Sagar Karandikar
|
a5616eb5c3
|
Switch to using XDMA instead of EDMA. One catch is that we always have
to flash all of the FPGAs on an instance, otherwise XDMA hangs. See
comment in run_farm.py. See performance results below:
latency
--------
old (edma):
[[0.021875, 2.05], [0.5009374999999999, 21.78], [0.9996875, 26.3], [2.0015625, 29.14], [2.9990625, 30.26], [4.0009375, 30.91], [5.000624999999999, 31.92], [6.0003125, 32.87], [7.0, 33.11], [7.9996875, 34.01], [8.999375, 34.03], [9.999062499999999, 34.31]]
new (xdma):
[(0.021875, 2.39), (0.5009374999999999, 26.13), (0.9996875, 31.41), (2.0015625, 35.97), (2.9990625, 37.85), (4.0009375, 38.98), (5.000624999999999, 39.71), (6.0003125, 40.17), (7.0, 40.39), (7.9996875, 40.72), (8.999375, 40.75), (9.999062499999999, 40.89)]
scale (truncated)
-------
old (edma):
[(64, 8.84), (32, 27.54), (16, 28.77), (8, 29.74), (4, 33.32), (2, 35.4), (1, 37.02)]
new (xdma):
[(64, 9.3), (32, 34.58), (16, 35.06), (8, 35.93), (4, 40.09), (2, 43.1), (1, 44.49)]
|
2018-11-16 20:19:08 +00:00 |