[Milkymist-devel] Memory performance

[Sébastien Bourdeauducq]

Hi,

I did some benchmarking and measurements of the memory controller's real 
performance (the system now has an additional core which measures that). Here 
are the results of the idle system (demo firmware started, no preset 
rendering, video output enabled 640x640 at 60Hz) and when rendering some presets.

The performance of the memory system is reasonable (it's not excellent, but it 
doesn't suck either and probably exceeds that of all other free designs). It 
might still be improved: it seems that it could go up to approximately only 
half of the peak bandwidth of the chips (with this access pattern). This is 
encouraging to experiment with an out-of-order memory controller in the far 
future (Milkymist 2.x).

Does anyone have typical results of proprietary SDRAM controllers to compare?

Sébastien

Idle:
Net memory bandwidth     : 293 Mbps
Memory bus occupancy     : 7%
Extrapolated max bandwith: 4185 Mbps
Avg. mem. access time    : 5.52 cycles

Geiss - Bright Fiber Matrix 1:
Net memory bandwidth     : 1067 Mbps
Memory bus occupancy     : 31%
Extrapolated max bandwith: 3441 Mbps
Avg. mem. access time    : 6.58 cycles

Geiss - Swirlie 3:
Net memory bandwidth     : 1160 Mbps
Memory bus occupancy     : 34%
Extrapolated max bandwith: 3411 Mbps
Avg. mem. access time    : 6.71 cycles

StudioMusic - Twisted Galaxy:
Net memory bandwidth     : 942 Mbps
Memory bus occupancy     : 25%
Extrapolated max bandwith: 3768 Mbps
Avg. mem. access time    : 5.94 cycles

Geiss - Spacedust:
Net memory bandwidth     : 1024 Mbps
Memory bus occupancy     : 30%
Extrapolated max bandwith: 3413 Mbps
Avg. mem. access time    : 6.67 cycles

Geiss - Anomaly 2:
Net memory bandwidth     : 1127 Mbps
Memory bus occupancy     : 33%
Extrapolated max bandwith: 3415 Mbps
Avg. mem. access time    : 6.58 cycles

Aderrasi - Candy Avian:
Net memory bandwidth     : 923 Mbps
Memory bus occupancy     : 27%
Extrapolated max bandwith: 3418 Mbps
Avg. mem. access time    : 6.52 cycles