AI INFRASTRUCTURE
A Perplexity lehetővé teszi billió paraméteres modellek futtatását AWS-en
Perplexity says it has created custom kernels that allow trillion-parameter models, like Kimi-K2, to run on AWS EFA (Elastic Fabric Adapter). This represents the first solution to make these massive, high-performance models viable on standard AWS infrastructure. By optimizing the way data moves through the hardware, Perplexity is lowering the technical barrier for deploying the world's most complex AI models in the cloud.
- Created custom kernels for trillion-parameter model support
- Optimized specifically for AWS Elastic Fabric Adapter (EFA)
- Makes models like Kimi-K2 viable on standard cloud infrastructure
- Solves significant scaling and latency issues for massive models
Miért fontos?
Enabling trillion-parameter models on standard cloud infrastructure like AWS means that cutting-edge AI is no longer restricted to companies with bespoke, private supercomputers.