Mac Pro 3GHz Benchmarks
Right after we started receiving results for Mac Pros at 2.66GHz, we’ve started receiving Mac benchmark benchmark results for Mac Pros at 3.0GHz. Since we’ve already compared the mid-range Mac Pro against the top-of-the-line Power Mac G5, we thought we’d compare the top-of-the-line Mac Pro against the same Power Mac G5. Let’s see if the 3.0GHz Mac Pro fares better than the 2.66GHz Mac Pro against a Quad Power Mac G5.
Setup
Here is the configuration of our two test machines:
- Mac Pro
- Intel Xeon 5160 @ 3.0GHz (two dual-core processors)
- 4096MB RAM
- Mac OS X 10.4.7 (Build 8K1079)
- Geekbench 2006 (Build 190)
- Power Mac G5
- PowerPC G5 @ 2.5GHz (two dual-core processors)
- 1024MB RAM
- Mac OS X 10.4.7 (Build 8J135)
- Geekbench 2006 (Build 180)
As per usual, we’re using the baseline scores (not the raw scores) from each benchmark. We’ve also computed the Mac Pro’s score as a percentage of the Power Mac G5’s score. Higher is better.
Results
Overall Score
| Power Mac G5 | Mac Pro | |
|---|---|---|
| Overall Score | 278.3 | 384.1 (138.0%) |
Integer Performance
| Benchmark | Power Mac G5 | Mac Pro |
|---|---|---|
| Emulate 6502 single-threaded scalar |
157.9 | 186.1 (117.9%) |
| Emulate 6502 multi-threaded scalar |
622.4 | 734.7 (118.0%) |
| Blowfish single-threaded scalar |
164.1 | 261.9 (159.6%) |
| Blowfish multi-threaded scalar |
651.5 | 1044.3 (160.3%) |
| bzip2 Compress single-threaded scalar |
157.0 | 250.7 (159.7%) |
| bzip2 Compress multi-threaded scalar |
572.5 | 964.9 (168.5%) |
| bzip2 Decompress single-threaded scalar |
152.8 | 274.4 (179.6%) |
| bzip2 Decompress multi-threaded scalar |
603.0 | 1152.0 (191.0%) |
Floating Point Performance
| Benchmark | Power Mac G5 | Mac Pro |
|---|---|---|
| Mandelbrot single-threaded scalar |
156.5 | 202.4 (129.3%) |
| Mandelbrot multi-threaded scalar |
623.5 | 808.5 (129.7%) |
| Dot Product single-threaded scalar |
143.2 | 408.0 (284.9%) |
| Dot Product multi-threaded scalar |
513.9 | 1214.9 (236.4%) |
| Dot Product single-threaded vector |
152.8 | 171.5 (112.2%) |
| Dot Product multi-threaded vector |
504.7 | 440.3 (87.2%) |
| JPEG Compress single-threaded scalar |
156.7 | 179.6 (114.6%) |
| JPEG Compress multi-threaded scalar |
621.6 | 719.6 (115.8%) |
| JPEG Decompress single-threaded scalar |
160.1 | 172.7 (107.9%) |
| JPEG Decompress multi-threaded scalar |
508.1 | 599.8 (118.0%) |
Memory Performance
| Benchmark | Power Mac G5 | Mac Pro |
|---|---|---|
| Read Sequential single-threaded scalar |
171.2 | 346.0 (202.1%) |
| Read Sequential multi-threaded scalar |
53.4 | 183.1 (342.9%) |
| Write Sequential single-threaded scalar |
196.2 | 556.7 (283.7%) |
| Write Sequential multi-threaded scalar |
185.1 | 272.7 (147.3%) |
| Stdlib Allocate single-threaded scalar |
119.6 | 320.8 (268.2%) |
| Stdlib Allocate multi-threaded scalar |
31.3 | 40.7 (130.0%) |
| Stdlib Write single-threaded scalar |
270.3 | 120.8 (44.7%) |
| Stdlib Write multi-threaded scalar |
367.4 | 206.4 (56.2%) |
| Stdlib Copy single-threaded scalar |
187.6 | 259.4 (138.3%) |
| Stdlib Copy multi-threaded scalar |
259.2 | 421.9 (162.8%) |
Stream Performance
| Benchmark | Power Mac G5 | Mac Pro |
|---|---|---|
| Stream Copy single-threaded scalar |
176.5 | 214.9 (121.8%) |
| Stream Copy multi-threaded scalar |
306.2 | 361.2 (118.0%) |
| Stream Copy single-threaded vector |
190.6 | 210.6 (110.5%) |
| Stream Copy multi-threaded vector |
277.4 | 332.3 (119.8%) |
| Stream Scale single-threaded scalar |
165.7 | 231.8 (139.9%) |
| Stream Scale multi-threaded scalar |
321.5 | 390.2 (121.4%) |
| Stream Scale single-threaded vector |
185.9 | 212.3 (114.2%) |
| Stream Scale multi-threaded vector |
274.9 | 340.2 (123.8%) |
| Stream Add single-threaded scalar |
147.0 | 192.8 (131.2%) |
| Stream Add multi-threaded scalar |
301.9 | 353.6 (117.1%) |
| Stream Add single-threaded vector |
188.4 | 209.9 (111.4%) |
| Stream Add multi-threaded vector |
290.7 | 341.5 (117.5%) |
| Stream Triad single-threaded scalar |
144.4 | 195.3 (135.2%) |
| Stream Triad multi-threaded scalar |
305.4 | 354.9 (116.2%) |
| Stream Triad single-threaded vector |
195.6 | 170.4 (87.1%) |
| Stream Triad multi-threaded vector |
307.2 | 275.7 (89.7%) |
Conclusion
Overall, the 3.0GHz Mac Pro is 35% faster than the Quad Power Mac G5 (and 28% faster than the 2.66GHz Mac Pro); it’s up to 242% faster in particular individual tests (and, unfortunately, up to 100% slower in others). Just like the 2.66GHz Mac Pro, the 3.0GHz Mac Pro excels at integer calulations and floating point calculations. Unlike the 2.66GHz Mac Pro, the 3.0GHz Mac Pro also excels at memory operations, although this might have more to do with the number of DIMMs installed in each machine rather than the processing power of each machine (the 3.0GHz has more DIMMs, and thus has a wider memory bus).
Vector Performance seems to be a weak point, despite the keynote claim the Xeon’s vector processing unit is more powerful than the G5’s. I’m surprised that performance is so erratic. stdlib memory access in particular is way too slow.
Considering the chip is so new, maybe GCC (if that is what you used) hasn’t been properly tuned for it yet. The official intel compilers might have a better idea what they are working with.
If the compiler is the problem, then we will have to wait a little longer to find out what these machines are really capable of.
Do you not think that testing a quad G5 with only 1Gb RAM against a Xeon with 4Gb skews the results a little? I’d like to see more realistic comparisons – I recognise the Xeon should trounce the G5, but having owned a quad for 6 months, I know that memory makes all the difference to that machine!
Regardless, thanks for the insight into the new Mac’s performance…
[OT] your top scores are littered with some invalid results. Maybe you should block faster-than-light “computers”?
Alan,
Geekbench uses about 100MB of memory; so long as a computer has more than 100MB of memory free, the total amount of system memory shouldn’t affect the results. The difference in performance between the G5 used in the comparison above and the fastest G5 we’ve seen is about 3%!
kkl,
We’ve found that the timers Geekbench for Windows uses are inaccurate when run on a virtual machine (e.g., Parallels). We’re looking into a solution for this issue.
in the last test you said: “Of course, we’ve tested a mid-range Mac Pro ($2500 US) against a top-of-the-line Power Mac G5 ($3000 USD), so even though the Mac Pro is only a little bit faster than the Power Mac G5, it’s significantly cheaper!”
You can’t think of it this way! you are comparing last years top-of-the-line machine against this years. Plus 500mhz is like in 2004 when the 2.5 ghz G5 was introduced. The test of the 2.66 against the G5 quad was a fair comparison because Apple don’t use the newest available G5 processor.
Perhaps I’m missing something, but it seems like the percentage performance increases over the G5 machines are pretty much in line with the percentage increase in clock speed of the G5 – which to me seems like a testament to the good design of the IBM chip in the old machine – that a design that old, per clock cycle is pretty much as good as the bleeding edge up-to-date Intel chip. Even the power consumption reduction mostly could be down to the much smaller fabrication process used by Intel, although I don’t know the details of either.
Its a shame IBM decided they didn’t want to provide Apple with what they wanted – a decision in light of how popular Apple is becoming they may regret.
I guess they thought they had enough on their plate developing chips for the PS3, Xbox 360 and Wii…
Here’s a question…I’ve seen comparisons all over the place between the quad power pc g5 and the new mac pro 3 ghz g5. But here’s the thing: I’ve got an “older” dual power pc g5 (bought a little less than 1 year ago), and I’m wanting to know how much faster the new mac pro is than that box. That to me seems like a comparison between an older machine and a newer one that I’d really be interested in knowing, when you consider that people like me with those “older” and slower boxes are probably the ones most excited about a new upgrade. The folks who only 3-6 months ago bought a quad g5 aren’t probably rushing out to spend another large sum of money for “modest” perfomance improvement. So anyone got some numbers on that? Thanks!
A note on the operations measured.
Many of the tested loops are apparently hand written simple loops doing the same thing over and over. Which is normal for benchmarks. Sad thing is they are written poorly.
As an example the 2.5GHz G5 vector unit can sustain about 19 Gflops when calculating a dot product. Since the benchmark uses a single accumulator and has no loop unrolling whatsoever, it only reaches a mere 2.1Gflop. (”vmaddfp” has a latency of 8, so the loop is limited by 2.5GHz * 8 [flop/cylce] / 8 [cycles/instruction] -> about 1 flop/cylce instead of 8 flop/cycle) The intel chip can do these additions with the same overall throughput. But as they have a shorter latency, they are less affected by the programming style.
So take the number here with a lot of salt as they do not represent what the machines are capable of. They are still useful, since they represent what poor programming yields on them;-)
PS: The maximum vector throughput of these two machines is about the same (per cycle). Which is a huge improvement over previous intel chips.
I dunno, I can see the Xeon’s faster but if you think about it, it has twice the graphics mem and faster acces to it via the newer video card-plus it has twice the memory bandwidth (256vs128) AND it’s 500MHz faster. Yet it only achieves a 33 percent performance increase. Makes me wonder how the G5 would fare if Apple had the same, new 256bit wide RAM backing it up.
Additionally, Freescale is about to come out with 2.0+GHz G4 dual core chips which use less power (and make less heat!) than their new intel counterparts. I think that, long term, Apple kinda sold out to grab a few more MHz right now, oh and the cowtowing to Microsoft Windows I guess doesn’t hurt either.
I think its only a matter of time before the new, Vista capable PeeCees with their similar EFI BIOS are hacked (or OS X is hacked, or both) so that it will be possible to just buy a clone PeeCee and stick your pie rat copy of X on there. Bye-Bye Apple profit margins on OS X! And Apple doesn’t have that huge business sector monopoly for the license fees to soften the blow either like M$ does.
They’re feeling all phat now because iPod music sales are huge, but who knows how long that’ll last? I think that, long term, it was a mistake to make it potentially so easy to end their unique quality of Apple controlled hardware running Apple made OS. More than one Mac geek has told me that can’t happen, that you can’t pirate OS X onto non-Apple intel hardware. But I don’t think that’s the case, hackers can be pretty ingenious. When X can be shoehorned onto any beige PC, it will lose a lot of its cache’ among the non-Mac elitists who will then run it for free. FWIW I’m not being a hater here, I like Mac and OS X, its what I use daily. I won’t touch Windows with my hands if I can help it.
I think Apple will one day be sorry they did this. Even if it is faster.
Wait a second, it’s nothing new to run OS X on normal PCs… Intel and AMD. There are a few problems, you can’t simply update it, some software won’t run (mostly those relying on the video card), and for most video cards there are no drivers (same with sound cards). Driver support really is the biggest issue. So, yes, you can run OS X on normal PCs, but it will be nowhere near as nice as the real thing. Think of it as a trial, maybe. There were many people who did this and have bought or a going to buy a Mac. And I don’t think this will change, unless Apple wants it to happen.
Do you think that Dual G4 can really beat the Intels? Just look at the numbers and take them times two (which would be really optimistic). They take less power, but they have less power too.