PDA

View Full Version : Bandwidth


Frank
29-Sep-2006, 22:02
If you build a GPGPU, you can do the same thing CPUs do: add more local buffers to reduce the latency by having more threads in flight (and store more states and the data requested by them), or go for more bandwidth.

Instead of more local stores and/or cache, isn't it better to add as much SRAM to the package as possible? That might offer more gain than increasing the GP bit, when you want fast results.

3dcgi
30-Sep-2006, 01:45
What do you propose using the SRAMs for if not local stores or cache?

Rolf N
04-Oct-2006, 15:16
If the effective instruction latency is 1 cycle, there's no point in trying to reduce it further.