Why not just make CPUs bigger?

LoneWolf121188 · Jan 17, 2005

In the need for speed, both AMD and Intel are trying to reduce the size of their transistors so they can fit more transistors in a die. Why not just make the dies bigger? I mean, my 3500+ is what, 2cm square or something? but thats just because there has to be that number of pins. the core itself is much, much smaller than the whole processor unit with the IHS and everything. Why not just make the core take up more room under the HSF? there wouldn't be any change in the socket type, and it wouldn't require any more room on the mobo, but a 1.5x to 2x increase in the size of the die (is that the right term or am I getting die and core confused?) should roughly equal a 1.5x to 2x increase in speed when really no more space is being used. heck, I would rather buy a 939 .18 micron 3500+ than my .13 because it would be so much cheaper.

TimoneX · Jan 17, 2005

Economics of course. Larger cores produce fewer per wafer and hence are less profitable.

Unreal89 · Jan 17, 2005

wouldn't it also draw much more power?

cack01 · Jan 17, 2005

Also the longer the pathways the harder it is to get a clean signal to each area of the CPU.

bandit390 · Jan 17, 2005

Would take more voltage and the electons would have to go further. Suposed to make comparatively less heat too. Unless your a prescott.
*edit I took to long to spit it out and everyone beat me to it lol*

LoneWolf121188 · Jan 17, 2005

TimoneX said:
Economics of course. Larger cores produce fewer per wafer and hence are less profitable.

Yeah, but don't forget that current yeilds on the newest chips are like under 20% (I know the nvidia NV40 and 45 have about 22% and 18%, and those are .13s...yes, I know thats GPU, but its similar, at least at the fab level). That means for every wafer, you get only one or two usable chips. Also, would it make that much of a difference, in terms or signal degration and voltage?

Steven4563 · Jan 17, 2005

i think he meant instead of making the die smaller y not just make the die bigger with more transistors instead of it being 50million for exmaple if the made the die with 0.90NM but made it larger with say 100million transistors wouldnt that be faster ?? hmmm maybe im not good with the inside of a cpu lol

Aphex_Tom_9 · Jan 17, 2005

LoneWolf121188 said:
Yeah, but don't forget that current yeilds on the newest chips are like under 20% (I know the nvidia NV40 and 45 have about 22% and 18%, and those are .13s...yes, I know thats GPU, but its similar, at least at the fab level). That means for every wafer, you get only one or two usable chips. Also, would it make that much of a difference, in terms or signal degration and voltage?

wow, i didnt know the yeilds were that low!

TimoneX · Jan 17, 2005

LoneWolf121188 said:
Yeah, but don't forget that current yeilds on the newest chips are like under 20% (I know the nvidia NV40 and 45 have about 22% and 18%, and those are .13s...yes, I know thats GPU, but its similar, at least at the fab level). That means for every wafer, you get only one or two usable chips. Also, would it make that much of a difference, in terms or signal degration and voltage?

nVidia's GPU problems are an extreme example of a die shrink gone bad. If you take a look at AMD's shrink from 130nm to 90nm you're talking about theoretically 72% more chips per wafer. 145x145 vs 84x84 for the A64 mobile. That's a big increase and it's what drives these companies to risk these die shrinks. For us it usually means lower power usage & greater OC headroom.

Dylruss · Jan 17, 2005

Cannot forget the idea that the bigger the die of the chip, the better chance to have impurities in it.

perfectturmoil · Jan 19, 2005

LoneWolf121188 said:
...under 20% ...for every wafer, you get only one or two usable chips...

I believe theres a bit of misinformation. Like someone said, yields under 20% are terrible.. and that must be a limited case.

Second, even if its 20%, how does that relate to 1-2 chips? Now, I'm not positive, but in most cases wafers do not contain 10 or so chips. Wafers are big, and chips are small.

I just looked online, and saw that AMD uses the 6 or 8 inch wafers in their fab process. We use the same size (well.. 6 inch) wafers here where I work, and there are thousands of parts per wafer. Of course, depending on the part, they are differnent sizes, but I unless AMD has a huge spread between chips, or the chips are about the size of the actual part we put in our computer, they must have at least more than 10 per wafer :-]

Sentential · Jan 19, 2005

perfectturmoil said:
I believe theres a bit of misinformation. Like someone said, yields under 20% are terrible.. and that must be a limited case.

Its *possible* that the yeilds are that bad, but frankly I would have to see it myself in person to belive it. I belive Xbit/The Inquirer said that AMD's .13 process is about 60-80% yeild rate and the .09 closer to 40%.

Naturally these numbers change daily so thee is no real way of knowing

attack · Jan 19, 2005

well adding more transistors doesn't increase they're speed...you can add more cache, but that has dimishing returns. What's happening now is your going to see dual core chips emerge (very very soon...as in less than 2 years you can easily buy them), and this will hopefully force software engineerings to code for SMP.

When you make the chip bigger (add more transistors that is), the only benefit you can get within the x86 instruction set that's usefull is more cache...I'm pretty sure adding registers, and stuff like that wouldn't matter because the instruction set wouldn't use them.

LoneWolf121188 · Jan 20, 2005

perfectturmoil said:
I believe theres a bit of misinformation. Like someone said, yields under 20% are terrible.. and that must be a limited case.

I know for a fact that the NV45s are about that low every time...IBM fabs nVidia chips and my dad is high enough up in IBM to have reliable info about that kinda stuff...can't say much more, I'm not sure exactly whats under the NDA (non disclosure agreement) and whats not...in fact, im not even sure that I was allowed to say that, but anyway, im pretty sure thats OK

...

@ajrettke: Um, isn't adding transistors to the chips what Intel and AMD have been doing all these years to make them run faster? I mean, thats where moore's law comes in...

@md0Cer · Jan 20, 2005

LoneWolf121188 said:
In the need for speed, both AMD and Intel are trying to reduce the size of their transistors so they can fit more transistors in a die. Why not just make the dies bigger? I mean, my 3500+ is what, 2cm square or something? but thats just because there has to be that number of pins. the core itself is much, much smaller than the whole processor unit with the IHS and everything. Why not just make the core take up more room under the HSF? there wouldn't be any change in the socket type, and it wouldn't require any more room on the mobo, but a 1.5x to 2x increase in the size of the die (is that the right term or am I getting die and core confused?) should roughly equal a 1.5x to 2x increase in speed when really no more space is being used. heck, I would rather buy a 939 .18 micron 3500+ than my .13 because it would be so much cheaper.

It is extremely inefficient in terms of cost. You make less CPU's and have less of a yeild rate meaning costs for these gigantic CPU's would skyrocket and they still would not be making as much money.

Another, is that the smaller the gate length, or process, or whatever you want to call it, the smaller the transistors are, meaning the smaller the core. Unfortunatly, the smaller this is, the better the CPU is with frequency and voltage. This means the CPU can run faster with less voltage. That is why they are trying for smaller and smaller. If we went back to .28micron, and still had the same number of transistors, you would have one huge CPU. A massive one. It would be a massive CPU that had a stock vcore of 2.3volts, topped out at 300MHz, and created just as much or more heat. Now we are bassically back to a modernized version of a Pentium with tons of transistors and modern technology packed into it.

deathman20 · Jan 21, 2005

Only problem with todays CPU's is even though by it getting smaller = less voltage. Yet we are putting out more and more heat. I know I saw a graph posted around here a while ago and I'm sure most of you seen it with it comparing temps back from the 486 days to the future with current trends of heat.

Yet by making the clocks faster on the CPU's they still have to legthen the pipes which in turn makes the CPU slower up to a certain speed then it gets better then the previous CPU (look at the 3.0C and 3.0E as examples).

@md0Cer · Jan 22, 2005

We are putting out more and more heat because CPU's are becoming less and less efficient. Back in the times of the Pentium it wouldn't have mattered much since clockspeeds and transistor counts were low. Now that we have more and more and more transistors and higher clockspeeds, using a long pipline is not very good in terms of thermal output. That is why the 31 stage Prescott at 90nm and 3.8GHz puts out MUCH more heat than for example, a 12 stage Winchestor at 90nm and even 2.2GHz. Or even the 10 stage Pentium M 90nm at 2.0.GHz. The thing could outdo a much higher clocked Prescott in performance, yet still has a maxTDP of 21 thermal watts vs. the Prescott's 100something.

So once the Prescott is no more and they use the Pentium M architecture, the amount of heat should go down.

Nandro · Jan 24, 2005

Isnt the Itanium a massive (comparatively) CPU in that regards?

TimoneX · Jan 24, 2005

Yeah McKinley was around 421(x2), Madison is around 374(x2) and I believe Itanium 1 was even bigger 450(x2)? These are large and expensive chips to produce.

LoneWolf121188 · Jan 24, 2005

kinda going along with this, though kinda off topic too (don't know if this deserves another thread, but anyway)...Why are AMD procs so much bigger than P4s? I mean, we're talking 478 pins vs 939/940! thats almost twice as many pins! is it cause of the onboard mem controller on the AMDs? but then again, P4 478 Prescotts have 2x as much L2 compared to the Clawhammers, right? 2MB vs 1MB?

Why not just make CPUs bigger?

Member

Closet Elitist Member

Member

Member

Member

Member

Member

Member

Closet Elitist Member

Member

Member

Contributing Member

Member

Member

Senior Member

High Speed Premium Senior

Senior Member

Disabled

Closet Elitist Member

Member

Similar threads