In early 2020, Argonne National Laboratory started collaborating with scientists from around the world to fight the SARS-CoV-2 virus with accelerated drug discovery. Using artificial intelligence (AI) initiatives, they created machine learning (ML) models of the virus and used them to screen a database including billions of candidate drug molecules. These simulations enabled them to identify high-potential lead compounds to be used in clinical therapy trials.
Argonne has partnered with Groq over the last year to realize hardware acceleration possible by using GroqChip™ and the associated software suite, GroqWare™. Specifically, Groq’s solution delivers 333X better performance versus their existing GPU solution, reducing the time to solution for these algorithms from days to minutes.
“Using the Groq platform at Argonne, we were able to accelerate our efforts to identify promising COVID-19 drug candidates from a vast number of small molecules,” said Argonne computational scientist Tom Brettin. “The system’s AI capabilities enabled us to achieve significantly more inferences a second, reducing the time needed for each search from days to minutes.”
In response Jonathan Ross, CEO of Groq commented, “There are some problems you just can’t solve with legacy architecture. We love working with customers like Argonne who are motivated and willing to apply new tools to solve difficult problems, and proud to have contributed, even if just a little, to the COVID response.”
Argonne leveraged GroqChip accelerator, a high-performance chip with 750Tops @900 MHz and 80 terabytes/sec of memory bandwidth. Instead of the inefficient reloading and context switching required by GPU solutions, Groq’s unique architecture allows far more models to fit within its unified, global, coherent on-chip memory. Another innovation that improves performance is Groq TruePoint™ technology, which maintains superior accuracy while taking advantage of lower precision. Together, these two Groq capabilities delivered a 333x GroqChip advantage for Argonne’s workload. The simple single core Tensor Streaming Processor architecture, which is the foundation for all of Groq’s products, helps reduce data movement, consumes less energy, and improves performance for optimized scalability.
Coupled with the architectural advantages, Groq has also optimized the end-to-end system data pipeline to realize further performance benefits. For these algorithms, data needs to be quantized and reshaped. Solutions based on GPUs and other emerging ML accelerators often overlook this and introduce inefficiencies by offloading these tasks to the slower CPU. In contrast, Groq avoids these inefficiencies by accelerating data manipulation tasks on GroqChip.
“Data processing and compute performance are key to achieving high performance for inference tasks for scientific applications. GroqChip, together with GroqWare, provides an optimized end-to-end system needed to glean faster insights,” said Venkat Vishwanath, lead of the Argonne Leadership Computing Facility’s data science team.
Along with performance, Groq was able to help simplify Argonne’s development flow utilizing GroqWare, which focuses on providing ease of use resulting in high developer velocity. The resulting accelerated platform then enables scientists to iterate quickly with new ideas and new models – accelerating time to innovation. Traditionally, developers have to go through a slow, dynamic profiling process after deployment to optimize the hardware that will be run in production. Groq simplifies the development process in two primary ways:
- Groq’s deterministic architecture, which extends single core performance across multiple chips, enables GroqWare to deliver consistent predictable performance across an ensemble of ML models with zero overhead for context switching.
- Groq’s TruePoint technology maintains accuracy while exploiting the efficiency of lower precision, further enhancing performance of Argonne’s workload.
Groq accelerates development from the scientist to the production engineer, delivering developer velocity. At Argonne, Groq has empowered the globally distributed group of researchers to efficiently collaborate on updating the virus model and near-instantly screening the 4 billion drugs in their database for efficacy against the virus. This has led to the accelerated discovery of strong drug combination candidates to combat the pandemic.