AMD Athlon64 Clock Speed Wall

[VIAN]

AMD doesn't seem to like to push their Athlon over 2.2 and their new A64 doesn't go past 2.2 either. It feels like adding the memory controller was just another trick, like the addition of cash and FSB, to gain some performance out of the Athlon. Watch next upgrade have 1MB L2 cache 533 memory support and 2.0GHz for A64 3400+ or something. Any thoughts?

 

[Elcs]

To be perfectly honest, AMD could stick a pair of fluffy dice in their next processor.... as long as they produce a chip that gives the best performance at the price point i want, then I'll buy their chips.

 

[Shinei]

So what if they don't ramp their clocks as high? Are they still squeezing the most performance possible out of their CPUs? I'd rather see a more efficient CPU than one with a bajillion gigahertz of clock speed...

Edit: I'm fairly certain the reason the clocks aren't going very high is because K8 is basically a modified K7, which ALSO apparently has a 2.2GHz clock ceiling...

 

[VIAN]

I understand and agree with you, but it just makes me think about the future of AMD. They haven't announced anything new and if that is the wall, then what happens next? Maybe they're doing something good by upping the memory speed support since there system will just get significantly faster by how much memory it supports since the FSB is the CPU speed. Maybe they are smart after all.

I wonder why noone ever ups the L1 cache.

 

[Jeff7181]

It has more to do with the limitations of the manufacturing process. The .09 micron SOI Athlon-64's should scale much higher than 2.2 Ghz, as stated in AMD's roadmap.

 

[zephyrprime]

The A64 has 20% more pipeline stages that the Athlon which by itself should let it scale higher than the Athlon.

 

[Jeff7181]

Originally posted by: zephyrprime
The A64 has 20% more pipeline stages that the Athlon which by itself should let it scale higher than the Athlon.

Also true... and by the way... who said the Athlon64/Opteron is limited to 2.2 Ghz? Is this being assumed because 2.2 Ghz is the highest speed processor available right now? If you honestly believe that the first processor in the family to be released is as high as the architecture is capable of scaling, you need to do a little research.

 

[VIAN]

Well, its basically an athlon with extra features and the athlon tops seems to top off at 2.2 since AMD has never released on past that speed.

 

[klah]

According to most speculation I've seen 2.4GHz (FX-53 & 64-3700+) will be the last speed bump for 130nm.

 

[Jeff7181]

Originally posted by: VIAN
Well, its basically an athlon with extra features and the athlon tops seems to top off at 2.2 since AMD has never released on past that speed.

You're forgetting one major difference... the Athlon-64 is manufactured using SOI, which is supposed to extend the scalability. There's also the extra stages in the pipeline as was mentioned.

 

[JeremiahTheGreat]

people ever fancy the possibility that all the chips we're buying are still the stockpiled chips? Old steppings = works at the required speed but not much more ...

 

[VIAN]

didnt know about the SOI thing

 

[myocardia]

Originally posted by: JeremiahTheGreat
people ever fancy the possibility that all the chips we're buying are still the stockpiled chips? Old steppings = works at the required speed but not much more ...

Not very likely, assuming it's true that the chip isn't much more than a modified XP. Of course, now that I think about it, that doesn't seem very likely at all, given all of the major design differences...

 

[alexruiz]

2.6 GHz...... mark my words (Yes, as someone said, SOI is the key...)

 

[mamisano]

Just an FYI, AMD just released a special XP chip to HP that runs at 2.33Ghz. So, I doubt 2.2 is the wall.

 

[Jeff7181]

Originally posted by: VIAN
didnt know about the SOI thing

Here's a pretty good explanation of what SOI is and what it's benefits are.

Silicon on Insulator

*EDIT* For you Intel guys, here's the best explanation of Strained Silicon I can find... it's not very technical, but it gets the point across.

Strained Silicon

 

[KF]

>I'm fairly certain the reason the clocks aren't going very high is because K8 is basically
> a modified K7, which ALSO apparently has a 2.2GHz clock ceiling..

It is time to become uncertain.

There is nothing much in common with the K7 die, because there can't be. SOI is a completely different way of building chips that uses different chip processing. A complete redesign would be necessary to manufacture even an identical logic circuit (whch the K8 isn't). The reason for using SOI is to reduce substrate capacitance, which in turn enables an otherwise identical circuit to operate at much higher speed. SOI processes are much harder to execute, which drives up the cost. SOI has not been used at the level of complexity together with the small design rules of the K8 chip before. Therefore working out all the bugs in the processes to get chips as perfect as they need to be to operate at maximum speed is going be difficult. There is no garantee it can be done successfully. The same can be said for any process pushed to its limits. AMD is spending huge sums of money on new sorts of equipment to get this going. Whatever clock speed they are shipping is the result of how close to ideal they have been able to get. Because of the nature of SOI, if the design were identical to the k7, the clock speed would go much higher, provided the quality of processing required can be achieved. In short , if the K7 maxes at 2200 GHz, then the K8 with SOI maxes at something much higher. Let me repeat: the ONLY reason for going the SOI route is to get much higher speeds. It is otherwise a total waste of money.

The K8 is an elaboration of the K7 in conception. In actual implementation, the CPU resources are designed to go to higher speed. Since the underlying process it is carried out in are inherently capable of higher speed, the speed of the K8 will go much higher provided AMD can execute the process engineering. I know people think AMD should try to "stomp" Intel, and if AMD doesn't, they are failing. But realistically the more margin of performance AMD strives for, the more it costs them, so the only economic thing to do is to slightly out-perform (if possible). Intel hasn't been sitting still, so the fact that AMD is achieving parity even with the K8 delays is excuting pretty well.

 

[BD231]

KF-

So in a sense this new Athlon64 is like going from a k6 to the K7? Slot-A Athlon, T-Bird, Pally, T-Bred, Barton. Are we going to see a progression with core modifications as we did the K7?

 

[Jeff7181]

Originally posted by: KF
>I'm fairly certain the reason the clocks aren't going very high is because K8 is basically
> a modified K7, which ALSO apparently has a 2.2GHz clock ceiling..

It is time to become uncertain.

There is nothing much in common with the K7 die, because there can't be. SOI is a completely different way of building chips that uses different chip processing. A complete redesign would be necessary to manufacture even an identical logic circuit (whch the K8 isn't). The reason for using SOI is to reduce substrate capacitance, which in turn enables an otherwise identical circuit to operate at much higher speed. SOI processes are much harder to execute, which drives up the cost. SOI has not been used at the level of complexity together with the small design rules of the K8 chip before. Therefore working out all the bugs in the processes to get chips as perfect as they need to be to operate at maximum speed is going be difficult. There is no garantee it can be done successfully. The same can be said for any process pushed to its limits. AMD is spending huge sums of money on new sorts of equipment to get this going. Whatever clock speed they are shipping is the result of how close to ideal they have been able to get. Because of the nature of SOI, if the design were identical to the k7, the clock speed would go much higher, provided the quality of processing required can be achieved. In short , if the K7 maxes at 2200 GHz, then the K8 with SOI maxes at something much higher. Let me repeat: the ONLY reason for going the SOI route is to get much higher speeds. It is otherwise a total waste of money.

The K8 is an elaboration of the K7 in conception. In actual implementation, the CPU resources are designed to go to higher speed. Since the underlying process it is carried out in are inherently capable of higher speed, the speed of the K8 will go much higher provided AMD can execute the process engineering. I know people think AMD should try to "stomp" Intel, and if AMD doesn't, they are failing. But realistically the more margin of performance AMD strives for, the more it costs them, so the only economic thing to do is to slightly out-perform (if possible). Intel hasn't been sitting still, so the fact that AMD is achieving parity even with the K8 delays is excuting pretty well.

Cliff Notes: SOI is new, give the process time to mature and you'll see the "predicted" speeds.

 

[Jeff7181]

Originally posted by: BD231
KF-

So in a sense this new Athlon64 is like going from a k6 to the K7? Slot-A Athlon, T-Bird, Pally, T-Bred, Barton. Are we going to see a progression with core modifications as we did the K7?

Most definately... the first core modification being the FX switching to the 939 pin layout. Then the .09 micron process. There will most likely be many more modifications, but those are two known to be in the making right now.

 

[KF]

>So in a sense this new Athlon64 is like going from a k6 to the K7? Slot-A Athlon, T-Bird, Pally, T-Bred, Barton.
So it seems to me. The K7 was an elaboration of the K6. A massive elaboration. With the K7, AMD decided to do a chip that was a big enough leap that Intel couldn't just up the timetable on their next generation 6 months and quickly leave AMD again as a second tier chip maker. I think people have said that AMD designs along these lines began with one of the K5's. There were at least 2 K5s that were totally different except for the name. The one that was designed by Nexgen, after AMD bought them, is the one that started this line.

People forget that Nexgen, before Intel, used a RISC processor to execute x86 instructions translated on the fly, in hardware, to behind-the-scenes RISC instructions, although it was not 100% x86 compatible. AMDs 100% x86 compatible K6 continued this. The K6 was designed to compete with Intel's Pentium, which was not RISC based. Intel didn't take long to get out the PII line, starting with the astronomically priced, oddball Pentium Pro and some stupendously expensive PII cartridge kludges, eventually progressing to a single chip socketed design. The PII dropped the K6 to second tier status a little too quickly. The K7 was a big enough stretch to keep AMD neck-and-neck with Intel for a long time. It was unnecessary for AMD to switch course like Intel did from the Pentium to the RISC based PII. The very high price of the PIIs worked out to AMDs advantage. It meant that AMD could make money even as expensive a stretch the K7 was for AMD. I'm sure AMD is looking forward to Intels K8 killer when it appears. By then the K8 process bugs should be worked out and AMD will be able to charge similar phenominal prices for the top speed K8s that will parallel Intel's killer.

>Are we going to see a progression with core modifications as we did the K7?

That's the way it alway is, isn't it? There were some pretty significant K7 performance redesigns at the Thunderbird and XP points. It hasn't been all die-shrinks to get the K7 up to 2000MHz. I forget where the first K7, slot A's started. 500MHz? It would be reasonable to expect the K8 engineers to have designed it for a similar run-up. 8000MHz? I don't think AMD will need a process shrink to get the K8 way above 2200MHz. Somewhere along the way to 8000MHz, IBM and AMD have something in the works, judging by the annoucements. It is interesting that both strained-silicon, now being exploited by Intel, and SOI, now being exploited by AMD, are both developments that came out of IBM. Actually, it seems like all developments somehow lead back to IBM.

One thing I found about from reading some Intel pdfs, is that the first silicon always has a lot of errors. Before they think about silicon, they create virtual chips as software that runs on vast banks of computers. (I'd like to see Intel's computer room. The one at Nvidia is quite impressive.) Unfortunately the software versions run too slowly to test thoroughly. It may take weeks or months to get to a DOS prompt. Only when it goes to silicon do they get to check it completely. They just revise as small a section as possible for the first retail product, provided it can be done. Some revisions are done with micro-code, which is a sort of software that can be loaded onto the chip. (In Windows XP Device Manager, I see there is a microcode driver.) I have no idea how that could fix anything. Redoing the thing without the patches costs too much and takes too long. Therefore the second production version, when they get around to it, is always the good one, the one the engineers really wanted to make.

 

[mrgoblin]

Originally posted by: VIAN
I understand and agree with you, but it just makes me think about the future of AMD. They haven't announced anything new and if that is the wall, then what happens next? Maybe they're doing something good by upping the memory speed support since there system will just get significantly faster by how much memory it supports since the FSB is the CPU speed. Maybe they are smart after all.

I wonder why noone ever ups the L1 cache.

Amd never announces anything new. When its coming you hear about it. Theyre not very good buisness people. Amd could hit high ghz just not on .13 process. Theyll go to 2.4 when prescott comes out and youll start seeing 2.6 fxes by the 90 nano rollout in the summer of next year.

 

[aka1nas]

Originally posted by: VIAN
Well, its basically an athlon with extra features and the athlon tops seems to top off at 2.2 since AMD has never released on past that speed.

AMD hasn't released one past that speed because then their "old" processor would have more clockspeed and thus more "speed" than their new one. Intel did the same thing with the pentium 3 when the P4 came out, turning the nicely performing tualatin core into the Celeron for a while. A 1.2Ghz tualatin celeron would trounce the original Willamette P4 core@anything less than 1.8Ghz. The tualatin also overclocked very well as Intel wouldn't increase it's official clockspeed.

This actually happens almost every time Intel or AMD comes out with a radically different processor architecture. The original Pentiums sucked compared to the 486s of the day. The fact that the K8 performs on par and sometimes better than the k7 is mainly because it was delayed so long and AMD had time to work out bugs and improve yields.

 

[BD231]

One thing I found about from reading some Intel pdfs, is that the first silicon always has a lot of errors.

This makes a lot of sense. I've been watching processor progression over the past few years and if you take a real good look at the starting point of the P4 and Athlon, they pale in comparison to what they've become. Slot-A Athlon with that half speed off die cache, P4 Wilmette that didn't even out perform the P3.

If the Athlon 64 is truly in such an infancy I think it's safe to say the A64 is going to soar. On a side note this means it's a horrible time purchase one of these processors though.

 

[ToxicWaste]

Just like with Microsoft, you wanna wait for the first service pack to come out before you use one of their OS's.