[LightWaver-67]

Okay... I'm trying to mentally justify an upgrade from my current system to a newer one. I really would like to understand what kind of speed increases I "may" see by doing so.

Now, by using "dumb" math, I can make some guesses. I am running a G4 (AGP) 533Mhz with 1.5 Gb of RAM. Now, to upgrade to a G5 dual 2.0 Ghz machine using "dumb" math, I figured:
2 X 2.0Ghz = 4.0Ghz
1Ghz = 1,024Mhz
4.0Ghz = 4,096Mhz
4,096Mhz ÷ 533Mhz (my machine) = 7.68 (rounded-up)

So using "simple math"... on paper, a dual 2Ghz should be 7.7-times faster than my current machine (processor speed only). Now, we know that's not REALLY how it works... but it gets me in a "ballpark".

Here's what I'm trying to figure out:
Currently, on this 533Mhz machine, it takes roughly 22-24-Hours to encode 2-Hours of video footage using a two-pass variable bit rate encoding with a decent bit rate. I am trying to guess what possible speed gains I would see from upgrading. Would it really be close to 8x the speed, thus bringing me down to 3-hours of encoding...? or would it flatten-out to something like maybe only 4x as fast and only yield me 6-7-Hour encoding time...?

I would love to be able to justify such a purchase, but I don't want to jump-in to a new tower and sink a few/several grand into it, only to get modest increases in encoding time. It needs to be substantial in order for me to consider it. Again, if it only bumped from 24-Hours of encoding to 16-hours... I'd be extremely ticked-off at myself for spending the $$$ so soon. I'd rather wait for the next rev.

Any others have 'better' math results for me to ponder...?

[Detrius]

This isn't a simple question with a simple answer.

The 533MHz G4 was actually about as fast as the 733MHz G4. Motorola had to do some tricks to get the MHz bump we all wanted. (this means the 533MHz was faster than the 667MHz). Based on this, you could say that a dual 1.4GHz G4 is going to be about 4x faster than your machine in processor speed.

Going from there to the G5... well, the dual 2GHz G5 ranges from 50% to 110% faster than the dual 1.4GHz, according to www.barefeats.com . The biggest speed boost was in the straight-up processor intensive applications. Basically, the dual 2GHz G5 would probably be 6 to 8 times faster than your machine, depending on what you are doing.

That's a pretty big difference. That turns 24 hours into 3 to 4 hours. Of course, you also have to consider where your bottleneck is... is it the processor, RAM, or hard drive? Those all make a difference.

[MindFad]

Oh, how I long for thee, dual 3GHz G5.

[LightWaver-67]

Originally posted by Detrius:
Of course, you also have to consider where your bottleneck is... is it the processor, RAM, or hard drive? Those all make a difference.

My best presumption is processor speed, because when I'm encoding the video, the CPU usage is maxed-out in the Compressor application, hard drive activity is average... the light flickers but is not constantly lit like it is during capture. I am maxed-out to 1.5GB of RAM... so my guess is that procesor speed and/or bus speed will be major advantages to my needs.

Also, looks like my "dumb" math was coincidentally similar to your assesment of speed gains.

If I really could get it so that I could encode 2-Hours of footage as a 2-Pass VBR and have it done before I get-up in the morning... I would be a VERY happy camper...!!! I could even live with 6-Hours... but again, I don't know for a FACT that it can be done that quick until either I try it... or someone else here who DOES use 2-Pass VBR and a G5 can confirm its speed.

Thanks.

[Commodus]

Just remember: a G5 update is likely within a month (possibly this Tuesday, even), so you might not want to rush out and buy a dual 2 GHz machine even if it does end up being dramatically faster. A dual 2.5 would leave no question as to whether or not you can justify a new system.

[Luca Rescigno]

2 GHz is not the same as 2048 MHz. That's only for RAM and hard drive space, where they should actually use the "gibi" prefix instead of the "giga" prefix.

Detrius, why do you say the 733 MHz G4 was the same speed as the 533 MHz one? The 7450 that the 733 uses has four AltiVec units, as opposed to the 7410's two. I'd think that the 667 and 733 would be faster per MHz than the 466 and 533. But I haven't actually seen benchmarks comparing them so maybe I'm wrong.

Having dual processors also does not mean you get twice as fast a machine. My dual 450 isn't as fast as my brother's single 867... I'd say it's between a single 533 and the 867. There are other benefits to dual processor machines - for example, less slowdown when doing a processor-intensive task, or when doing many things at once. But a dual 2 GHz G5 is not going to behave the same as a single 4 GHz G5.

Wait for the next G5 revision, then buy.

[Catfish_Man]

Originally posted by Luca Rescigno:
2 GHz is not the same as 2048 MHz. That's only for RAM and hard drive space, where they should actually use the "gibi" prefix instead of the "giga" prefix.

Detrius, why do you say the 733 MHz G4 was the same speed as the 533 MHz one? The 7450 that the 733 uses has four AltiVec units, as opposed to the 7410's two. I'd think that the 667 and 733 would be faster per MHz than the 466 and 533. But I haven't actually seen benchmarks comparing them so maybe I'm wrong.

Having dual processors also does not mean you get twice as fast a machine. My dual 450 isn't as fast as my brother's single 867... I'd say it's between a single 533 and the 867. There are other benefits to dual processor machines - for example, less slowdown when doing a processor-intensive task, or when doing many things at once. But a dual 2 GHz G5 is not going to behave the same as a single 4 GHz G5.

Wait for the next G5 revision, then buy.

Both model G4s had 4 vector units, but the 740x and 741x had some instruction dispatching restrictions that made them a bit weaker than the 745x. However, the 745x also had a pipeline 3 stages longer, and the early ones had some weird problems that made them a bit slower than one would think. I'm not sure that a 533MHz 7410 equals a 733MHz 7450, but it's faster than it seems (especially for non-altivec code).

[Lateralus]

Originally posted by Catfish_Man:
Both model G4s had 4 vector units...

If that is true, then you are the only person I have ever heard say it. From every other account, the 7450 has 4 units and the 7400 had 2.

[reader50]

Most of this info comes from Hannibal's execellent CPU analysis articles on Ars.

All of the 74xx series CPUs (so far) have four vector execution units:
Vector Permute
Vector Simple Integer
Vector Complex Integer
Vector Floating Point

The 7400/7410 have a dispatch limitation: one dispatch queue feeds the Permute unit, and the other dispatch queue feeds the other three units. Thus, you could have 2 vector instructions run at the same time, but only if one is a Permute command.

The 744x/745x have separate dispatch queues feeding each unit, so you could theoretically execute four vector instructions in parallel - provided you had to execute one instruction of each type.

Thus Apple could say that the new CPUs had "twice the AltiVec", but I do not remember them actually saying there were twice as many execution units. It would be more correct to say they had twice the maximum theoretical execution capacity - but many other factors limit this improvement to a more modest boost in performance. Besides, advertising that they had "twice the maximum theoretical execution capacity under ideal conditions" doesn't seem very catchy from a marketing standpoint.

In practice, the CPU & Memory bus bandwidths tended to be the limiting factor on vector performance. Even doubling the CPUs did not come close to doubling the AltiVec performance - I seem to recall the bump was more like 50%. The PPC970 has the same AltiVec dispatch limitation as the early G4s, but performs wonderfully on AltiVec code because of wide system bandwidth and higher clock speed.

The 7400/7410 have 4-stage execution pipelines while the 744x/745x has a 7-stage pipeline (I'm not sure on the 7447/7457). The longer pipeline means longer instruction latency and longer stalls from branch mispredicts, this caused the early 7455's to have lower performance per clock until Apple added the L3 cache. Thus, the early 733MHz G4 tower without L3 (7-stage CPU + 256K L2@733) might well be equivalent to a 533 G4 tower (4-stage CPU + 2MB L2@266). More branch misses plus a smaller available cache means more unexpected code fetches from main memory - this is where the real delay happens.

This is for average use with mixed threads - typical of a home system's usage. In practice, certain types of code, such as compact DC clients that fit entirely in the L1/L2 cache, will benefit from the full clock speed - provided nothing much else is running to bump them out of cache.

[Luca Rescigno]

But the early 733 MHz G4 tower DID have L3 cache, 1 MB of it. So did the 667 MHz model. It was the LATER G4/733 that had no L3 - performance suffered in the Quicksilver 733, 800, and 867 (educational), because they only had 256 kb of L2 cache, without any L3 cache at all.

Another thing I've heard is that the G5s aren't as great at AltiVec as the G4s. As in, a dual 1.25 GHz G4, despite having the choke-a-rific 167 MHz bus, is still better than the 1.6 GHz G5. G5s also have a longer pipeline than the G4s, so that could hurt performance slightly.

Anyway, thanks for the great explanation. Very helpful for someone like me who doesn't know much about the inner workings of CPUs.

[LightWaver-67]

My bad... it was my assumption that the Mhz numbers were... oh.. duh.. of COURSE they're not based on 1,024... why was I thinking that...?

Oh well...

Either way... I don't plan on buying a new tower until the next rev. of faster machines. I would NEVER buy this long after a release date. I would rather wait until the next batch of speed-bumps hit the store shelves.

Thanks for all the info.

[willab]

Originally posted by LightWaver-67:
My bad... it was my assumption that the Mhz numbers were... oh.. duh.. of COURSE they're not based on 1,024... why was I thinking that...?

Oh well...

Either way... I don't plan on buying a new tower until the next rev. of faster machines. I would NEVER buy this long after a release date. I would rather wait until the next batch of speed-bumps hit the store shelves.

Thanks for all the info.

If you tell me what bit rate to use I will try it for you on my dp 1.8.

[Detrius]

Originally posted by Luca Rescigno:
Detrius, why do you say the 733 MHz G4 was the same speed as the 533 MHz one? The 7450 that the 733 uses has four AltiVec units, as opposed to the 7410's two. I'd think that the 667 and 733 would be faster per MHz than the 466 and 533. But I haven't actually seen benchmarks comparing them so maybe I'm wrong.

Hate to say it, but this is old news.


http://www.barefeats.com/G4733.htm
http://www.barefeats.com/GGG.html
http://www.barefeats.com/pm05.html


Okay, I admit, my memory was a wee bit biased towards my purchase of the dual 533. However, the single 533 tests virtually identical to the single 667. Also, the dual 533 spanks a single 967... but two processors aren't supposed to be twice as fast...

Like the guy said... the Dual 533 was cheaper than the single 667. Now why would I not have bought the 533 (other than the fact that OS X hadn't been released yet)?

I haven't seen any recent benchmarks that take things like QE or gcc 3.3 into account.