There is a 8-core GCN implementation in Kaveri that consumes a substantial 47% of the 2.41 billion transistor die. Each of those 8 GCN cores has 64 stream processors, for a total of 512.
Based on GCN architecture whitepaper, one compute unit includes 4 separate SIMD units. And one SIMD unit simultaneously executes a single operation across 16 work items.
Processing elements inside one GCN cores. |
0 意見:
張貼留言