Welcome, Guest. Please login or register.

Author Topic: Linux PowerPC Blender benchmark  (Read 9244 times)

Description:

0 Members and 1 Guest are viewing this topic.

Offline KimmoK

  • Sr. Member
  • ****
  • Join Date: Jun 2004
  • Posts: 319
    • Show all replies
Re: Linux PowerPC Blender benchmark
« on: May 04, 2012, 06:18:59 AM »
No big surprice on tests so far.
PA6T performs as expected (notebook caliber chip).

So far it seems that FPU/SIMD performance is not as good as PASemi estimated. IIRC, according to PASemi PA6T should perform like G5 at the same clock rate.

PA6T was technically the best available option for custom high end PPC board, it just is too expensive and the insanely complex motherboard design makes the price even worse when compared to non AOS4 HW.
Definitely the best available option as the SMP development platform for AOS4.

Other than that, summing best parts...
Very good results in:
- fileoperations (2x speed even with old drivers vs others)
- mpeg decoding (20% faster than G4)
- RAM speed (10x speed when compared to G4s)
- slightly faster L2 than on any other

So far we have not seen tests where fast RAM access is truly put in use (lack of such SW?).

I expect to see "better than ever before" results from cards on PCIe, once drivers get beyond handling x1 speed.

update/
Blender might require some rework, as also the 2.3Ghz G5 seems pathetic against the G4... and PA6T almost as it would not be using Altivec. Was the same blender binary in use in all cases?
/update

@Iggy
"You gave them what they asked for."
Hmm???
« Last Edit: May 04, 2012, 07:58:23 AM by KimmoK »
- KimmoK
// Windows will never catch us now.
// The multicolor AmigaFUTURE IS NOW !! :crazy:
 

Offline KimmoK

  • Sr. Member
  • ****
  • Join Date: Jun 2004
  • Posts: 319
    • Show all replies
Re: Linux PowerPC Blender benchmark
« Reply #1 on: May 04, 2012, 07:46:42 AM »
Quote from: WolfToTheMoon;691532
very poor results, indeed. unless there is a software glitch in there somewhere, I'd say it becomes increasingly clear why Apple moved on to x86. It seems PA Semi's party piece was power savings, but they forgot sbout the power


They were initially doing "G5 laptop" chip. But they had roadmap to have up to 16 cores, IIRC.

@Trev

Dualcore 970 gives 1.85x the performance of single core.
PA6T gives 2.0x the performance of single core.

Might show that PA6T has a more modern MP aware design. There starts to be some use for the high bandwidth and large cache when there are more cores to use them.

Some older tests: http://www.barefeats.com/g5sum.html
http://www.barefeats.com/g5.html
That would hint that G5 should perform better than how it did in blender tests (vs G4).
(it should be up to almost 2x better in rendering??)
« Last Edit: May 04, 2012, 08:18:51 AM by KimmoK »
- KimmoK
// Windows will never catch us now.
// The multicolor AmigaFUTURE IS NOW !! :crazy:
 

Offline KimmoK

  • Sr. Member
  • ****
  • Join Date: Jun 2004
  • Posts: 319
    • Show all replies
Re: Linux PowerPC Blender benchmark
« Reply #2 on: May 04, 2012, 10:54:35 AM »
@takemehome...
No fetish.
Just that there is no way forward without going multicore. World knows it, every Amiga flavor OS team knows it.

For most desktop needs 2Ghz single core seems enough, while there are a lot of productivity SW that benefit from more cores as well as other advancements.

btw. PA6T was planned to have 8 cores, not 16 as I said before.

@zylesea
"No, it wasn't. It's performing poor. ..."
MPC8641D system would have been cheaper, yes.
But slower (it's slower than those powerBook G4's) in almost every aspect and with less modern technology for developers to tinker with.
 
IIRC, for G5 macs they managed to get up to 2x speeds in rendering with single 1.6Ghz G5 when compared to 2xG4 systems (Bryce test), so there must be some optimizations ahead before we have seen the maximum out from PA6T.
« Last Edit: May 04, 2012, 11:59:48 AM by KimmoK »
- KimmoK
// Windows will never catch us now.
// The multicolor AmigaFUTURE IS NOW !! :crazy:
 

Offline KimmoK

  • Sr. Member
  • ****
  • Join Date: Jun 2004
  • Posts: 319
    • Show all replies
Re: Linux PowerPC Blender benchmark
« Reply #3 on: May 04, 2012, 12:01:44 PM »
@ppcamiga1

Is there OpenCL benchmark where the effect of system bandwidths are observed?
For example PCIex4 vs PCIex16.
- KimmoK
// Windows will never catch us now.
// The multicolor AmigaFUTURE IS NOW !! :crazy:
 

Offline KimmoK

  • Sr. Member
  • ****
  • Join Date: Jun 2004
  • Posts: 319
    • Show all replies
Re: Linux PowerPC Blender benchmark
« Reply #4 on: May 05, 2012, 09:22:46 AM »
@thread

Interesting observation by pavlor:
http://amigaworld.net/modules/newbb/viewtopic.php?mode=viewtopic&topic_id=35671&forum=34&start=100&viewmode=flat&order=0#664371

"Edit - exact numbers:
PowerBook G4 1667 MHz, Blender 2.49: 325 s
PowerBook G4 1667 MHz, Blender Piru: 205 s

Edit2 (Blender 2.48 - same as used on OS4):
PowerBook G4 1500 MHz, Blender 2.48a: 425 s
AmigaOne XE G4 1000 MHz, Blender 2.48: 591 s"

AOS4 users would be happy to get those Blender optimizations, as it seems to make Piru's Blender 2x faster.
- KimmoK
// Windows will never catch us now.
// The multicolor AmigaFUTURE IS NOW !! :crazy:
 

Offline KimmoK

  • Sr. Member
  • ****
  • Join Date: Jun 2004
  • Posts: 319
    • Show all replies
Re: Linux PowerPC Blender benchmark
« Reply #5 on: May 07, 2012, 09:25:32 AM »
According to shown benchmarks of this thread I wonder how anyone could see G5 as a good option for Amigalike OS. If it was true, that G4 performs better than liguid cooled G5, why would anyone want G5 over G4? For house heating purposes?


Some PowePC Results from fastest to slowest:
115s blender 2.43 G5 2.3Ghz Dualcore (erroneously reported singlecore?)
117s blender 2.44 G5 2.3Ghz Dualcore
117s blender 2.44 G5 1.3Ghz Dualcore (erroneously reported singlethread?)
120s blender 2.49 G5 2.3Ghz Dualcore
121s blender 2.49 G5 2Ghz Dualcore (erroneously reported singlecore?)
127s blender 2.49 G5 3Ghz Dualcore (erroneously reported singlecore?)
202s blender 2.5a PowerBook 1667Mhz
205s blender latest PowerBook 1667Mhz (Piru)
206s blender 2.48? x1000 1800Mhz Dualcore
223s blender 2.49 G5 2.3Ghz
234s blender 2.49 G5 2.1Ghz single core
235s blender 2.49 G5 3Ghz SingleCore
256s blender 2.42 PowerBook 1667Mhz
305s blender 2.43 PowerBook 1667Mhz
318s blender 2.44 PMac 1600Mhz
325s blender 2.49 PowerBook 1667Mhz
331s blender 2.49 PowerBook 1667Mhz
368s blender 2.43 MacMini 1420Mhz
420s blender 2.48? x1000 1800Mhz single core
425s blender 2.48a PowerBook 1500Mhz
540s blender 2.48 MacMini 1420Mhz
591s blender 2.48 A1XE 1000Mhz
1180s blender 2.48 SAM440 800Mhz (AOS4)


And when we observe with blender versions...

Then with blender 2.49 or 2.48 & dualcore:
120s blender 2.49 G5 2.3Ghz Dualcore
121s blender 2.49 G5 2Ghz Dualcore (erroneously reported singlecore?)
127s blender 2.49 G5 3Ghz Dualcore (erroneously reported singlecore?)
206s blender 2.48? x1000 1800Mhz Dualcore

Then with blender 2.49 or 2.48 & singlecore:
223s blender 2.49 G5 2.3Ghz single core
234s blender 2.49 G5 2.1Ghz single core
235s blender 2.49 G5 3Ghz SingleCore
325s blender 2.49 PowerBook 1667Mhz
331s blender 2.49 PowerBook 1667Mhz
420s blender 2.48? x1000 1800Mhz Single core
425s blender 2.48a PowerBook 1500Mhz
540s blender 2.48 MacMini 1420Mhz
591s blender 2.48 A1XE 1000Mhz
1180s blender 2.48 SAM440 800Mhz (AOS4)

With same blender version G4 is a slower than G5, but faster than PA6T.

Performance scaling etc.:
G5-2300single vs G4-1667  - 1.3 x Mhz - 1.4 x result - G5 better

G5-2300single vs PA6T-1800single - 1.28 x Mhz - 1.88 x result - G5 better

G4-1667 vs PA6T-1800single - 1.08 x Mhz - 0.77 x result - G4 better
G4-1667 vs PA6T-1800dual - 1.08 x Mhz - 1.6 x result - PA6T is better

The last part shows PA6T advantage, one day, after AOS is updated to use both cores.
« Last Edit: May 07, 2012, 09:35:05 AM by KimmoK »
- KimmoK
// Windows will never catch us now.
// The multicolor AmigaFUTURE IS NOW !! :crazy:
 

Offline KimmoK

  • Sr. Member
  • ****
  • Join Date: Jun 2004
  • Posts: 319
    • Show all replies
Re: Linux PowerPC Blender benchmark
« Reply #6 on: May 07, 2012, 01:44:40 PM »
Quote from: Linde;692036
...Then again, some people say that Amiga-like operating systems will only ever run on single cores if they want to retain some level of backwards compatibility. I have no idea if this is true, though.


To my understanding every Amigaflavour has plans to go multicore in the future (with and without sandboxing, etc...).
I hope it happens during my lifetime....

(I vote for compatibility break initially (perhaps nonSMP and SMP versions of kernell available in parallel for a while) while building the sandbox during longer period of time, if the compatibility need remain)
- KimmoK
// Windows will never catch us now.
// The multicolor AmigaFUTURE IS NOW !! :crazy:
 

Offline KimmoK

  • Sr. Member
  • ****
  • Join Date: Jun 2004
  • Posts: 319
    • Show all replies
Re: Linux PowerPC Blender benchmark
« Reply #7 on: May 07, 2012, 04:36:30 PM »
Quote from: Iggy;692058
@KimmoK
I'm not sure I get your point.
From your own quotes, the G5 comes off looking pretty good.


Yes.
But before that it seemed like G4 is as fast as single core G5. That's what I meant and that's why I collected more results to see more clearly.
More detailed study reveals that G5 actually is not that bad. Better than what I expected.
Also, in Bryce tests that Barefeats once did it seems that with optimizations G5 can be far ahead of G4. (up to 2x faster per core, even per Mhz, IIRC).
Time will tell if that kind of optimizations can be done on 32bit SW on AOS for PA6T.
- KimmoK
// Windows will never catch us now.
// The multicolor AmigaFUTURE IS NOW !! :crazy:
 

Offline KimmoK

  • Sr. Member
  • ****
  • Join Date: Jun 2004
  • Posts: 319
    • Show all replies
Re: Linux PowerPC Blender benchmark
« Reply #8 on: May 07, 2012, 09:48:07 PM »
G5 FSB...

Was it 32in + 32out hypertransport kind of bus?
For example on 2Ghz G5 the FSB ran at 1Ghz (2x4GB/s).

While G4 can do 1.3GB/s(one direction at a time) and PA6T 32 GB/s (IIRC 16+16).
- KimmoK
// Windows will never catch us now.
// The multicolor AmigaFUTURE IS NOW !! :crazy: