Architecture - Compute numbers
1.5 TeraOPS (1.5x1012) 8 bit multiplies and adds
- LEE + PE
- 12K bit data-path (8x1536)
600 Gigabytes/sec on chip DRAM bandwidth
- 50k bit wide data-bus (32x1536)
1.2 Terabytes/sec PE register file data bandwidth
1.2 Terabytes/sec inter PE data bandwidth
ɯ GigaFlops (IEEE compatible)
- Can be faster with custom format
150M 3D Transformations/sec