Silicon Investor (SI) -- The First Internet Community

STOCKTALK

We've detected that you're using an ad content blocking browser plug-in or feature. Ads provide a critical source of revenue to the continued operation of Silicon Investor. We ask that you disable ad blocking while on Silicon Investor in the best interests of our community. If you are not using an ad blocker but are still receiving this message, make sure your browser's tracking protection is set to the 'standard' level.

Technology Stocks : Advanced Micro Devices - Moderated (AMD) -- Ignore unavailable to you. Want to Upgrade?

To: Sarmad Y. Hermiz who wrote (228902)	3/25/2007 5:55:26 PM
From: neolib	Read Replies (1) \| Respond to of 275872

It is an admission that things are approaching effects due to small numbers. IIRC, it claimed that something on the order of 100 ions relatively shallow in the channel set the threshold. So yes, that is getting down there. Still a long ways from being concerned with an individual atom. Did you see the 3D graph of ion distribution in the gate & channel (blue & red dots)? Moving any single atom in there is not going to do anything. Unfortunately, the article didn't provide any data on what variation is typical out of the implanters. Since both Intel & AMD buy these from 3'rd parties, I would not expect Intel to have any advantage over AMD in channel implanting.

To: Sarmad Y. Hermiz who wrote (228902)	3/25/2007 7:19:44 PM
From: pgerassi	Read Replies (2) \| Respond to of 275872

Dear Sarmad: You are missing something about APM. It gathers results based on parameters and the machines they run on. It is entirely possible that certain machines do a better job on certain stages of the die fabrication. No one needs to know why it works, just that statistically it is better to do steps xx to yy on these pieces of equipment. That is for step 14 equipment serial no xxxx produces parts 5% faster at same power on average that the same equipment serial no yyyy does. So APM would direct those steps to be done with xxxx and other steps where it doesn't matter so much with yyyy. So the CPUs from APM's control come out 5% faster than those with copy exact. Copy exact doesn't determine which equipment does best just that you use ACML scanner model 1234 for steps 14, 24, 36, ... and which scanner serial number doesn't matter. Intel just picks one and fixates that one to do all for line 1, another for line 2 and a third for line 3. Because of this copy exact misses opportunities. Sometimes a tweak just works better. Copy exact doesn't find it, because once developed, it stays the same, thus giving far fewer chances to find it. Because APM always gathers information, as number of runs increases, so does the chance of a tweak being better is recognized. Also a why a tweak works can be found out by various means or even by deliberately incrementing or decrementing parameters and looking at the results. You might stumble over a better set of parameters or looking at the data have an inspiration which causes you to try a different set of parameters not in the original search grid and boom it turns out even better. If it doesn't, that inspiration went down the wrong path and is discarded. If the inspiration works, you apply it to not only to that step, but others and see if it has global implications. If so, it is added to AMD's IP and is another reason why APM is better than copy exact. This does happen in industry. I remember once an induction furnace yoke cracked and it wasn't discovered for a while. That cracked yoke somehow improved the furnace's efficiency by 1.5%. Given the enormous amount of energy used per batch, 1.5% turned out to be $250K per year in electricity savings. They had a study done to find out why this worked, how could it be done deliberately and how much better could the savings be. It turned out that the yoke design wasn't as good as initially thought. The long and short was that a different yoke with bigger air gaps had a 4.5% better efficiency due to a better impedance match to the type and composition of the typical batches used. Copy exact would never have found that out. So the company now saves $750K per quarter for each of its 12 induction furnaces easily paying the $150K for the study and $1.2 million for the implementation. A 4.5% CPU performance improvement alone is worth hundreds of millions over time. It is the equivalent of doubling the cache or about a third of a process generation. It appears that APM does that every 4 months or so. Thats 14% in one year, 31% in two years and 50% in three, more than enough for AMD's 45nm to be ahead of Intel's 32nm. It is also why AMD's process switches yield less apparent improvement than Intel's. It isn't that AMD gets any less of an improvement than Intel on process switches, its that AMD has more improvements from APM on the previous process. It is why AMD needs to start making the next process generation sooner and put it to making high volume low end CPUs. It will take APM some time to get the newer APM process improvements added to the inherent process improvement to cross over the older's process APM improvements. Once that happens you switch over the high margin high performance stuff. For 65nm, the Barcelona design improvements will cross over the 90nm APM improvements. The same thing happens to Intel, but to a far lesser extent. The C2D design also trumped the Intel 90nm process maturity over Intel's 65nm process. Given the same, look for Penryn not to be faster than Conroe for the first year. Only the redesign for ODMC and CSI will trump Woodcrest and Cloverton enough to switch servers and high end workstations to 45nm. By that time they might be able to get to initial 65nm Barcelona. ZRAM2, Torrenza and Fusion will do it for AMD, placing AMD well ahead of Intel again, perhaps for good. By that time look for copy exact to be dumped for something akin to APM (renamed of course, due to Intel NIMBY syndrome). Intel belatedly will follow AMD again. Pete