You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
First thanks a lot for your contribution and useful simulator.
I have a question about the estimated ofmap bandwidth, it seems quite high for me.
I used the tool to simulate a simple fully connected layer which has 1 1 1 1 256 72 1 hyperparameters. In an output stationary model, I set the output buffer to 4 KB. For a double buffer, 2KB is the active buffer size. And a 2KB buffer is enough for my 72 B output right? So we don't need very high bandwidth to shift the things output.
However, when I run the code, I get an estimated 72 bytes per cycle bandwidth, and I check the code, it seems like in the estimated model, the bandwidth for ofmap is directly assigned to the column length of the PE array, which is quite high.
Did I misunderstand something? Since when I run in estimation model, I get quite a high estimated bandwidth (72) for zero stalls, but when I set the bw to 10, there seems like still no stall for the fc layer which I run.
Best wishes!
Thanks in advanced!
The text was updated successfully, but these errors were encountered:
Hi @Oliverwang11, I could not understand your hyperparameters. Can you please provide the exact config and topology files? I can try to reproduce the ofmap bandwidths that you are getting.
@Oliverwang11
We only estimate ifmap and filter bandwidths through read_buffer_estimate_bw.py file. We currently don't have write_buffer_estimate_bw.py, therefore the numbers that you are getting is invalid. We are working on adding write_buffer_estimate_bw.py file.
Hello,
First thanks a lot for your contribution and useful simulator.
I have a question about the estimated ofmap bandwidth, it seems quite high for me.
I used the tool to simulate a simple fully connected layer which has 1 1 1 1 256 72 1 hyperparameters. In an output stationary model, I set the output buffer to 4 KB. For a double buffer, 2KB is the active buffer size. And a 2KB buffer is enough for my 72 B output right? So we don't need very high bandwidth to shift the things output.
However, when I run the code, I get an estimated 72 bytes per cycle bandwidth, and I check the code, it seems like in the estimated model, the bandwidth for ofmap is directly assigned to the column length of the PE array, which is quite high.
Did I misunderstand something? Since when I run in estimation model, I get quite a high estimated bandwidth (72) for zero stalls, but when I set the bw to 10, there seems like still no stall for the fc layer which I run.
Best wishes!
Thanks in advanced!
The text was updated successfully, but these errors were encountered: