How fast is PearPC: Benchmarks

[Ribtorus]

I don't know the answer to that.

However I would expect that whichever host OS made best and most efficient use of the Host CPU, hard drive controller and memory bus would probably perform best at this stage. Obviously, sound and video performance of the Host OS is pretty irrelevent to PearPC and the guest OS.

Also, the Host OS which places the least overhead burden on the hardware may have an advantage as well.

Both linux and Windows XP/2000 can be leaned out by shutting down unneeded services. perhaps that's what I'll benchmark next.

I'll run the Photoshop, XBench and Bryce tests with an [i:4f07d62bc8]extremely[/i:4f07d62bc8] leaned out boot of Windows XP and see if that gives a measurable boost in performance.

[Ribtorus]

Running XBench, Photoshop and Bryce 5 on an Windows XP host with [b:d3c0558e99]all[/b:d3c0558e99] default services running showed no performance loss over a Windows XP boot that had [b:d3c0558e99]minimal [/b:d3c0558e99]services running. It would appear that Windows background services have no appreciable effect on PearPC performance. On my installation, anyway.

Xbench actually improved its overall score by about 6% when running on an XP host that had all the default services running.

Photoshop showed an almost identical score, as did Bryce 5.

Booting Windows with minimal services did not improve things in my tests, and may actually have lowered performance.

[Ribtorus]

After a bit more benchmarking and comparisons on the web, and considering a bunch of average scoring machines, I have determined that [b:0d676f5f87]Photoshop 7[/b:0d676f5f87] performance With PearPC and Panther on my Windows XP host, running either PearPC .3.1 or a .4pre build gives roughly the same performance as a G3 266 - G3 300 running OS X.

So far, using [i:0d676f5f87]Photoshop 7[/i:0d676f5f87], I see:

a [b:0d676f5f87]20X[/b:0d676f5f87] slowdown over the host machine.

a rough equivalent of a [b:0d676f5f87]G3 266-300[/b:0d676f5f87] with at least 512MB ram running OS X.x.x

a [b:0d676f5f87]4X[/b:0d676f5f87] slowdown compared to the [i:0d676f5f87]stated CPU speed[/i:0d676f5f87] shown in OS X's System Profiler.

My experience with Altivec builds has been spotty at best, with some apps failing to even run. I have limited my tests to the non-altivec builds. The exception has been XBench, which liked the Altivec build and showed a considerable performance increase.

With XBench, I see roughly a [b:0d676f5f87]4.5X[/b:0d676f5f87] slowdown over the [i:0d676f5f87]stated CPU speed[/i:0d676f5f87] in the System profiler when running an Altivec build, and a [b:0d676f5f87]8X to 9X[/b:0d676f5f87] slowdown when running a non-Altivec build. That's [i:0d676f5f87]very[/i:0d676f5f87] approximate.

I don't know if Xbench is a practical benchmark for comparison because of the huge variability in posted scores among the same model macintoshes. It is also not cross-platform, so I cannot relate its results to the host machine.

Bryce 5 shows a consistent slowdown over the host of about [b:0d676f5f87]19X [/b:0d676f5f87]to [b:0d676f5f87]20X[/b:0d676f5f87], Altivec build or not.

Strata3D CX (demo) is impractical to use as a benchmark at this point because the render time of the test images available to me take a very long time (over ten hours), and the application itself has taken over an hour just to load! Maybe some wierd Altivec thing causes this, as I had a similar experience with Quark and Photoshop.

[Guru3]

Very nice work.

[Guest]

A new benchmark:

Stuffit Expander on the host WinXP vs the guest 10.3.5 Panther.

I created a.sit archive with 99 jpgs at a size of 229 MB and timed how long to uncompress it.

Host machine: 1 min 5 sec

Guest OS: 23min 58 sec

That's a slowdown of 22X

That's in pretty good agreement with the other benchmarks so far.

[Ribtorus]

I re-ran the stuffit benchmark with the same 229 MB file, but this time I used the [b:0cbaaa743b]full[/b:0cbaaa743b] version of Stuffit 8, for both the Host Windows XP and the guest, Panther.

Quite a different score this time:

Host WinXP: create archive, 1 minute 28 seconds

Host WinXP: unstuff archive, 1 minute 12 seconds

Guest Panther: create archive, 3 minute 15 seconds

Guest Panther: unstuff archive 3 minute 28 seconds

Slowdown for creating an archive, [b:0cbaaa743b]2.2X[/b:0cbaaa743b]

Slowdown for unstuffing an archive, [b:0cbaaa743b]2.9X[/b:0cbaaa743b]

Either the full version of stuffit for the mac is [b:0cbaaa743b]significantly[/b:0cbaaa743b] faster than the stripped-down version included with Panther, or something was seriously wrong with my first trial. But I ran it three times and each time was consistent. I'm going to look into this discrepancy a bit more.


I think I have found the reason for the difference. When I create the archive in windows and move it to the mac, it takes a lot longer to unstuff. An archive created on the mac takes far less time (10X less).

[prasys]

Panther OSX (PearPC) create archive : 3min 23 sec

Panther OSX (iBook) create archive : 1min 49 sec

Its a 200 MB MP3 Files compresed in sit...Well , well.What can you tell...Both of them "rock"....

[HFU]

This is a very nice site. Good job everyone!

Ribtorus's dual Opteron is truely amazing for PPC.

Here are some comparisons between Linux and Windows on dual Xeon. I'd like to make another closer comparison between Linux and Windows when 0.40Pre source is available for download. It looks like 0.4Pre does run faster according to some preliminary XBench test. It even showed faster G3 emulated speed on guest system (1.36GHz vs 1.26GHz). Linux on the other hand seemed to run a bit faster on the file system test.

[b:1a3be32ccc]Host:[/b:1a3be32ccc] [url=http://www.ctechnet.com/hardware/Asus/NCCH-DL/Small/Xeon3.4GHz_PPC_OSX1.36GHz.jpg:1a3be32ccc]3.4MHz/800FSB/512k Xeon[/url:1a3be32ccc]
[b:1a3be32ccc]Guest:[/b:1a3be32ccc] PPC 0.3.1 on Linux Kernel-2.6.8smp | PPC 0.4Pre on WinXP/SP2

[b:1a3be32ccc]Panther 10.3.5:[/b:1a3be32ccc] 1.26GHz/G3 (Linux) | 1.36GHz/G3 (Windows)

[b:1a3be32ccc]CPU Test:[/b:1a3be32ccc] 0.55 | [b:1a3be32ccc]0.56[/b:1a3be32ccc]
- GCD Loop: 40.23 | [b:1a3be32ccc]43.98[/b:1a3be32ccc]
- Floating Point Basic: 2.67 | [b:1a3be32ccc]2.74[/b:1a3be32ccc]
- vecLib FFT: 0.15 | 0.15
- Floating Point Library: 13.57 | [b:1a3be32ccc]14.60[/b:1a3be32ccc]
[b:1a3be32ccc]Thread Test:[/b:1a3be32ccc] 24.09 | [b:1a3be32ccc]29.21[/b:1a3be32ccc]
- Computation: 58.83 | [b:1a3be32ccc]59.62[/b:1a3be32ccc]
- Lock Contention: 15.15 | [b:1a3be32ccc]19.34[/b:1a3be32ccc]
[b:1a3be32ccc]Memory Test:[/b:1a3be32ccc] 37.98 | [b:1a3be32ccc]44.85[/b:1a3be32ccc]
- System: 43.39 | [b:1a3be32ccc]55.85[/b:1a3be32ccc]
-- Allocate: 42.87 | [b:1a3be32ccc]57.82[/b:1a3be32ccc]
-- Fill: 45.03 | [b:1a3be32ccc]55.95[/b:1a3be32ccc]
-- Copy: 42.36 | [b:1a3be32ccc]53.91[/b:1a3be32ccc]
- Stream: 33.77 | [b:1a3be32ccc]37.47[/b:1a3be32ccc]
-- Copy: 37.74 | [b:1a3be32ccc]40.43[/b:1a3be32ccc]
-- Scale: 26.35 | [b:1a3be32ccc]28.62[/b:1a3be32ccc]
-- Add: 36.54 | [b:1a3be32ccc]42.11[/b:1a3be32ccc]
-- Triad: 37.54 | [b:1a3be32ccc]42.83[/b:1a3be32ccc]
[b:1a3be32ccc]Disk Test:[/b:1a3be32ccc] [b:1a3be32ccc]92.37[/b:1a3be32ccc] | 81.70
- Sequential: [b:1a3be32ccc]53.38[/b:1a3be32ccc] | 49.82
-- Uncached Write: 30.44 | [b:1a3be32ccc]31.06[/b:1a3be32ccc]
-- Uncached Write: [b:1a3be32ccc]208.54[/b:1a3be32ccc] | 129.05
-- Uncached Read: [b:1a3be32ccc]30.58[/b:1a3be32ccc] | 28.93
-- Uncached Read: [b:1a3be32ccc]218.33[/b:1a3be32ccc] | 173.0
- Random: [b:1a3be32ccc]342.60[/b:1a3be32ccc] | 227.0
-- Uncached Write: 201.65 | [b:1a3be32ccc]202.24[/b:1a3be32ccc]
-- Uncached Write: [b:1a3be32ccc]352.01[/b:1a3be32ccc] | 125.96
-- Uncached Read: [b:1a3be32ccc]599.82[/b:1a3be32ccc] | 545.85
-- Uncached Read: [b:1a3be32ccc]452.88[/b:1a3be32ccc] | 344.13
[b:1a3be32ccc]Quartz Graphic Test:[/b:1a3be32ccc] N/A
[b:1a3be32ccc]OpenGL Test:[/b:1a3be32ccc] N/A
[b:1a3be32ccc]User Interface Test:[/b:1a3be32ccc] N/A

HFU

[Ribtorus]

Good stuff.

I ran Cinebench 2003.

Here's the results



CINEBENCH 2003 v1
****************************************************

Tester : Rib

Processor : PearPC
MHz : 1006
Number of CPUs : 1
Operating System : Panther

Graphics Card : PearPC
Resolution : 1024x768
Color Depth : 32 bit

****************************************************

Rendering (Single CPU): 6 CB-CPU
Rendering (Multiple CPU): --- CB-CPU


Shading (CINEMA 4D) : 8 CB-GFX
Shading (OpenGL Software Lighting) : 8 CB-GFX
Shading (OpenGL Hardware Lighting) : 8 CB-GFX

OpenGL Speedup: 1.00


Time to render the final image was 4770.8 seconds.

****************************************************
Cinebench can also be run in Windows, so I can see its slowdown factor over the host machine. In my case, the slowdown for the single CPU render is [b:7986cd14a8]50X[/b:7986cd14a8].

That's more than twice the slowdown factor for Photoshop and Bryce, but more along the lines of Strata 3D CX. Perhaps this is an OpenGL issue?

My CineBench score is approximately [b:7986cd14a8]4.5X[/b:7986cd14a8] slower than an actual G3 300 running Jaguar, in the CPU benchmark, but only about [b:7986cd14a8]1.8X[/b:7986cd14a8] slower in the OpenGL rendering than the same G3 with a Rage 128 video card.