heterogeneous CPU-GPU simulator, gem5-gpu, to model the NVIDIA Jetson TX2 and AGX Xavier platforms. They meticulously tuned the simulator by incorporating publicly available architectural parameters and, where necessary, reverse-engineered or estimated missing details through synthetic experiments. The study involved a systematic design space exploration, where individual hardware parameters (like cache size, associativity, number of CUDA cores, SMs, etc.) were varied one at a time, followed by experiments where multiple parameters were changed concurrently. This allowed for a comprehensive analysis of their impact on performance.
Datasets
The paper utilizes the Rodinia benchmark suite for its experiments. Rodinia is a collection of benchmarks designed for heterogeneous computing, targeting both CPU and GPU platforms. The authors selected Rodinia because it includes kernels relevant to autonomous driving tasks such as image processing and pattern recognition, which share similarities with ADASMark benchmarks. While ADASMark was considered, it was not accessible at the time of writing.
Key Insights and Ideas
- Not all GPU resources scale linearly: Doubling CUDA cores doesn't necessarily double performance, and reducing certain resources (like register file size) can have a disproportionate negative impact on specific benchmarks.
- Cache parameters have nuanced effects: While increasing L1 cache size generally improves performance, increasing L2 cache size beyond a certain point yields diminishing returns. Cache associativity, however, shows less significant impact on performance in isolation.
- Software granularity matters: The effectiveness of hardware changes is heavily dependent on how the application is parallelized. Optimizing thread block and grid distributions can unlock performance gains from hardware modifications.
- Automotive GPUs have unique constraints: Unlike general-purpose GPUs, automotive GPUs operate under strict power envelopes, making optimization for both performance and efficiency crucial.
- Trade-offs are key: The study highlights that optimizing automotive GPUs involves careful trade-offs between performance, die area, and power consumption.
Key Results
The paper proposes two types of improved setups for both TX2 and AGX Xavier:
- Reduced die space with similar performance: For the TX2, this involved halving the register size, warp schedulers, and shared memory, resulting in a performance degradation of less than 5% on average. For the AGX Xavier, similar reductions were proposed for warp schedulers and register file size.
- Increased performance with comparable die space: This was achieved by increasing the number of SMs and L1 cache sizes while decreasing L2 cache size. For the TX2, one variation showed a 15% reduction in execution time, while another achieved a 9% reduction. For the AGX Xavier, similar improvements were observed.
Limitations and Future Work
The authors acknowledge that they lack access to complete, detailed architectural parameters for these GPUs, necessitating estimations and reverse-engineering. They also note that the proposed hardware optimizations might require corresponding software modifications (e.g., grid and thread block distribution) to fully realize their potential, which was explored as a proof of concept but not exhaustively for all benchmarks. Future work could involve a more in-depth analysis of software-hardware co-optimization and exploring the impact of these changes on power consumption, a critical factor in automotive applications.
Summary updated