[ChillieMac]

The Power Mac G5 motherboard has brought us (in a limited way) faster memory architecture and more importantly a whopping 800 MhZ, 900 MhZ and 1 Gig Front Side Bus! This has been AT LEAST as important a leap to the Mac since getting to a 1 Gig processor. All of Apple's hardware consumer and desktop still have no more than 167 MhZ FSB! Why is that?! Is it cost? Does it cost more to build larger capacity FSB motherboards? Is it more or less expensive than having a large Level 2 cache? These two things along with faster standard (let me speak up STANDARD) hard drives would definitely speed up the Mac experience while waiting for processor speeds to catch up. Including 512 MB RAM STANDARD (1 chip, how much does Apple REALLY make off of people too impatient or too lazy to research and buy cheaper elsewhere?) would give the impression to new users (ie not too tech savy) that Macs are way faster than PCs or at least not slower as they have heard so often. Processor speed IS important, but isn't the ONLY factor orf a computer's speed. RAM speed and ammount, Front Side Bus, Video Card processor and ammount of it's memory, hard drive speed, optical drive speed and 1st, 2nd and someties 3rd Level cache memory.

Hadn't Apple made a single motherboard design during the first generation iMac in order to make all of their computers cheaper? (at least one for desktops, one for laptops). Shouldn't all of the Macs have very near idenical on the board abilities: FSB, RAM type used, PCI interface for expansion, sound chips, USB and Firewire and Bluetooth controlers, etc? The differences to each model to be made by the graphics card, number of PCI card slots, hard drives etc and of course form factor.

My point is, the G5 processor, while terrific, shouldn't make Apple hold back other Macs by not incorporating other features that make it fast to the entire Mac line.

Anyone with inside knowledge as to why low FSB (167 MhZ and lower) is normal on most Macs, please post a message.

[tooki]

It doesn't take inside knowledge: that's the well-documented limit of the G4 chip's bus.

tooki

P.S. It's Hz not hZ -- it's an abbreviation of Hertz, the name of the German physicist Heinrich Hertz, who proved electricity can be transmitted through electromagnetic waves. --> KHz, MHz, GHz, THz (well, that last one we don't use in computing, at least not yet)

[miksu]

It doesn't make any sense to develop any new mobos for G4. Desktops will be G5 "quite soon", whatever it means, but for sure, G5 is coming into eMac and iMac. Current laptop mobo is more than adequate. G5 board is best ever made for personal computers, workstations or even rather high end servers.

Apple has had same problem as AMD, megahertz count doesn't tell everything. That's actually something even Intel has admitted nowadays, since no new Intel processors will be marketed using megahertz, but instead model number. It also helps Apple and AMD, since Intel has to tell people that their new model 730 is actually faster than 3.4GHz Pentium 4, even still that 730 is less than 3.4Gig.

G5 has best bus ever used on any personal computer. Intel netburst is really only 200Mhz quadpumped, G5 has two(!) independent busses, one downstream, one up. That's 2x500MHz doublepumped. Most beatiful thing about IBM's solution is that you can easily build 2, 4, or say 8-processor system (even more), as they can easily have required inter-processor connectors. But U can't do that on Pentium 4 or any of it's Xeon versions. Four-processor Xeon systems do exist, but because of it's poor bus design, they are only marginally faster than two processor systems. They all share same bus bandwidth. Apple doesn't have this limit, nor does AMD. It will be even better, when IBM implements on-chip memory controller and for sure, it's coming.

I wouldn't be surprised, if Apple introduces 4 processor (970FX-based) system in a current or slightly taller tower case. Only restriction is currently limited availability of 970/970FX.

G4 isn't really bad chip. It's actually rather good. Biggest part of problem with it was that Moto made it, not IBM. We had faster G3s made by IBM, but they didn't have Altivec. Had G4 been made by IBM, it would have reached 1,5GHz two years ago... And have at least quad pumped 200MHz bus, which in PC marketing equals 800MHz.

[Leonard]

There are a few reasons why the consumer machines don't have a faster FSB.

- The G4 chip can handle a maximum 167Mhz FSB. It can't handle anything higher than a 167Mhz FSB and it can't handle a DDR bus.
- The current G5, the PowerPC 970 runs too hot to put it in a laptop, maybe even an iMac or eMac. The PowerPC 970FX or a future PowerPC 970 variant may solve this problem. But it's not ready yet.

Apple probably has motherboard designs ready for the various product lines should either of these barriers be solved. They probably have prototypes of the motherboards, but until they get a better G4 chip or a cooler G5, they can't do anything.

[Lateralus]

Originally posted by miksu:
G4 isn't really bad chip. It's actually rather good.

Yeah, the G4 is probably the most important PowerPC release to date.

And I have seen many tests that show that at the same clock speed, the G4 is faster than the G5.

And frankly, if the G5 were crippled to a 6:1 bus (Put the 1.6GHz G5 on 266MHz bus for example), the G4 would wipe the floor with it.

[Peder Rice]

Two reasons for that:

(1) The G4 has a much better Altivec unit, that runs circles around the G5s
(2) The 970 adds so many pipeline stages, that G5 to G4 is worthy of the comparison of the P4 to the PIII

[Thinine]

FUD!

The G5's AltiVec unit is only slightly worse than the PPC7450+ but the rest of the G5's architecture differences make the chips nearly even in AltiVec performance.

The G5 is vastly superior to the G4 is floating point performance and has slightly higher integer performance, per clock. At the same clock, properly optimized code will perform better on the G5. Of course, there are always situations where one chip is better than another.

[Link]

When the altivec enhanced version of seti@home comes out, I think the g4 might very well become THE processor to have much like what happened to rc5.

[Waragainstsleep]

I don't really know much about Altivec, other than it has great benefits for multimedia applications.
How can it speed up Fast Fourier Transforms? (Thats what SETI does with your CPU)

Given the importance of Fourier transforms in computing, I am surprised any use of Altivec to speed them up didn't happen long ago....

[Catfish_Man]

Originally posted by Waragainstsleep:
I don't really know much about Altivec, other than it has great benefits for multimedia applications.
How can it speed up Fast Fourier Transforms? (Thats what SETI does with your CPU)

Given the importance of Fourier transforms in computing, I am surprised any use of Altivec to speed them up didn't happen long ago....

Apple includes a vector FFT library with OSX, and has for a few years.

[cmoney]

show me a G4 that's running at 2GHz with a 1GHz bus. can't? then the point's moot!

doesn't matter if the G4 is a "better" processor at the same clock speed as a G5 since they don't run at the same clock speed. (btw, that's the same fallacy as the MHz myth, yeah the G4 is faster than a similarly clocked P4 but P4s run 2x or 3x the clock speed making the point moot.)

[Eug Wanker]

Originally posted by ChillieMac:
The Power Mac G5 motherboard has brought us (in a limited way) faster memory architecture and more importantly a whopping 800 MhZ, 900 MhZ and 1 Gig Front Side Bus! This has been AT LEAST as important a leap to the Mac since getting to a 1 Gig processor. All of Apple's hardware consumer and desktop still have no more than 167 MhZ FSB! Why is that?! Is it cost?

No, it's because Freescale's G4 does not support DDR.

Including 512 MB RAM STANDARD (1 chip, how much does Apple REALLY make off of people too impatient or too lazy to research and buy cheaper elsewhere?) would give the impression to new users (ie not too tech savy) that Macs are way faster than PCs or at least not slower as they have heard so often. Processor speed IS important, but isn't the ONLY factor orf a computer's speed. RAM speed and ammount, Front Side Bus, Video Card processor and ammount of it's memory, hard drive speed, optical drive speed and 1st, 2nd and someties 3rd Level cache memory.

It would be financially stupid to include 512 MB with every machine at this time. That would dramatically increase the cost of every machine. And it's not a big deal to run out and buy $60 worth of memory for your laptop.

L1 and L2 cache are integrated into the CPU, so that can't be changed unless Freescale changes it. Apple can add L3, but it's expensive, and it shouldn't be necessary for desktop usage there is adequate memory speed. Unfortunately, the G4 can't provide adequate memory speed support.

And frankly, if the G5 were crippled to a 6:1 bus (Put the 1.6GHz G5 on 266MHz bus for example), the G4 would wipe the floor with it.

That is a strange argument. One of the main points of the G5 is that it can run with better memory. The G4 cannot. In essence, you're arguing that if the G5 didn't have the features of the G5, then it would not be as good. Well... of course. That's obvious. Also, overall the G5 has other significant design advantages, so I don't buy your argument, even if the G5 were to run a 266 bus (which is still much faster than the fastest G4).

One big thing is the G5's excellent FP, which is MUCH faster than any G4, clock-for-clock. The G4's FP speed is so bad that Freescale refuses to publish any FP benchmarks.

While the G4 may be faster clock-for-clock with some integer code, it's not as good as you think it is. Freescale itself says that the G4 7455 1.25 GHz can get about 500 in SPECint2000, whereas IBM says the G5 970 1.8 can get 838. Even if the 7455 scaled perfectly up to 1.8 GHz it would only get 720. It would get a boost from the 7447A's increased cache size, but that would be negated by the fact that SPEC never scales linearly with clock speed.

As for Altivec code, the G4 7447A scores 2.8 MIPs/MHz, whereas the G5 970 scores 2.9 MIPs/MHz or possibly higher. So while the G4's Altivec unit is technically superior in many ways, the Altivec units of the G4 and G5 have performance clock for clock probably in the same ballpark in actual implementation, with the actual clock for clock speed advantage possibly going to the G5.

And finally, all of this is moot for desktops, since the G5 runs at so much higher GHz anyway. Where the G4 does well is low power, which is good for the embedded market and laptops (and maybe iMacs). But even here, Freescale is running scared, because the 970FX is already appearing, and it does pretty well in the power department too at the high end. And at the low end, IBM's G3 offerings do pretty well too both from the perspective of power and from the perspective of performance if Altivec isn't needed.

All is not lost however. Freescale has announced the e600, which is basically a G4, but with DDR support and higher GHz (up to 1.8 GHz). It will compete directly against the G5 970FX at those speed, but if Freescale can put it out relatively soon and with good power specs, then it may make for a fine chip for iBooks. By the time it's out though, I think it will already be too late for the iMac and PowerBook, which would have already have long since gone to the G5.

[Waragainstsleep]

So SETI@home doesn't currently use the OS X FFT library?

[djohnson]

Originally posted by Waragainstsleep:
So SETI@home doesn't currently use the OS X FFT library?

Not that I am aware of. I think that might in BOINC...

[Lateralus]

Originally posted by Thinine:
FUD!

The G5's AltiVec unit is only slightly worse than the PPC7450+ but the rest of the G5's architecture differences make the chips nearly even in AltiVec performance.

FUD!

My G4 Dual 1.47 gets nearly 12,000 in the AltiVec Fractal Carbon. The Dual 2 G5 gets nowhere near that.

The G5 inherited it's AltiVec subsystem from the 7400, and it is only about 1/2 as powerful as the AltiVec subsystem of the 7450.

[Eug Wanker]

Originally posted by PowerMacMan:
My G4 Dual 1.47 gets nearly 12,000 in the AltiVec Fractal Carbon. What does a G5 Dual 2.0 get?

Since when is Altivec Fractal a real world Altivec test?

The G5 inherited it's AltiVec subsystem from the 7400, and it is only about 1/2 as powerful as the AltiVec subsystem of the 7450.

Even Freescale would not be so bold as to claim that. The Altivec unit of the G4 7450 series definitely is a good one, but it does not exist independent of the rest of the chip. The G5 has an older Altivec design, but overall it's still a pretty good performer, given the overall design of the G5.

Check out this page and the conclusion:

I'm glad I got a chance to talk to IBM, because the discussion definitely cleared up some questions and concerns I had about the processor's design. The VMX/Altivec unit turns out to be quite a bit less restricted and more powerful than I'd initially thought, which is good news for the Mac platform. However, a vector unit design along the lines of the Motorola G4e would be even better (though how much better would depend on the application), and it's nice to see that IBM is exploring the possibility of moving in that direction.

[Lateralus]

Originally posted by Eug Wanker:
[B]Since when is Altivec Fractal a real world Altivec test?/B]

Since when isn't it?

Every other platform that uses G4s, Pegasos in particular, as well as Amiga, flaunts Fractal Carbon tests.

[Catfish_Man]

Originally posted by PowerMacMan:
FUD!

My G4 Dual 1.47 gets nearly 12,000 in the AltiVec Fractal Carbon. The Dual 2 G5 gets nowhere near that.

The G5 inherited it's AltiVec subsystem from the 7400, and it is only about 1/2 as powerful as the AltiVec subsystem of the 7450.

I'm afraid you're wrong.

On the G4+ (744x, 745x): instructions can be dispatched freely to any vector unit, and latencies are low, BUT, memory bandwidth and load store bandwidth are also low, and out of order execution is nonexistent (or at least almost nonexistent)

On the G5: vector simple integer operations can't be dispatched at the same time as permute instructions (iirc, the exact dispatch details are a little foggy). While this is a limitation, it's fairly slight. OTOH, memory and load store bandwidth are very high, and altivec relies on them heavily. Another slight potential weakness is that the weaker (than the G4) non-vector integer unit might not be able to do pointer arithmetic fast enough for certain vector algorithms. I don't think this has ever been demonstrated though.

Overall: It's possible to find tasks that will run faster on either, but I think the G4+ may be a little ahead on most, discounting clock frequency. Counting clock frequency, I doubt that most altivec tasks run faster on the G4+ (and I realize that RC5 and fractal demo do). Certainly the realtime h264 decoding that's been demoed recently could never have been done on an existing G4 (from what I've read, anyway).

[Eug Wanker]

Originally posted by Catfish_Man:
On the G4+ (744x, 745x): instructions can be dispatched freely to any vector unit, and latencies are low, BUT, memory bandwidth and load store bandwidth are also low, and out of order execution is nonexistent (or at least almost nonexistent)

On the G5: vector simple integer operations can't be dispatched at the same time as permute instructions (iirc, the exact dispatch details are a little foggy). While this is a limitation, it's fairly slight. OTOH, memory and load store bandwidth are very high, and altivec relies on them heavily. Another slight potential weakness is that the weaker (than the G4) non-vector integer unit might not be able to do pointer arithmetic fast enough for certain vector algorithms. I don't think this has ever been demonstrated though.

Overall: It's possible to find tasks that will run faster on either, but I think the G4+ may be a little ahead on most, discounting clock frequency. Counting clock frequency, I doubt that most altivec tasks run faster on the G4+ (and I realize that RC5 and fractal demo do). Certainly the realtime h264 decoding that's been demoed recently could never have been done on an existing G4 (from what I've read, anyway).

You might find this post interesting:

The one really impressive demo for the G5 at NAB was the 1920x1080 24p H.264 Main Profile @ 8 Mbps. Real time decoding of that has never been demonstrated before that I know of. Another vendor at NAB needed THREE dual Opteron boxes ganged together to do this, and it was quite glitchy.

And this:

One of the improvements in H.264 is that instead of using a floating point DCT, it uses a DCT-like algorithm that can be fully implemented in integer. This is good, because precision issues go away.

That said, there is a LOT of math to do with Main Profile H.264 - its MIPS per pixels per second are many times higher than for, say, MPEG-2. That the 2.0 G5 can do it at all is quite surprising to a number of codec engineers I talked to, and I haven't heard anyone say that this would be possible on a similarly configured Opteron.

Eug asked: "So, in terms of the H264 decoder, were the theoretical so-called limitations of the 970's Altivec unit a limitation at all in the real world?"

name99 (one of the programmers) answered:

For H264, not really. There are a few places where it would have been nice to be able to have a multiply-add operate in parallel with a simple altivec integer op, but not many. Much more of an issue is the complaint I have made before and will doubtless make again --- lack of zero cycle forwarding. It's a testament to the amazing OoO of the machine that the single-cycle forwarding doesn't impose more of a penalty, but it still bugs me, especially when the codec has long squirrelly pieces of code like entropy decode where every instruction depends on the previous one.

I can't say I understand all this stuff, but it does seem that the G5 does quite well for Altivec.

[Waragainstsleep]

So does the G5 use 64-bit Altivec? Or will it?

[Cadaver]

Originally posted by Waragainstsleep:
So does the G5 use 64-bit Altivec? Or will it?

The AltiVec design of the Motorola G4 is a 128-bit unit anyway.

[Catfish_Man]

Originally posted by Eug Wanker:
You might find this post interesting:And this:Eug asked: "So, in terms of the H264 decoder, were the theoretical so-called limitations of the 970's Altivec unit a limitation at all in the real world?"

name99 (one of the programmers) answered:I can't say I understand all this stuff, but it does seem that the G5 does quite well for Altivec.

Actually, I read that thread once, if not more than once, per day. It was the primary source for that post, as well as one of the major reasons I found the comment about "1/2 as powerful altivec" rather odd

<edit> in response to the above post:
He may have been asking about double precision support in Altivec. Current it can handle 4 32-bit numbers, but not 2 64-bit ones. From the thread linked to above I gather that it would be unlikely for double support to be added without widening Altivec to 256 bits so that you could have 4 64-bit numbers.
</edit>