Cheers, FormerLurker. That final paragraph on the importance of public perception was right on. I've been arguing for over a year that Steve's explaining away MHz helps nothing. As you said, it's preching to the choir. We need to make Macs appear fast, and be able to back up that claim. Then we'll be competitive.
<strong>Cheers, FormerLurker. That final paragraph on the importance of public perception was right on. I've been arguing for over a year that Steve's explaining away MHz helps nothing. As you said, it's preching to the choir. We need to make Macs appear fast, and be able to back up that claim. Then we'll be competitive.
Sorry, shakedown flub (what a way to make a bow...)
I've never been too terribly concerned with the PPC/Pentium comparisons, but when there's NO relative chip speed progress in and of itself, that's painful.
<strong>The G5 is the only way to not be ridiculous in the mhz war, by lenghtening the pipeline bus (14 stages in many rumors, unstead seven for the 7450), it's their only way to go above one ghz (still stuck at 867 mhz)
probabily in many aspect excepting video games (because of the bandwitch) the G5 will be slower at equal Mhz than the G4 (like the P4 versus the PIII).
Even with the G5 the mhz war is loose by Motorola the only goal is to not be too ridiculous and to find some benchmarks who will show the superiority of their chips (even if in most of the tasks the chip is slower ...)</strong><hr></blockquote>
1) The G5 will kick the shit of the G4.....I will personally guarantee that. And there will be no G4 at Equivalent MHZ.
2) The pipleline will probably not be 14 stages....
3) G4 chips will break one gig without increasing the pipeline.
[quote]I don't have any sources to quote or link, but I distinctly remember hearing, a while ago, that the G5 ran 32-bit natively, even though it is supposed to be a 64-bit processor. Thus, no recompilations necessary for OS X, or apps, classic or otherwise.
Someone please correct me if I'm wrong....
<hr></blockquote>
No, you are right. I heard the same thing, many times, from different people.
1) The G5 will kick the shit of the G4.....I will personally guarantee that. And there will be no G4 at Equivalent MHZ.
2) The pipleline will probably not be 14 stages....
3) G4 chips will break one gig without increasing the pipeline.
Wait and see....</strong><hr></blockquote>
1) I doubt that the G5 will beat the G4 at equal Mhz especially the 7410. Of course the G5 wil start clocking higher than the G4 (same as the 7450 starting at 733 vs the 7410 stuck at 533, but remember excepting altivec stuff the benchmarks of this too later chips where very similar...)
2) Have you got any info about the number of stage of the pipeline ?
3) well you are right undead the G4 will break the GHz wall, but is not able to go far beyond.
1) I doubt that the G5 will beat the G4 at equal Mhz especially the 7410. Of course the G5 wil start clocking higher than the G4 (same as the 7450 starting at 733 vs the 7410 stuck at 533, but remember excepting altivec stuff the benchmarks of this too later chips where very similar...)
2) Have you got any info about the number of stage of the pipeline ?
3) well you are right undead the G4 will break the GHz wall, but is not able to go far beyond.</strong><hr></blockquote>
"Getting to those clock speeds [1.6 GHz] involved increasing the G5's pipeline from the 7450's seven stages to ten.
Initial benchmarks supplied by our source suggest the G5 is pretty fast indeed:
GHz 1.2 1.4 1.6
SpecInt2000 987 1151 1340
SpecFP2000 1005 1173 1359
By comparison, Intel's 2GHz Pentium 4 has recorded SpecInt2000 and SpecFP2000 scores of 656 and 714, respectively, according to <a href="http://www.specbench.org"; target="_blank">www.specbench.org</a> If accurate, the G5 figures are impressive indeed. "
This compares to
POWER4 (1 core, 1.3 GHz) 814 int, 1169 fp
Athlon MP1800+(1,53 GHz) 607 int, 547 fp
If the numbers for G5 are roughly right, it runs circles around any G4!
I do not know what their plans are with some of the earlier features I spoke of earlier in the year, but they must have made some big changes to their schedule.
I expect them to blow everyone away at MWSF.
The current Quicksilver' s are obviously buying them 6 months more time. More time that will be completely worth it. Those models are crippled versions of what I earlier have worked with. More on specifics later.....
As excerpted <a href="http://forums.appleinsider.com/cgi-bin/ultimatebb.cgi?ubb=get_topic&f=1&t=000012&p="; target="_blank">here</a> by havanas, Dorsal claimed to be working on machines with DDR RAM, ATA/100, integrated AirPort, and various other enhancements over the current offering all the way back before MWNY. The board didn't materialize, but the processors he mentioned did.
As I recall, there has also been a case redesign in the works.
Hmm. I don't know for sure, obviously, but I seriously doubt that a desktop G5 is going to come anywhere near the performance of IBM's super-hot, super HUGE ultimate floating point optimized Power4. This point alone tends to cast doubt on the veracity of the Register's claims. If true, the 1.6GHz G5 would match the performance of the Athlon 4400+ nearly a year before the Athlon's introduction. I tend to doubt this, though...
Still, I suppose that almost anything is possible.
Those mythical SPEC scores for the G5 are BS. If you do the math, it's scaling superlinearly, which is totally impossible.
Just to get perfectly linear scaling, you would need to make sure that absolutely every scrap of code is housed in cache; for which you'd need a few hundred megs of it. And even then you wouldn't get linear scaling, but you'd get closer to it.
Superlinear scaling, however, could not happen under any situation in a given rigidly defined set of circumstances -- i.e. same motherboard, same hardware, same programs, same CPU revision; same everything except clockspeed.
Edit: Furthermore, you would not do a SPEC run on anything but a final production chip and motherboard, because otherwise it would potentially be unstable and ruin your run. Also, the SPEC source code isn't free.
Altogether, it's BS. My guess is the Register's "source" pulled those numbers directly from his ass without thinking too clearly about them.
[Those mythical SPEC scores for the G5 are BS. If you do the math, it's scaling superlinearly, which is totally impossible.<hr></blockquote>
The linear variance is < 2% which is realistic as run to run fluctuations. It is also possible and maybe even likely that these are chips from different batches with different revisions and internal tweaks.
[quote]Just to get perfectly linear scaling, you would need to make sure that absolutely every scrap of code is housed in cache; for which you'd need a few hundred megs of it. And even then you wouldn't get linear scaling, but you'd get closer to it.<hr></blockquote>
Not necessarily true at all. It depends heavily on the nature of the data that is being used for the calculations as well as the compiler used to create the code. If the data displays good temporal and spacial locality and the compiler uses pre-fetch instructions there is no reason for the processor to ever have to wait for a memory fetch. Of course we cannot know if this is the case or not, but if the benchmark number ratios are accurate I wouldn't be surprised.
Assuming that Motorola did in fact run these tests it is possible that they didn't run the entire suite and are basing the numbers on just a few of the tests. In any case I'm sure Motorola has put them into motherboards at some point at least, especially if they are shipping engineering samples to Apple for testing.
<strong>I do not know what their plans are with some of the earlier features I spoke of earlier in the year, but they must have made some big changes to their schedule.
I expect them to blow everyone away at MWSF.
The current Quicksilver' s are obviously buying them 6 months more time. More time that will be completely worth it. Those models are crippled versions of what I earlier have worked with. More on specifics later.....
dorsal</strong><hr></blockquote>
It's been a while since Apple "blew" anyone away with features. A few I can think of were AirPort (with the iBook), Firewire (with the B&W G3). Maybe the dual-processors, too. Design, Apple's always done well with that (love the TiBook). Features-- I won't be holding my breath.
1) I doubt that the G5 will beat the G4 at equal Mhz especially the 7410. Of course the G5 wil start clocking higher than the G4 (same as the 7450 starting at 733 vs the 7410 stuck at 533, but remember excepting altivec stuff the benchmarks of this too later chips where very similar...)
2) Have you got any info about the number of stage of the pipeline ?
3) well you are right undead the G4 will break the GHz wall, but is not able to go far beyond.</strong>[/QUOTE]
Powerdoc, how can you really comment on the relative speed of the G5 as it relates to the G4. By coincidence, you may be correct. However, without seeing the white paper on the G5 design, all speed predictions are completely guesses.
All things being equal, yes, a longer pipeline would mean slower mhz per mhz. However, if additional processing units are added to the chip, such as extra INT of FP units, then these very well offset the extra pipeline length. We don't know what affect Rapid IO will have, etc.
I don't really believe those published SPEC results for the G5, but, for the sake of argument, if they are even close to being true, then the G5 will be MUCH faster than the G4 mhz for mhz.
Also, regarding the 7410 at 533mhz and the 7450 at 733mhz... The 7450 scales proportionally against the 7410 in Altivec and Int operations. It just doesn't scale as well at FP. Of course, that's because additional INT and Altivec units were added, but FP was left the same but with a longer instruction pipeline.
[quote]Powerdoc, how can you really comment on the relative speed of the G5 as it relates to the G4. <hr></blockquote>
How can anyone really comment on it?
It's just a bunch of rumors. Anyone with any licence to say anything is most likely constrained by NDA's. Anyone else is just speculating or passing on third hand information.
All in all, If I had to make my own educated guess on who's predictions are right, I'd say Powerdoc's more on the right track.
Will the G5 be faster than the G4, yes.
1.2-1.6 Gighz, most probably, but you won't see it in a Powermac in January.
Will it have that kind of IPC that makes it competitive with a Power4?
Highly Unlikely. it takes more than just strapping on extra execution units for that kind of performance. You need better BP, more reorder resources, Probably a recompile on existing code to maximize new capabilities, basically large improvements accross the board in everything to get that kind of IPC boost while also virtually doubling clock-speed.
And I don't recall Motorola having a history of being able to bring something like this to table in this manner.
<strong>Those mythical SPEC scores for the G5 are BS. If you do the math, it's scaling superlinearly, which is totally impossible.
Just to get perfectly linear scaling, you would need to make sure that absolutely every scrap of code is housed in cache; for which you'd need a few hundred megs of it. And even then you wouldn't get linear scaling, but you'd get closer to it.
Superlinear scaling, however, could not happen under any situation in a given rigidly defined set of circumstances -- i.e. same motherboard, same hardware, same programs, same CPU revision; same everything except clockspeed.
Edit: Furthermore, you would not do a SPEC run on anything but a final production chip and motherboard, because otherwise it would potentially be unstable and ruin your run. Also, the SPEC source code isn't free.
Altogether, it's BS. My guess is the Register's "source" pulled those numbers directly from his ass without thinking too clearly about them.
Comments
- Pook
<strong>Cheers, FormerLurker. That final paragraph on the importance of public perception was right on. I've been arguing for over a year that Steve's explaining away MHz helps nothing. As you said, it's preching to the choir. We need to make Macs appear fast, and be able to back up that claim. Then we'll be competitive.
- Pook</strong><hr></blockquote>
[QB][/QB]<hr></blockquote>
Sorry, shakedown flub (what a way to make a bow...)
I've never been too terribly concerned with the PPC/Pentium comparisons, but when there's NO relative chip speed progress in and of itself, that's painful.
<strong>The G5 is the only way to not be ridiculous in the mhz war, by lenghtening the pipeline bus (14 stages in many rumors, unstead seven for the 7450), it's their only way to go above one ghz (still stuck at 867 mhz)
probabily in many aspect excepting video games (because of the bandwitch) the G5 will be slower at equal Mhz than the G4 (like the P4 versus the PIII).
Even with the G5 the mhz war is loose by Motorola the only goal is to not be too ridiculous and to find some benchmarks who will show the superiority of their chips (even if in most of the tasks the chip is slower ...)</strong><hr></blockquote>
1) The G5 will kick the shit of the G4.....I will personally guarantee that. And there will be no G4 at Equivalent MHZ.
2) The pipleline will probably not be 14 stages....
3) G4 chips will break one gig without increasing the pipeline.
Wait and see....
Someone please correct me if I'm wrong....
<hr></blockquote>
No, you are right. I heard the same thing, many times, from different people.
<strong>1) The G5 will kick the shit of the G4.....I will personally guarantee that. And there will be no G4 at Equivalent MHZ.
2) The pipleline will probably not be 14 stages....
3) G4 chips will break one gig without increasing the pipeline.
Wait and see....</strong><hr></blockquote>
In Valen's name, I pray that you are right!
<strong>
1) The G5 will kick the shit of the G4.....I will personally guarantee that. And there will be no G4 at Equivalent MHZ.
2) The pipleline will probably not be 14 stages....
3) G4 chips will break one gig without increasing the pipeline.
Wait and see....</strong><hr></blockquote>
1) I doubt that the G5 will beat the G4 at equal Mhz especially the 7410. Of course the G5 wil start clocking higher than the G4 (same as the 7450 starting at 733 vs the 7410 stuck at 533, but remember excepting altivec stuff the benchmarks of this too later chips where very similar...)
2) Have you got any info about the number of stage of the pipeline ?
3) well you are right undead the G4 will break the GHz wall, but is not able to go far beyond.
<strong>
1) I doubt that the G5 will beat the G4 at equal Mhz especially the 7410. Of course the G5 wil start clocking higher than the G4 (same as the 7450 starting at 733 vs the 7410 stuck at 533, but remember excepting altivec stuff the benchmarks of this too later chips where very similar...)
2) Have you got any info about the number of stage of the pipeline ?
3) well you are right undead the G4 will break the GHz wall, but is not able to go far beyond.</strong><hr></blockquote>
<a href="http://www.theregister.co.uk/"; target="_blank">http://www.theregister.co.uk/</a>;
"Getting to those clock speeds [1.6 GHz] involved increasing the G5's pipeline from the 7450's seven stages to ten.
Initial benchmarks supplied by our source suggest the G5 is pretty fast indeed:
GHz 1.2 1.4 1.6
SpecInt2000 987 1151 1340
SpecFP2000 1005 1173 1359
By comparison, Intel's 2GHz Pentium 4 has recorded SpecInt2000 and SpecFP2000 scores of 656 and 714, respectively, according to <a href="http://www.specbench.org"; target="_blank">www.specbench.org</a> If accurate, the G5 figures are impressive indeed. "
This compares to
POWER4 (1 core, 1.3 GHz) 814 int, 1169 fp
Athlon MP1800+(1,53 GHz) 607 int, 547 fp
If the numbers for G5 are roughly right, it runs circles around any G4!
I expect them to blow everyone away at MWSF.
The current Quicksilver' s are obviously buying them 6 months more time. More time that will be completely worth it. Those models are crippled versions of what I earlier have worked with. More on specifics later.....
dorsal
<strong>More on specifics later.....</strong><hr></blockquote>
Later, later, later.... always later. <img src="graemlins/hmmm.gif" border="0" alt="[Hmmm]" />
Escher
[ 11-15-2001: Message edited by: Escher ]</p>
As I recall, there has also been a case redesign in the works.
<strong>[...] As I recall, there has also been a case redesign in the works.</strong><hr></blockquote>
Yeah, I the thread. He wrote about a case looking like the old Performa 6400... or was it the case like the MacTrashCan on following site? <a href="http://homepage.mac.com/benvp/PhotoAlbum.html"; target="_blank">http://homepage.mac.com/benvp/PhotoAlbum.html</A>;
Anyway, nice to see Dorsal back on AI.
[ 11-15-2001: Message edited by: M'Ashan ]</p>
Initial benchmarks supplied by our source suggest the G5 is pretty fast indeed:
GHz 1.2 1.4 1.6
SpecInt2000 987 1151 1340
SpecFP2000 1005 1173 1359
*snip*
POWER4 (1 core, 1.3 GHz) 814 int, 1169 fp
Athlon MP1800+(1,53 GHz) 607 int, 547 fp<hr></blockquote>
Hmm. I don't know for sure, obviously, but I seriously doubt that a desktop G5 is going to come anywhere near the performance of IBM's super-hot, super HUGE ultimate floating point optimized Power4. This point alone tends to cast doubt on the veracity of the Register's claims. If true, the 1.6GHz G5 would match the performance of the Athlon 4400+ nearly a year before the Athlon's introduction. I tend to doubt this, though...
Still, I suppose that almost anything is possible.
Just to get perfectly linear scaling, you would need to make sure that absolutely every scrap of code is housed in cache; for which you'd need a few hundred megs of it. And even then you wouldn't get linear scaling, but you'd get closer to it.
Superlinear scaling, however, could not happen under any situation in a given rigidly defined set of circumstances -- i.e. same motherboard, same hardware, same programs, same CPU revision; same everything except clockspeed.
Edit: Furthermore, you would not do a SPEC run on anything but a final production chip and motherboard, because otherwise it would potentially be unstable and ruin your run. Also, the SPEC source code isn't free.
Altogether, it's BS. My guess is the Register's "source" pulled those numbers directly from his ass without thinking too clearly about them.
[ 11-15-2001: Message edited by: TheAlmightyBabaramm ]</p>
[Those mythical SPEC scores for the G5 are BS. If you do the math, it's scaling superlinearly, which is totally impossible.<hr></blockquote>
The linear variance is < 2% which is realistic as run to run fluctuations. It is also possible and maybe even likely that these are chips from different batches with different revisions and internal tweaks.
[quote]Just to get perfectly linear scaling, you would need to make sure that absolutely every scrap of code is housed in cache; for which you'd need a few hundred megs of it. And even then you wouldn't get linear scaling, but you'd get closer to it.<hr></blockquote>
Not necessarily true at all. It depends heavily on the nature of the data that is being used for the calculations as well as the compiler used to create the code. If the data displays good temporal and spacial locality and the compiler uses pre-fetch instructions there is no reason for the processor to ever have to wait for a memory fetch. Of course we cannot know if this is the case or not, but if the benchmark number ratios are accurate I wouldn't be surprised.
Assuming that Motorola did in fact run these tests it is possible that they didn't run the entire suite and are basing the numbers on just a few of the tests. In any case I'm sure Motorola has put them into motherboards at some point at least, especially if they are shipping engineering samples to Apple for testing.
<strong>I do not know what their plans are with some of the earlier features I spoke of earlier in the year, but they must have made some big changes to their schedule.
I expect them to blow everyone away at MWSF.
The current Quicksilver' s are obviously buying them 6 months more time. More time that will be completely worth it. Those models are crippled versions of what I earlier have worked with. More on specifics later.....
dorsal</strong><hr></blockquote>
It's been a while since Apple "blew" anyone away with features. A few I can think of were AirPort (with the iBook), Firewire (with the B&W G3). Maybe the dual-processors, too. Design, Apple's always done well with that (love the TiBook). Features-- I won't be holding my breath.
So why did you pick "dorsal?"
<strong>
1) I doubt that the G5 will beat the G4 at equal Mhz especially the 7410. Of course the G5 wil start clocking higher than the G4 (same as the 7450 starting at 733 vs the 7410 stuck at 533, but remember excepting altivec stuff the benchmarks of this too later chips where very similar...)
2) Have you got any info about the number of stage of the pipeline ?
3) well you are right undead the G4 will break the GHz wall, but is not able to go far beyond.</strong>[/QUOTE]
Powerdoc, how can you really comment on the relative speed of the G5 as it relates to the G4. By coincidence, you may be correct. However, without seeing the white paper on the G5 design, all speed predictions are completely guesses.
All things being equal, yes, a longer pipeline would mean slower mhz per mhz. However, if additional processing units are added to the chip, such as extra INT of FP units, then these very well offset the extra pipeline length. We don't know what affect Rapid IO will have, etc.
I don't really believe those published SPEC results for the G5, but, for the sake of argument, if they are even close to being true, then the G5 will be MUCH faster than the G4 mhz for mhz.
Also, regarding the 7410 at 533mhz and the 7450 at 733mhz... The 7450 scales proportionally against the 7410 in Altivec and Int operations. It just doesn't scale as well at FP. Of course, that's because additional INT and Altivec units were added, but FP was left the same but with a longer instruction pipeline.
Steve
How can anyone really comment on it?
It's just a bunch of rumors. Anyone with any licence to say anything is most likely constrained by NDA's. Anyone else is just speculating or passing on third hand information.
All in all, If I had to make my own educated guess on who's predictions are right, I'd say Powerdoc's more on the right track.
Will the G5 be faster than the G4, yes.
1.2-1.6 Gighz, most probably, but you won't see it in a Powermac in January.
Will it have that kind of IPC that makes it competitive with a Power4?
Highly Unlikely. it takes more than just strapping on extra execution units for that kind of performance. You need better BP, more reorder resources, Probably a recompile on existing code to maximize new capabilities, basically large improvements accross the board in everything to get that kind of IPC boost while also virtually doubling clock-speed.
And I don't recall Motorola having a history of being able to bring something like this to table in this manner.
<strong>Those mythical SPEC scores for the G5 are BS. If you do the math, it's scaling superlinearly, which is totally impossible.
Just to get perfectly linear scaling, you would need to make sure that absolutely every scrap of code is housed in cache; for which you'd need a few hundred megs of it. And even then you wouldn't get linear scaling, but you'd get closer to it.
Superlinear scaling, however, could not happen under any situation in a given rigidly defined set of circumstances -- i.e. same motherboard, same hardware, same programs, same CPU revision; same everything except clockspeed.
Edit: Furthermore, you would not do a SPEC run on anything but a final production chip and motherboard, because otherwise it would potentially be unstable and ruin your run. Also, the SPEC source code isn't free.
Altogether, it's BS. My guess is the Register's "source" pulled those numbers directly from his ass without thinking too clearly about them.
[ 11-15-2001: Message edited by: TheAlmightyBabaramm ]</strong><hr></blockquote>
wouldn't they be possible if the G5 were multicore with 2 cores running at the said clockspeeds?
or is the G5 no longer planned for multicore?