The latest accelerators and networking platforms boost performance in advanced simulation, AI, quantum computing, data analytics, and more.
From quantum computing to drug discovery and fusion energy, with advances in accelerated computing and AI, major breakthroughs are emerging around the world, and physics-based simulation-based scientific computing promises to advance dramatically in all areas for the benefit of humanity.
At GTC in March, NVIDIA unveiled the NVIDIA Blackwell platform, which enables generative AI on trillion-parameter large language models (LLMs) at as little as 1/25th the cost and power consumption of the NVIDIA Hopper architecture.
Blackwell's technical capabilities are significant for AI workloads, and will help advance the exploration of a wide range of scientific computing applications, including traditional numerical simulations.
Accelerated computing and AI drive sustainable computing by reducing energy costs. Many scientific computing applications have benefited from this. Compared to traditional CPU-based systems and other systems, the cost and energy consumption of weather simulation are reduced to 1/200 and 1/300, respectively, and the cost and energy consumption of digital twin simulation are reduced to 1/65 and 1/58, respectively.
Make it happen with Blackwell
Scientific computing simulation performance is doubled
Scientific computing and physics-based simulations often rely on so-called double-precision formats or FP64 (floating point) to solve problems. Blackwell GPUs deliver up to 30% better FP64 and FP32 FMA (fusion multiplication) performance than Hopper.
Physics-based simulation is essential for product design and development. Whether it's airplanes, trains, bridges, semiconductor chips, and pharmaceuticals, testing and refining products in simulations can save researchers and developers billions of dollars.
Today's application-specific integrated circuits (ASICs) are designed almost entirely on CPUs, and the entire process is lengthy and complex, including simulation analysis to determine voltage and current.
But that's changing. For example, the Cadence SpectreX simulator is a typical analog circuit design solver. SpectreX circuit simulations are expected to run up to 13 times faster than traditional CPUs on the Grace Blackwell superchip, which consists of a Blackwell GPU and a Grace CPU.
In addition, GPU-accelerated computational fluid dynamics (CFD) has become an important tool. It is used by engineers and equipment designers to predict the behavior of various designs. Cadence Fidelity is expected to run CFD simulations on NVIDIA's Grace Blackwell systems up to 22 times faster than traditional CPU-based systems, capturing unprecedented flow detail.
In another application, Cadence Reality's digital twin software was used to create a virtual replica of a physical data center, including all its components, including its servers, cooling system, and power supply. This virtual model saves time and money by allowing engineers to test configurations and scenarios in advance of real-world applications.
Cadence Reality excels with physics-based algorithms that simulate the impact of heat, airflow, and electricity usage on data centers. This helps engineers and data center operators manage capacity more effectively, anticipate potential operational issues, and make informed decisions that improve efficiency and capacity utilization by optimizing the layout and operation of the data center. With Blackwell GPUs, these simulations are expected to run up to 30 times faster than CPUs, resulting in faster completion times and improved energy efficiency.
AI in Scientific Computing
The new Blackwell accelerator and networking platform will dramatically improve simulation performance.
NVIDIA Grace Blackwell ushers in a new era of high-performance computing (HPC). Its architecture uses a second-generation Transformer engine that is optimized to accelerate LLM inference workloads.
The Blackwell architecture enables a 30x acceleration of resource-intensive applications such as the 1.8 trillion parameter GPT-MoE (Generative Pretrained Converter-Expert Hybrid) model compared to the previous generation of Hopper architecture GPUs, opening up new possibilities for HPC. By enabling LLMs to process and interpret massive amounts of scientific data, high-performance computing applications can accelerate scientific discovery by gaining valuable insights faster.
Sandia National Laboratories is building an LLM smart assistant for parallel programming. Traditional AI can efficiently generate basic serial code, but when it comes to parallel code for HPC applications, LLMs are inadequate. Sandia researchers are tackling this problem with an ambitious project that will automatically generate parallel code with Kokkos. Kokkos is a programming language designed by multinational laboratories to run tasks on tens of thousands of processors of the world's most powerful supercomputers.
Sandia National Laboratories is using an AI technique called Retrieval Enhanced Generation (RAG) to combine information retrieval capabilities with language generation models. The project team is creating a Kokkos database and integrating it with AI models using RAG.
The initial results are very encouraging. The different RAG methods employed by Sandia National Laboratories have autonomously generated Kokkos code for parallel computing applications. They hope to open up new HPC possibilities for the world's leading supercomputing facilities by overcoming hurdles in AI-parallel code generation.
Other examples include renewable energy research, climate science, and new drug development.
Advancing quantum computing
Quantum computing has brought significant acceleration to fields such as fusion energy, climate research, and new drug discovery. As a result, researchers are working to simulate future quantum computers on NVIDIA GPU-based systems and software to develop and test quantum algorithms faster than ever before.
The NVIDIA CUDA-Q platform enables quantum computer simulation and hybrid application development through a unified programming model that enables CPUs, GPUs, and QPUs (quantum processors) to work together.
CUDA-Q is accelerating simulations for BASF's chemistry workflows, Stony Brook University's high-energy and nuclear physics research, and NERSC's quantum chemistry.
The NVIDIA Blackwell architecture will take quantum simulation to new heights. With the latest NVIDIA NVLink multi-node interconnect technology, data can be streamed faster to increase the speed of quantum simulation.
Accelerate data analytics to drive scientific breakthroughs
The way data is processed using RAPIDS is very common in the world of scientific computing. Blackwell includes a hardware decompression engine that decompresses compressed data and speeds up analysis in RAPIDS.
该解压缩引擎可将性能提升至 800GB/s，使 NVIDIA Grace Blackwell 在查询基准测试中的性能较 CPU（在 Sapphire Rapids 上）快 18 倍，较 NVIDIA Hopper Tensor Core GPU 快 6 倍。
With a high memory bandwidth of 8TB/s and Grace CPU's high-speed NVLink interconnect technology, the engine dramatically increases data transfer speeds, speeding up the entire database query process. Blackwell delivers superior performance in data analytics and data science use cases to accelerate data insights and reduce costs.
NVIDIA networking platforms
Extreme performance for scientific computing
The NVIDIA Quantum-X800 InfiniBand networking platform delivers the highest throughput for scientific computing infrastructure.
The platform includes NVIDIA Quantum Q3400 and Q3200 switches, as well as NVIDIA ConnectX-8 SuperNIC, which combine to deliver twice the bandwidth of the previous generation. The Q3400 platform delivers 5x more bandwidth capacity and 14.4 Tflops of network compute power with NVIDIA's SHARPv4 (Scalable Tiered Aggregation and Reduction Protocol) technology, a 9x increase over the previous generation.
Leaps in performance and improved energy efficiency have resulted in significant reductions in workload completion time and energy consumption for scientific computing.
To learn more about NVIDIA Blackwell, click: