Assume a hypothetical GPU with the following characteristics . Clock rate 1 GHz . 32 simultaneous instruction streams . Contains 8 SIMD cores, each containing 32 single-precision floating-point units, each instruction performs one single -precision floating -point operation (either a multiply or an add ) . 1024 concurrent (but interleaved) instruction streams What is the peak single -precision floating -point throughput for this GPU in GFLOP /sec , assuming that all memory latencies can be hidden ?

Solution:

The clock rate of a hypothetical GPU is 1.5 GHz, exists 16 SIMD processors, each processor contains 16 single-precision floating point units and off-chip memory bandwidth is 100 GB/sec. For this GPUs the peak single-precision floating-point throughput is, core frequency ( GHz ) x number of cores x number of operations per clock = 1.5x16x16 = 384 Thus, the peak single-precision floating-point throughput for this GPUs is 384 GFLOP/sec.

Assume that 4-byte two operands and outputs one 4-byte result requires to each single-precision operation. The memory bandwidth of sustaining this throughput will be, (2x4+1x4) bytes/Flopx384GFLOP/s =12 bytes/Flop x 384GFLOP/s = 4.6 TB/s Thus, this throughput is not sustainable.