How much faster is the 1Ghz unit - what we know


craigix

Mega GP Mania
Joined
Feb 3, 2003
Messages
11,008
Location
England
Website
twitter.com
Right, many many emails, PMs and forum questions about this.


How much faster is it?


Well, firstly the difference isn't just the clock speed:


What's new:


CPU is 1Ghz (overclockable)


GPU is now 200MHz instead of 110MHz


RAM is now 512MB instead of 256MB


Memory is clocked at 200MHz instead of 166MHz


CPU L1 cache increased from 16KB/16KB to 32KB/32K


How does this effect performance?


The 1Ghz Pandora is around 66% faster than the 600Mhz Pandora.


Of course both overclock and this can vary the results too.


Notaz has made a table showing the differences in speed:



Code:
______________________________-3530-__|__-3730-__|_[1]____[2]

clock_speed__________________600M_1GHz_1GHz_1280M|_28%____67%

PSX_Crash_Bandicoot_3_(fps)____91_133*__154___184|_<_38%__69%

PSX_Tekken3_(fps)_____________~42_~60*__~67___~77|_<_28%__60%

genesis_sonic1_(fps)_________~245_~405_~416__~520|_28%____70%

32X_Virtua_Racing_(fps)________46___80___84___107|_33.7%__82%

zip_compresion_(-9,_seconds)_28.3_17.7_16.1__12.8|_38%____76%

bzip2_compression_(seconds)___141___94___86____71|_32%____64%

"memspeed",_MB/s______________310__326__473___474|_45%____53%


[1] - 3530@1GHz vs 3730@1.28GHz

[2] - 3530@600MHz vs 3730@1GHz										  

* had to run at 960MHz, unstable otherwise because of heavy use of NEON
 
Last edited by a moderator:
I'm very interested in Battery consumption compared to the old Pandoras. Do 1GHz standard clocking has a big impact on this or do all the new powersave features of the new Kerrnel help to have similar battery times like we know from the old Pandoras?
 
Last edited by a moderator:
Ok anyone want to chime in here as to how the 1Ghz pandy could help with amiga 1200/cd32 emulation in theory at least.I know commander beef is working on uae-ex- A1200 emulation should benefit from this right? Commander beef please optimise your miggy emulator for the 1GHZ model.
 
Ok anyone want to chime in here as to how the 1Ghz pandy could help with amiga 1200/cd32 emulation in theory at least.I know commander beef is working on uae-ex- A1200 emulation should benefit from this right? Commander beef please optimise your miggy emulator for the 1GHZ model.

There's no optimization to be done for the 1GHz model.. it'll just run things faster (to varying degrees, I don't have a lot of faith in 1.25GHz running things 50% faster than the 1GHz original Pandora.. maybe if the memory latency is much better)
 
Last edited by a moderator:
Ok anyone want to chime in here as to how the 1Ghz pandy could help with amiga 1200/cd32 emulation in theory at least.I know commander beef is working on uae-ex- A1200 emulation should benefit from this right? Commander beef please optimise your miggy emulator for the 1GHZ model.

There's no optimization to be done for the 1GHz model.. it'll just run things faster (to varying degrees, I don't have a lot of faith in 1.25GHz running things 50% faster than the 1GHz original Pandora.. maybe if the memory latency is much better)

What about the doubled L1 cache? That seems like it'd make a big difference in certain usages cases.


I'd be interested in mupen64 benchmarks with the improved CPU speed and L1 cache, and for GPU comparison, maybe something like Doom3 (once the port is finished..).
 
What about the doubled L1 cache? That seems like it'd make a big difference in certain usages cases.

I'd be interested in mupen64 benchmarks with the improved CPU speed and L1 cache, and for GPU comparison, maybe something like Doom3 (once the port is finished..).

Ari64 actually tested Mupen64plus on a DM3730 (Beagleboard xM) running at the same clock speed and found zero improvement. I was surprised.





Note that it's using the same dynarec as PCSX-reARMed, so you'd expect some of the performance characteristics to be similar, but of course this could vary a lot depending on what you're emulating in both cases. Just having a lot more RAM on N64 could change things substantially. The other big performance consumer for reARMed is the GPU emulation, and of course I would like to know if gpu-neon got a huge boost here. 50% improvement for 28% clock is about 17% improvement in IPC which is very substantial for only an L1 improvement, when you have a much larger and not terribly high latency L2 cache in front of it. 16KB to 32KB at 4-way set associativity isn't past the knee for diminishing returns or anything but I still wouldn't expect such a huge difference for most things.


Memory latency could play a bigger role, since it was very bad on OMAP3530.. but at 160ns (according to this http://www.7-cpu.com/cpu/Cortex-A8.html) it doesn't seem that dramatic. OMAP3530 was something like 250ns but that measurement is end to end, I think you'd measure around 200ns for the same test on DM3730. This would still probably make a bigger impact than the L1.


When testing performance for gpu-neon I would find that the memory clock did hold it back somewhat, making the improvements a little less than linear. I don't know how much this applies to other software. It could be worse because gpu-neon pretty aggressively prefetches, but prefetching on Cortex-A8 can be rather dodgy. Suffice it to say though, DM3730 only improving the memory clock by 20% would mean it'd need an even bigger improvement in other places to see such a big gain at clocks increased by 28%.
 
One improvement I forgot to mention before was that the DSP now clocks up to 800MHz, which is a huge increase over the baseline in Pandora, nearly 2x.. but of course people will have to start actually using the DSP for this to be pertinent.
 
Ok anyone want to chime in here as to how the 1Ghz pandy could help with amiga 1200/cd32 emulation in theory at least.I know commander beef is working on uae-ex- A1200 emulation should benefit from this right? Commander beef please optimise your miggy emulator for the 1GHZ model.

There's no optimization to be done for the 1GHz model.. it'll just run things faster (to varying degrees, I don't have a lot of faith in 1.25GHz running things 50% faster than the 1GHz original Pandora.. maybe if the memory latency is much better)
I see thanks for the response exo.I wonder if the memory latency is much better then on the new units? not that i really know what that means in technical terms.
One improvement I forgot to mention before was that the DSP now clocks up to 800MHz, which is a huge increase over the baseline in Pandora, nearly 2x.. but of course people will have to start actually using the DSP for this to be pertinent.
Ahh and could this be useful for amiga emulation or such.?
 
An example of overclocking extremes:


*Original Pandora overclocked to 1Ghz* Vs. *New DM Pandora overclocked to 1.28Ghz*


The new DM Pandora is 50% faster than the original Pandora in this test.
This appears to be completely wrong, I have done some "unscientific" tests again and it's not even close. Actually it was me who came up with that 50% figure, but this was just and impression after preliminary testing, when I saw Crash Bandicoot running significantly faster on DM3730 at the same clock. The correct figure is closer to 30%. Oh well, at least it's proportional to CPU clock increase, meaning the memory is not slowing it all down too much..



Code:
______________________________-3530-__|__-3730-__|_[1]____[2]

clock_speed__________________600M_1GHz_1GHz_1280M|_28%____67%

PSX_Crash_Bandicoot_3_(fps)____91_133*__154___184|_<_38%__69%

PSX_Tekken3_(fps)_____________~42_~60*__~67___~77|_<_28%__60%

genesis_sonic1_(fps)_________~245_~405_~416__~520|_28%____70%

32X_Virtua_Racing_(fps)________46___80___84___107|_33.7%__82%

zip_compresion_(-9,_seconds)_28.3_17.7_16.1__12.8|_38%____76%

bzip2_compression_(seconds)___141___94___86____71|_32%____64%

"memspeed",_MB/s______________310__326__473___474|_45%____53%


[1] - 3530@1GHz vs 3730@1.28GHz

[2] - 3530@600MHz vs 3730@1GHz										  

* had to run at 960MHz, unstable otherwise because of heavy use of NEON


EDIT: sorry for the underscores, forum software mangles spaces in some braindead way, clean version at http://pastie.org/3998172


The emulators were run here with frame limiter and frameskip disabled.


Same exact binaries, firmware and kernel used on both units.


memspeed is NEON portion of Måns Rullgård's test
 
Last edited by a moderator:
An example of overclocking extremes:


*Original Pandora overclocked to 1Ghz* Vs. *New DM Pandora overclocked to 1.28Ghz*


The new DM Pandora is 50% faster than the original Pandora in this test.
This appears to be completely wrong, I have done some "unscientific" tests again and it's not even close. Actually it was me who came up with that 50% figure, but this was just and impression after preliminary testing, when I saw Crash Bandicoot running significantly faster on DM3730 at the same clock. The correct figure is closer to 30%. Oh well, at least it's proportional to CPU clock increase, meaning the memory is not slowing it all down too much..


-3530- | -3730- | 3530@1GHz vs 3730@1.28GHz


clock speed 600M 1GHz 1GHz 1280M| 28%



PSX Crash Bandicoot 3 (fps) 91 133* 154 184| < 38%



PSX Tekken3 (fps) ~42 ~60* ~67 ~77| < 28%



genesis sonic1 (fps) ~245 ~405 ~416 ~520| 28%



32X Virtua Racing (fps) 46 80 84 107| 33.7%



zip compresion (-9, seconds) 28.3 17.7 16.1 12.8| 38%



bzip2 compression (seconds) 141 94 86 71| 32%



"memspeed", MB/s 310 326 473 474| 45%



EDIT: damn unable ro make it do monospace font


The emulators were run here with frame limiter and frameskip disabled.


memspeed is NEON portion of Måns Rullgård's test
So generally speaking it's a 30 per cent gain in performance.Still a good increase i think.
 
The emulators were run here with frame limiter and frameskip disabled.
memspeed is NEON portion of Måns Rullgård's test

Thanks for the numbers, really informative. Bandwidth tests here tend to be limited by latency at least to some extent, so the improved latency probably accounts for the better than expected increase in bandwidth. Or it could actually be a real increase in attainable bandwidth.
 
Last edited by a moderator:
I will update the original post, do you think you could give us a % increase on the TI clock speeds, ie 600mhz vs. 1Ghz? As not all Pandoras will overclock to 1Ghz.
 
So the 1Ghz is between 53% and 82% faster on the PSX emulator at stock speeds, is this correct?
 
Last edited by a moderator:
Back
Top