Llama 3.1 405B achieves 1.5x throughput improvement with NVIDIA H200 GPU and NVLink.

Rapid advances in large language models (LLMs) continue to drive innovation in artificial intelligence, with NVIDIA at the forefront. According to the NVIDIA Technology Blog, recent developments show a 1.5x increase in throughput for the Llama 3.1 405B model with NVIDIA’s H200 Tensor Core GPUs and NVLink switches.

Advances in parallelism technology

The improvements are primarily due to optimized parallel processing techniques, including tensor and pipeline parallel processing. These methods allow multiple GPUs to operate simultaneously, sharing computational tasks efficiently. Tensor parallelism focuses on reducing latency by distributing model layers across GPUs, while pipeline parallelism minimizes overhead and leverages the high bandwidth of NVLink switches to improve throughput.

In effect, these upgrades deliver a 1.5x improvement in throughput for throughput-sensitive scenarios on NVIDIA HGX H200 systems. The system leverages NVLink and NVSwitch to facilitate powerful inter-GPU interconnection and ensure maximum performance during inference workloads.

Comparative Performance Insights

Performance comparisons show that tensor parallelism excels at reducing latency, while pipeline parallelism significantly improves throughput. For example, in the minimum latency scenario, tensor parallelism outperforms pipeline parallelism by 5.6x. Conversely, in the maximum throughput scenario, pipelined parallelism increases efficiency by a factor of 1.5, highlighting its ability to effectively handle high-bandwidth communications.

These results are supported by recent benchmarks, including a 1.2x speedup on the MLPerf Inference v4.1 Llama 2 70B benchmark achieved through software improvements to TensorRT-LLM using NVSwitch. These advances highlight the potential to optimize AI inference performance by combining parallelism techniques.

NVLink’s role in maximizing performance

NVLink switches play an important role in this performance increase. Each NVIDIA Hopper architecture GPU is equipped with NVLink, which provides significant bandwidth, facilitating high-speed data transfer between stages during parallel execution of the pipeline. This feature minimizes communication overhead, allowing you to effectively scale throughput with additional GPUs.

Strategic use of NVLink and NVSwitch allows developers to tailor parallel processing configurations to their specific deployment requirements and balance compute and capacity to achieve desired performance results. This flexibility is essential for LLM service operators seeking to maximize throughput within fixed latency constraints.

Future outlook and continuous optimization

Looking ahead, NVIDIA’s platform continues to evolve with a comprehensive technology stack designed to optimize AI inference. The integration of NVIDIA Hopper architecture GPUs, NVLink, and TensorRT-LLM software provides developers with excellent tools to improve LLM performance and reduce total cost of ownership.

As NVIDIA continues to improve these technologies, the potential for AI innovation expands, promising breakthroughs in generative AI capabilities. In future updates, we will further investigate latency thresholds and GPU configuration optimizations, and leverage NVSwitch to improve online scenario performance.

Image source: Shutterstock