Posted 3/26/2024, 12:00:20 PM
Phison Unveils New Platform to Enable AI Model Training on Workstations, Cutting Hardware Costs for SMBs
- Phison demos new software that uses SSDs and DRAM to expand GPU memory for large AI model training on a single workstation
- Platform allows training massive 70 billion parameter model on 4 GPUs instead of typical 24 GPU cluster, lowering barriers to AI training
- Special SSDs and software store/manage data, freeing GPU VRAM for compute when needed
- Significantly cuts hardware costs but increases training time 4x vs optimized cluster
- Aims platform at SMBs who want private model training without need for fastest speed