Key results
- 5x More experiments per week
- 3x Faster training jobs
Challenge
Epsilon Health is the world’s first AI-native radiology practice reinventing how medical imaging is interpreted. They pair world-class radiologists with purpose-built AI to deliver faster, more consistent, high-quality care. In less than a year, Epsilon Health has served over 300,000 patients, and is on track to interpret 1% of all daily U.S. X-rays. Their AI models accelerate the radiologists' natural workflow without changing how they work. Meanwhile, their in-house AI research team is dedicated to pushing the frontier of ML model capabilities.
That work depends on training models on medical imaging, a kind of data that's unusually demanding to work with. Epsilon Health needed a specialized AI infrastructure partner to tackle the following:
Native-resolution imaging, not downscaled data. X-rays are natively high-resolution, and CT and MRI volumetric data add another order of magnitude on top of that. Training without downscaling requires large context windows, and large context windows require enough memory per GPU to support them, otherwise the models lose the fine-grained detail needed to catch faint findings.
Larger batch sizes for contrastive learning. Epsilon's models rely on contrastive learning approaches, which benefit directly from larger batch sizes. Larger batch sizes need more memory to support them, so memory capacity wasn't a nice-to-have, it was a direct lever on model performance.
Dedicated capacity. Epsilon Health needed a dedicated NVIDIA HGX™ B300 GPU cluster to scale its compute for the next generation of its models. There was no availability, it was hard to find long-term contracts at reasonable prices, and shared platforms had no capacity either.
A small team, without room to build ML infrastructure from scratch. As a small team, Epsilon Health couldn't dedicate significant human resources to building and maintaining its own ML infrastructure. Whatever partner they chose had to take that work off their plate, not add to it.
"We wanted a partner who could get us onto NVIDIA Blackwell Ultra fast, with the memory capacity our imaging models require."
Arjun Karpur, Head of Machine Learning, Epsilon Health
Solution
Epsilon Health partnered with Verda to implement a dedicated instant cluster on Verda’s platform, with NVIDIA Blackwell Ultra (B300) GPUs, chosen specifically for what the hardware generation unlocks, not just for raw compute.
Managed infrastructure, hands-off cluster ops. Managed Kubernetes and direct support from Verda's engineering team, including custom scheduling, automated hardware health checks, and help designing a dataset-caching approach, let Epsilon's team focus on model training rather than infrastructure management.
High-HBM capacity for native-resolution training. NVIDIA Blackwell Ultra's memory capacity lets Epsilon Health train on X-ray, CT, and MRI data at native resolution instead of downscaling, supporting the large context windows volumetric imaging requires and improving fine-grained detection of faint findings.
Blackwell-tuned kernels for training speed. Flash Attention 4 and other libraries built specifically for NVIDIA Blackwell hardware gave Epsilon Health meaningful speedups over an equivalent number of NVIDIA Hopper architecture nodes.
Multi-node training at scale. Training runs span multiple nodes at a time, using NVLink and InfiniBand for fast parameter sharding and weight updates across the cluster.
Fast availability, on workable terms. Verda provided capacity quickly, at a reasonable price and term length, at a point when dedicated NVIDIA HGX B300 capacity was difficult to find anywhere on the market.
"What we didn't expect was a team of engineers who could help us design our own data streaming and cluster management, not just hand us a cluster and walk away."
Arjun Karpur, Head of Machine Learning, Epsilon Health
Results
Epsilon Health's ability to keep improving its models depends on how fast it can train and how many ideas it can test in a given week. Verda's infrastructure changed both.
Less infrastructure work, more experiments. Managed cluster operations and Kubernetes scheduling freed up engineering time Epsilon's team previously spent on infrastructure, letting them run roughly 5x more experiments per week.
Faster training, shorter iteration loops. Upgraded hardware and Blackwell-tuned libraries cut training time significantly against Epsilon's previous Hopper-based setup, training jobs roughly 3x faster.
A shorter path from a research idea to a better-read scan. Faster training loops mean more research directions get tested sooner, feeding directly into the loop that makes Epsilon’s models and radiologists more accurate.
"Faster training loops mean more research ideas, more accurate models, and more accurate, more efficient radiologists."
Rustin Rassoli, Founder, Epsilon Health
Epsilon Health is trying to close a gap that's only getting wider: too few radiologists for too many scans, and patients waiting on results that should take hours, not days. Every improvement to how fast its models learn is a direct lever on how quickly and safely that gap closes.
Partner with Verda to build the dedicated GPU infrastructure that lets AI-native teams like Epsilon Health move as fast as their ambitions. Get in touch to get started.
