Wafer-Scale Integration
Small-chip accelerators are limited by slow off-chip DRAM to 2 megapixels. With 40GB of on-chip SRAM, Cerebras handles 25 megapixels with ease.
Weight Streaming
The key to large image capability, weight streaming allows us to stream layers into the CS-2 system without the penalty of off-chip memory.
Easy Model Size Changes
To easily leverage increased data sizes, make simple changes to both simple model depth and width in PyTorch or configuration files.
50 Megapixel Segmentation
Distributed memory with massive bandwidth enables segmentation training on multi-channel inputs up to 50 megapixels with deep and shallow networks.
Ultra-Simple Programming
Effortlessly train deep, wide classification and segmentation CV models on Cerebras systems without parallel programming using familiar frameworks.
Featured Resources
More Pixels, More Context, More Insight!
The Cerebras architecture overcomes the limitations of GPUs to train large computer vision models on high-resolution, 50 megapixel images.
Unlocking High-Resolution Computer Vision with WSI
A platform for accelerating CV workloads that allows users to easily scale CNN-based models to large images and volumes.
