Today, Lambda reached another milestone in its mission to build the best infrastructure for AI. Congratulations to the Lambda team, and we’re particularly excited about this investment coinciding with massive undertakings and growth projects already underway.
Picture this: you’re an AI researcher in 2017 trying to build a model that can accomplish something supremely cool. Maybe your ResNet model will classify images with an astounding 80% accuracy. Or your Deep Q-Network will play Atari games at human-level after sufficient episodes. You have the latest and greatest hardware and software that you –painstakingly–purchased and assembled yourself. After making sure Ubuntu is talking to your dual NVIDIA GTX 1080 Ti GPU (each with 8GB memory and 11 tera flops at FP32), downloading and configuring drivers, libraries and more, you’re ready to train your model. You have a ResNet model with 50 layers and 138 GB of ImageNet training data, and it looks like you’ll need about 3 hours to make it through each epoch (and weeks to fully train the model).
That was the reality AI developers faced when Lambda began building workstations and server blades with four GPUs, low-latency memory, and Lambda Stack (the latest OS, ML libraries and GPU drivers). As AI developers, brothers Stephen and Micahel Balaban wanted to build the infrastructure they wished existed for everyone. With Lambda, AI teams could finally get the best tool for ML training shipped to them and ready out-of-the-box.
The company caught our eye and we made a bet on Stephen and Michael. Their vision that AI would permeate computer applications and infrastructure, would prove prescient. (As Stephen liked to say: “A GPU for every person, just like the smart phone or computer.”) They put the AI developer first in every conversation and consideration, and dedicated themselves to the relentless pursuit of building the very best infrastructure for this customer.
Today, Lambda makes it possible for any developer to quickly access training capacity at an affordable rate, with no commitment or contract required. In minutes, users can access and start training a model on the highest capability H100 NVIDIA GPU (80GB memory, 67 tera flops at FP32) for the best price in the market. All the drivers and libraries are pre-installed and consistently deployed. Thanks to Distributed Data Parallel and a confluence of other software improvements, you can spin up multiple GPUs and get near linear scalability.
Back in 2017, training a performant model usually meant buying a $700 GPU and kicking off a job to run overnight. Today, with Lambda, it’s about $20 and can be completed during your lunch break.
In six short years, AI capabilities have come a long way and surprised even experts in the field. This is possible in no small part due to the visionaries that build the best in AI infrastructure. If you’re building in ML, we hope you try Lambda Cloud soon and bet you’ll continue to be impressed. Lambda is just getting started.