Intel shows the GPU D-series SI accelerator with 480 GB of LPDDR5X memory, to run models of up to 300 billion parameters on a single card
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
The choice of cheap, high-capacity memory is aimed at making large models fit on a single card.
The accelerator, previously known as Crescent Island, was shown at the Open Compute Project summit. According to Intel, one card can handle models of up to 300 billion parameters, and eight can handle models of more than a trillion. LPDDR5X memory is used instead of HBM; the stated specs are 1.3 petaflops in FP8, 1.5 TB/s of memory bandwidth and power consumption of up to 350 W. The card comes in a PCIe form factor with air cooling. Price and sales dates have not been announced, and response speed depends not only on memory capacity but also on compute units and software.
Sources 1