Cerebras has an entire AI training system built around its wafer-scale engine (WSE-3) processor, which has 900,000 AI processing cores and 44GB of on-chip memory. That can compete with a cluster of machines, and the company says it's a good choice for large language models and other AI work. The system also includes AI model services, cloud services for quick model training, and a software stack that integrates with frameworks like PyTorch.
Another contender is Together, a cloud platform for fast and efficient development and deployment of generative AI models. It includes new optimizations like Cocktail SGD, FlashAttention 2 and Sub-quadratic model architectures to speed up training and inference. Together supports a broad range of models and offers scalable inference, collaborative tools for fine-tuning, and optimized pricing with big discounts compared to other providers.
If you prefer a more integrated solution, check out NVIDIA AI Platform. It's a collection of accelerated infrastructure, enterprise software and AI models that accelerates the entire AI workflow. It includes multi-node training at scale, accelerated data science pipelines and streamlined deployment of production AI applications, making it a good option for businesses that want to build AI into their operations.