How AWS’s Custom AI Chips Are Fueling a Gene Editing Revolution at 56% Lower Cost

The AI-Powered Hunt for Genetic Cures

In the race to develop groundbreaking gene therapies, startup Metagenomi has discovered that smarter AI infrastructure can dramatically accelerate discovery while slashing costs. The company recently revealed that switching to AWS’s custom Inferentia 2 accelerators reduced their AI expenses by 56% compared to previous Nvidia GPU solutions, while maintaining the computational power needed to search for life-saving treatments.

“Gene editing represents a fundamental shift from treating symptoms to addressing disease at its genetic root,” explained Chris Brown, Metagenomi’s Vice President of Discovery. “This approach requires finding precise biological tools that can target and edit specific DNA sequences – essentially searching for molecular needles in genomic haystacks.”, according to market developments

Protein Language Models: The AI Behind the Discovery

Metagenomi’s approach leverages CRISPR technology, the Nobel Prize-winning gene editing system developed by Jennifer Doudna and Emmanuelle Charpentier. To identify suitable enzymes for their therapies, the company employs sophisticated protein language models (PLMs) – a specialized class of generative AI that designs novel protein sequences rather than generating text.

The startup specifically uses Progen2, a transformer model with approximately 800 million parameters that functions similarly to early language models like GPT-2. Unlike massive contemporary models requiring extensive resources, Progen2’s relatively compact architecture makes it ideal for cost-effective deployment on optimized inference hardware., as our earlier report, according to related news

Inferentia 2 vs. Traditional GPUs: The Cost-Performance Equation

When comparing AWS’s Inferentia 2 against Nvidia’s L40S GPUs, the performance specifications tell only part of the story. While the L40S boasts higher theoretical performance (362 teraFLOPS versus 190 teraFLOPS) and more memory bandwidth, real-world economics favored Amazon’s custom silicon., according to additional coverage

“The key advantage came from AWS’s ability to integrate spot instances with their batch processing pipeline,” noted Kamran Khan from AWS’s Annapurna Labs. “Spot instances typically cost about 70% less than on-demand pricing, and when combined with AWS Batch for scheduling non-interactive workloads, the savings become substantial.”, according to market analysis

The infrastructure efficiency extended beyond simple pricing. Inferentia 2 instances demonstrated significantly lower interruption rates – approximately 5% compared to 20% for Nvidia-based spot instances. This reliability translated into more consistent research cycles and fewer wasted computational resources.

Transforming Research Economics

The financial savings directly impacted Metagenomi’s research capabilities. “What would have been a single annual project became something my team could run multiple times per day or week,” Brown emphasized. The increased throughput means researchers can explore more potential enzyme candidates, dramatically improving the odds of discovering viable treatments.

The collaboration demonstrates that for batch-oriented AI workloads like protein generation, raw performance metrics don’t tell the whole value story. Older or specialized accelerators, when combined with optimized cloud infrastructure, can deliver superior economics for specific scientific applications.

Broader Implications for AI in Biotech

Metagenomi’s experience highlights several important trends for AI deployment in biotechnology:

• Specialized hardware can outperform general-purpose solutions for specific computational tasks
• Infrastructure reliability directly impacts research velocity and discovery potential
• Cloud economics enable smaller companies to compete with well-funded pharmaceutical giants
• Batch processing of non-interactive workloads represents a massive cost optimization opportunity

As AI continues transforming drug discovery and development, infrastructure choices are becoming increasingly strategic. The ability to run more experiments at lower costs doesn’t just save money – it potentially saves lives by accelerating the development of treatments for genetic diseases.

References & Further Reading

This article draws from multiple authoritative sources. For more information, please consult:

• https://arxiv.org/abs/2206.13517

This article aggregates information from publicly available sources. All trademarks and copyrights belong to their respective owners.

Note: Featured image is for illustrative purposes only and does not represent any specific product, service, or entity mentioned in this article.