AMD cant fab at 65nm but can at 55?

[dmens]

However in certain cases (such as this one, where the chip was rushed to market), the errata fix can represent a very large boost in both performance and scaling.

doesn't happen. thoroughbred b was not a "b-stepping" of anything, it added an entire new metal layer. calling that scope of change a stepping would be ridiculous. major performance deviations would have been caught on initial silicon and fixed. hence, once productized, there won't be any major performance increases.

also, thoroughbred b ran faster but was not any cooler than thoroughbred a.

 

[taltamir]

but that is his point isn't it? that stepping is an arbitrary name they decide to give... basically for modifying the processor in a way that is not reflected in advertising, only in the detailed specs that few people know to look for.
The nvidia 8800GT is a perfect example of this... and the upcomming version3 8800GTS also.

AMD might be releasing a new "stepping" for the phenom that is as significant as thoroughbred b and still calling it phenom.

 

[dmens]

Originally posted by: taltamir
but that is his point isn't it? that stepping is an arbitrary name they decide to give... basically for modifying the processor in a way that is not reflected in advertising, only in the detailed specs that few people know to look for.
The nvidia 8800GT is a perfect example of this... and the upcomming version3 8800GTS also.

AMD might be releasing a new "stepping" for the phenom that is as significant as thoroughbred b and still calling it phenom.

traditionally, steppings are small engineering modifications on an existing design to improve the various aspects of the product. a company can decide to call a major revision a "stepping", or whatever else they want.

an overhaul of a design, call it a stepping or redesign, is still bound by engineering reality and schedule. that is why i find the idea of a "magic stepping" ludicrous. it implies a quick fix yielding huge gains.

 

[Viditor]

Originally posted by: dmens

However in certain cases (such as this one, where the chip was rushed to market), the errata fix can represent a very large boost in both performance and scaling.

doesn't happen. thoroughbred b was not a "b-stepping" of anything, it added an entire new metal layer. calling that scope of change a stepping would be ridiculous. major performance deviations would have been caught on initial silicon and fixed. hence, once productized, there won't be any major performance increases.

also, thoroughbred b ran faster but was not any cooler than thoroughbred a.

You can call it whatever you want...major redesign, marketing, revision, stepping, whatever...

The only important points are:

1. The "A" Thouroughbred was released on 10 June 2002, the "B" T-Bred was released on August 21, 2002, that's just over 10 weeks for a major enhancement.

2. The most important factor isn't how much change is made on the design to fix an errata, the only important point is how much that errata effects the performance!

This isn't a "magic stepping" (whatever that is), it's a very nasty problem in the design that's being fixed. That's not magic, it's engineering...

 

[dmens]

1. so? A was a shrink, B was a redesign. they were done concurrently. B is not an enhancement, it was far more than that. are you comparing this supposed TLB fix (if it is even real) to that level of change? i hope not.

2. tell me how many errata fixes yield major performance gains... oh yeah, none.

if the problem were so nasty, the product would not have been released. end of story. you keep on claiming that B-3 or whatever will yield big performance gains, hence the "magic stepping". how gullible you are... too bad engineering reality is in the way.

 

[Viditor]

Originally posted by: dmens
if the problem were so nasty, the product would not have been released.

Talk about gullibility...do you really think they would hold the release on ALL chips just prior to Xmas because of the TLB errata?

We know what the problems are, and what the effect is. It's been talked about ad nauseum here...
Just like with T-bred, it's perfectly acceptable to go ahead with the already delayed release prior to the fix, just not at competitive clockspeeds.
Also, just like with T-bred, a fix has been found and will be released within a few months of launch.

What's "magic" about that?

This is very much common practise...Intel released Willamette before Netburst was really ready as well. To quote from Wikipedia:

"In November 2000, Intel released the Willamette Pentium 4 at speeds of 1.4 and 1.5 GHz. Most industry experts regarded the initial release as a stopgap product, introduced before it was truly ready."

 

[dmens]

look in the mirror... you've lapped up every PR soundbite from AMD ever since i read this forum. that is the definition of gullibility.

"we know what the problems are"? you're the main pusher of the TLB rumor on the forum, and you're also the main pusher of the B3=big gains rumor. neither of which have a shred of evidence. i think you have no idea what the problem is (just like everyone else), and you simply want to see an improvement because of your financial interests.

funny statement about with the t-bred. an extra metal layer isn't exactly a "fix". but hey, the phenom is in the same situation and a fix will be released in a few months, yessiree, absolutely.

nice strawman with the willamette, bringing up an issue that i never even contested. but now that you mention it, you really ought to look at its frequency progression. that is engineering reality right there.

 

[AlabamaCajun]

One year ago Brisbanes were not that great, now 2.7G versions are out meaning that 65nm does work. The big question now is will Phenom do this in under 6 months since the window is tighter with the current market. 2.2-2.4 sans-errata is not bad for an introduction except that the bar is at 3.0G that AMD set. This is what the Spider show was about, just to show it is possible. The hardware will run at 3G, just at this time it's a cherry pick and an overclock (not good for business or reviewers). A 5000 Brisbane has been on the web doing 4G with CO2 but it will run the frequency. I think in a year we may see a 3.2G phenom which should be eq to a 3.6G Athlon with the same number of core in action.

 

[CTho9305]

I don't know what the TLB issue is or anything about "B3". All I know, I read on the Inquirer.

Originally posted by: Viditor
Intel uses a method called "Copy Everything".

I thought it was "Copy Exact".

AMD uses a more automated process called APM. {snip} It can adjust the doping on an individual transistor of a single die on a specific wafer.

{{citation needed}} I'd believe all transistors on a given mask, but not individual transistors.

Originally posted by: taltamir
also, what is doping?

Transistors are made by taking very pure silicon and adding other atoms ("dopants") to "dope" the silicon. Some types of dopant atoms result in a net excess of electrons around them, and some result in a net shortage. By putting doped regions of different types next to each other, you can create diodes and transistors. How strongly you dope the silicon in the various parts of a transistor affects how fast it is and how leaky it is. If you discover a critical path on a chip (the slowest path on a chip - the path that limits the clock speed), you can dope the transistors on it differently so they're faster (at the expense of increased leakage power). The leakage difference is extreme; the fastest transistors have to be used very sparingly because the leakage difference is orders of magnitude (but the speed difference is probably less than a factor of 2).

Originally posted by: dmens

However in certain cases (such as this one, where the chip was rushed to market), the errata fix can represent a very large boost in both performance and scaling.

doesn't happen.

Oh really? FWIW, speed bin increases often come from new revisions that generally make small changes with large effects.

major performance deviations would have been caught on initial silicon and fixed. hence, once productized, there won't be any major performance increases.

What if the fix is an all-layer spin and you don't want to wait 3 more months to start shipping? It would be easy to have Fmax limiters that could be fixed with a relatively simple revision. Sure, once you've been through a few revs of a chip you'll have a good idea of where the real silicon critical paths are, but for a brand new design, you might just be unlucky and run into a case where the timing analysis didn't catch the worst path (spice models aren't perfect).

Originally posted by: dmens
an overhaul of a design, call it a stepping or redesign, is still bound by engineering reality and schedule. that is why i find the idea of a "magic stepping" ludicrous. it implies a quick fix yielding huge gains.

What if there's a single critical path that stands out far beyond all others, and you fix it? There are other issues that could put similar artificial limits on Fmax - I've heard stories of old Alpha chips which had an Fmin caused by leakage on dynamic nodes - for the fastest process corners, the Fmin was above the Fmax and the chips wouldn't work. If there were only a few problem nodes, simple changes could probably fix the Fmin issue, making the very fastest parts sellable again.

I don't know what the TLB issue is or anything about "B3". All I know, I read on the Inquirer. I'm not claiming there's a "magic stepping" coming - only that such a thing doesn't sound impossible. I'm not speaking for any company.

 

[taltamir]

Originally posted by: Viditor

Originally posted by: dmens
if the problem were so nasty, the product would not have been released.

Talk about gullibility...do you really think they would hold the release on ALL chips just prior to Xmas because of the TLB errata?

We know what the problems are, and what the effect is. It's been talked about ad nauseum here...
Just like with T-bred, it's perfectly acceptable to go ahead with the already delayed release prior to the fix, just not at competitive clockspeeds.
Also, just like with T-bred, a fix has been found and will be released within a few months of launch.

What's "magic" about that?

This is very much common practise...Intel released Willamette before Netburst was really ready as well. To quote from Wikipedia:

"In November 2000, Intel released the Willamette Pentium 4 at speeds of 1.4 and 1.5 GHz. Most industry experts regarded the initial release as a stopgap product, introduced before it was truly ready."

This reminds me a real conversation I happened to hear on a product that was released by a company I worked.

Employee: But we can't release it right now, we have a 30% failure rate and about 60% with one minor issue or another
Boss: Look, we can go on, lock ourselves with RnR for 3 years, and have a perfect product. And by then it will probably obsolete and we will be bankrupt. Or we can sell what we have now, and repair or replace failing machines with more reliable parts as we develop them...

And so it was.


Oh yea, and the cell processors on PS3 only uses 7 of 8 cells due to defects. They all come with one processor disabled, either a defective one or a good one. They get 10-20% yields when they count a 7/8 processor as good.
There is a very small chance you will get one with 8/8 processors, in which the last processor will become active if one of the other ones will fail (which has been known to happen according to sony).
Why did they sell it? because they couldn't afford to wait a year and a half until they got it really right...

 

[Phynaz]

even moving the wafers around is done completely via an automated sealed carrier system.

Every fab on the planet does this. People don't actually touch wafers you know...

In any modern chip fabrication facility people cannot go where the carriers go, it's physically not accessable.

 

[taltamir]

you know what. I actually go back to what I said about the 65nm process being problematic...

Yes, it has better yields then the 90nm process... but yields aren't the absolute measure of a process effectiveness... It has less defects that cause a die to be useless, but it has more electron leakage causing lower speeds... Now ofcourse that could be due to the processor design itself... but what about when the thoroughbred B added an enitre new layer for a revision? that is a change in manufacturing process. The manufacturing process and what is manufactured go hand in hand. Otherwise you could just make phenoms on the working 55nm process used for video cards.

 

[AlabamaCajun]

Die shrinks require a change in geometry of the circuits a change in masking and the materials used. It won't help AMD to jump to 55nm when work on 45nm has already begun. It is quite complex and expensive to setup. This is why once you start up a run, you want to get as much through as possible before obsoletion of the product. In this case Phenom arch at 65 comes within 3 to 6 months of obsoletion on release say. Now 6 months later they need to have something on 65nm and it should be the Brisbane Athlon X2 replacement then Opterons followed by Phenom. We only hope now that when 45nm pulls out, it's out quick and it works well. It will be at least 2009 when we see more improvements unless work is already underway to do what Penryn did and bump out a fast high doller chip with a real fast and large set of caches. Improved int process would be nice also. Brisbanes are already great, now lets see how much better the Phenoms are.

 

[Extelleron]

While I agree that while Phenom is not an ideal chip right now with the current stepping/release, I really think that it will get much better with future releases as mass production begins. A clear example of this is the Brisbane 65nm processors, which at launch were horrible in terms of clockspeed and didn't overclock anywhere near as well as the 90nm A64's did. Now, however, G2 Brisbanes are capable of hitting well over 3GHz consistantly and they have been great overclockers for a while. I bought an X2 3600+ back in May for $60~ and have it running 24/7 @ 2.8 GHz, no problems. I'm sure Phenom will be the same way; right now, the clockspeeds suck, but they will get much better as time goes on.

Also, I think that people are overplaying how large of a chip Phenom is. At 285mm^2, it is massive for a CPU, but remember that the 90nm AM2 chips are 230mm^2 in size, and AMD has been selling them for as low as $120 (X2 5200+). Phenom is a bit bigger, and the 65nm process is not as mature, but according to AMD yields are good and I don't think this will be as much of a problem as people think.

 

[coldpower27]

Originally posted by: taltamir
AMD has serious problems with the 65nm process. It is the first time ever that a new, smaller process cannot be used to make faster processors then the older process, instead making slower ones. This is why AMD is in so much trouble. They should be subsidizing the costs of the new upgrade by selling the parts at 300-1000$ per cpu. Which would slowly decrease... so that by the time the second plant is upgraded to 65nm they would be selling in the 150+ range with 100 and less bering reserved for 90nm oldest parts (in about 10 monthes from today). The problem is that the 65nm parts are SLOWER then the 90nm parts...

Not really remember AMD's 90nm process which had only up to 2.2GHZ at first with the Winchester core vs Newcastle which went to 2.4 GHz it wasn't till overtime on the second revision that they got higher speed bins out of the 90nm process.

This is the first time it is taking so long to ramp up 65nm process speed/performance in Single threaded scenarios is what is so alarming, even after so much time speed wise they are still around the same level.

We have a max of what 2.7 GHz on the 5200+ Brisbane core, when the introduction speed was like 2.6 GHz from the start back in December 2006. Even now we only have a mere 2.3 GHz with Agena hich given it's IPC improvements can be compared at about the same level as the 2.7GHz Brisbane.

Even though AMD did a test processor with Brisbane it is still a shame they were only able to get 2.3GHz with the Agena.

 

[coldpower27]

Originally posted by: AlabamaCajun
One year ago Brisbanes were not that great, now 2.7G versions are out meaning that 65nm does work. The big question now is will Phenom do this in under 6 months since the window is tighter with the current market. 2.2-2.4 sans-errata is not bad for an introduction except that the bar is at 3.0G that AMD set. This is what the Spider show was about, just to show it is possible. The hardware will run at 3G, just at this time it's a cherry pick and an overclock (not good for business or reviewers). A 5000 Brisbane has been on the web doing 4G with CO2 but it will run the frequency. I think in a year we may see a 3.2G phenom which should be eq to a 3.6G Athlon with the same number of core in action.

2.7 GHz is the best AMD can do now? It has been nearly a year since the 65nm process was introduced and your only looking at a 100MHz uptick in all that time?

In terms of Single Threaded performance were still stuck with 90nm products at 3.2GHz, but once your throw in Multi Threadedness a 2.3 GHz Agena doesn't appear too bad.

Yes just like how Core 2 Duo's can be overclocked to 3.5 GHz+, but that doesn't mean they will be released at that frequency. For AMD's sake I do hope that one day that frequency does arrive.

 

[Viditor]

Originally posted by: taltamir
you know what. I actually go back to what I said about the 65nm process being problematic...

Yes, it has better yields then the 90nm process... but yields aren't the absolute measure of a process effectiveness... It has less defects that cause a die to be useless, but it has more electron leakage causing lower speeds

It doesn't have more leakage...and that's not why it's only using slower clocks at the moment.

... Now ofcourse that could be due to the processor design itself... but what about when the thoroughbred B added an enitre new layer for a revision? that is a change in manufacturing process.

It's not actually, it's a design change. T-Bred B used exactly the same 65nm manufacturing process as TBred A...

 

[dmens]

Originally posted by: CTho9305
Oh really? FWIW, speed bin increases often come from new revisions that generally make small changes with large effects.

What if the fix is an all-layer spin and you don't want to wait 3 more months to start shipping? It would be easy to have Fmax limiters that could be fixed with a relatively simple revision. Sure, once you've been through a few revs of a chip you'll have a good idea of where the real silicon critical paths are, but for a brand new design, you might just be unlucky and run into a case where the timing analysis didn't catch the worst path (spice models aren't perfect).

What if there's a single critical path that stands out far beyond all others, and you fix it? There are other issues that could put similar artificial limits on Fmax - I've heard stories of old Alpha chips which had an Fmin caused by leakage on dynamic nodes - for the fastest process corners, the Fmin was above the Fmax and the chips wouldn't work. If there were only a few problem nodes, simple changes could probably fix the Fmin issue, making the very fastest parts sellable again.

small changes with large effects is the norm on early silicon, not when the product is already on the market. timing and marginality outliers also tend to be discovered early on.

 

[dmens]

Originally posted by: Viditor
It's not actually, it's a design change. T-Bred B used exactly the same 65nm manufacturing process as TBred A...

it's both a design and process change. an extra metal layer isn't exactly "the same manufacturing process", nor was t-bred at 65nm.

 

[stevty2889]

Originally posted by: taltamir
3. It is cheaper, it makes less heat, it takes less power... all very good reasons to use it to overclock... Rather then taking a maxed out 90nm part and trying to OC it.. oh yea, and because of the black edition 5000 prcessor having an unlocked multi but NOT being maxed out... (like the 90nm black edition part which sells for much more and comes with 3.2ghz default clock). The problematic part is that it achives max speeds lower then the previous process, which is unheard of... new processes always achieve MUCH faster max speeds.

Did you forget about Intel and Prescott..the 90nm P4's that at the beginning, not only did they not have higher clock speeds than the previous generation, but they used a lot more power, produced a lot more heat, and in many applications they were actually SLOWER at the same clock speed. And then they topped out at a whopping 3.8ghz, compared to 3.4ghz of the previous generation. Even when the shrunk to 65nm they STILL didn't pass 3.8ghz. And there was also the abomination known as the pentium-D as well. Meanwhile Dothan was also made on the 90nm process and did just fine. The process isn't always the problem, the chip design is a huge factor as well..

 

[coldpower27]

Originally posted by: stevty2889

Originally posted by: taltamir
3. It is cheaper, it makes less heat, it takes less power... all very good reasons to use it to overclock... Rather then taking a maxed out 90nm part and trying to OC it.. oh yea, and because of the black edition 5000 prcessor having an unlocked multi but NOT being maxed out... (like the 90nm black edition part which sells for much more and comes with 3.2ghz default clock). The problematic part is that it achives max speeds lower then the previous process, which is unheard of... new processes always achieve MUCH faster max speeds.

Did you forget about Intel and Prescott..the 90nm P4's that at the beginning, not only did they not have higher clock speeds than the previous generation, but they used a lot more power, produced a lot more heat, and in many applications they were actually SLOWER at the same clock speed. And then they topped out at a whopping 3.8ghz, compared to 3.4ghz of the previous generation. Even when the shrunk to 65nm they STILL didn't pass 3.8ghz. And there was also the abomination known as the pentium-D as well. Meanwhile Dothan was also made on the 90nm process and did just fine. The process isn't always the problem, the chip design is a huge factor as well..

The 65nm process not passing the 3.8 GHz frequency on NetBurst of the old process was a little different, most of the thermal headroom was used to make somewhat slower Dual Cores rather then a really fast Single Core.

By the time the 65nm process had rolled around Single Core was on it's way out, so they only needed to release something faster then the Pentium XE 840 which they did with the Pentium XE 955 right from the get go. So right from the get go the process did provide frequency headroom, due to reduced thermals primarily.

 

[stevty2889]

Originally posted by: coldpower27

Originally posted by: stevty2889

Originally posted by: taltamir
3. It is cheaper, it makes less heat, it takes less power... all very good reasons to use it to overclock... Rather then taking a maxed out 90nm part and trying to OC it.. oh yea, and because of the black edition 5000 prcessor having an unlocked multi but NOT being maxed out... (like the 90nm black edition part which sells for much more and comes with 3.2ghz default clock). The problematic part is that it achives max speeds lower then the previous process, which is unheard of... new processes always achieve MUCH faster max speeds.

Did you forget about Intel and Prescott..the 90nm P4's that at the beginning, not only did they not have higher clock speeds than the previous generation, but they used a lot more power, produced a lot more heat, and in many applications they were actually SLOWER at the same clock speed. And then they topped out at a whopping 3.8ghz, compared to 3.4ghz of the previous generation. Even when the shrunk to 65nm they STILL didn't pass 3.8ghz. And there was also the abomination known as the pentium-D as well. Meanwhile Dothan was also made on the 90nm process and did just fine. The process isn't always the problem, the chip design is a huge factor as well..

The 65nm process not passing the 3.8 GHz frequency on NetBurst of the old process was a little different, most of the thermal headroom was used to make somewhat slower Dual Cores rather then a really fast Single Core.

By the time the 65nm process had rolled around Single Core was on it's way out, so they only needed to release something faster then the Pentium XE 840 which they did with the Pentium XE 955 right from the get go. So right from the get go the process did provide frequency headroom, due to reduced thermals primarily.

True, but my point was that a die shrink, doesn't automaticly mean there is improved headroom for clock speed. Of course if prescott had been a direct dieshrink, it may have done better.

 

[coldpower27]

Originally posted by: stevty2889
True, but my point was that a die shrink, doesn't automaticly mean there is improved headroom for clock speed. Of course if prescott had been a direct dieshrink, it may have done better.

Agreed, if the architecture changes along the way all bets are off, the Northwood to Prescott represents somewhat of an architecture change with the amount that was added to the processor core vs the Prescott-2M to Cedar Mill which was pretty much an optical shrink.

But I think from what we seen so far as long as you hold the architecture factor constant then the optical shrink should yield at least some thermal and clockspeed benefits. How much is up for debate though.

It also depends which company your looking at, AMD on the 90nm process didn't change their architecture and didn't get any additional frequency headroom at first, that didn't come till later down the line. Though they did quite the reduction in power consumption. Performance improvements came along later down the line as seen with Venice/San Diego.