Tabula Rasa Semantics, in Microprocessor Burn-in. Part-I

CPU by KeithSuppe @ 2003-06-26

In my last article; "AMD Ingots, sliced TBread, with the crusts cut off" I proposed a hypothesis which purported to explain the propensity for "low speed" Thoroughbred-B's (1700 - 2100) to attain extraordinarily high clock-speeds. Athlon 1700's were achieving speeds as high as 2.7GHz on air cooling and minimal voltages. And in a few cases the coveted "double overclock" was achieved via LN2, and phase-change. Unfortunately, several of the premises on which my theory rested were incorrect. Here is correct version based on true facts, many hours and days of research has gone into this article. I hope you like it!

  • prev
  • Go to mainpage

More Corrections and SOI!

As I've already stated, many of the premises in my first paper were incorrect, however; there was also a great deal of validity in a few assertions. That there was a propensity for high overclocks among Lower Speed Tbred's due in part to AMD's binning process was correct. From the very start, low-end TBreds especially JIUHB 1700's either DUT3C or DLT3C, were attaining 2.5GH on air, and minimal Vcore. From the beginning AIUCB 2400 and higher, were failing to overclock past 2400MHz without excessive Vcore. And as I correctly contended in several posts at Xtremesys, articles had incorrectly identified the potential overclockers juxtaposing the order. It was not the "High End" Tbreds as Ed Stroglio asserted, but in fact the "Low End" Tbreds that were the best overclockers, ergo better quality cores.The reason for this overclocking prowess had not derived from some "error", but in fact the opposite phenomenon. AMD's mastery of the .13 micron die, (at least in so far as the TBred-B0) resulted in an ideal manufacturing environment. Clock rate, high performance ceilings, and exceptional wafer yeild.

Where one who is not an experienced overclocker or familiar with microprocessor manufacturing idiosyncrasies, overclocking results were confounding. One might expect higher end models to overclock at much higher frequencies, however in overclocking this is not necessarily so.
The experienced Overclocker, does not necessarily purchase the fastest processor in a specific line, but in fact the slowest. If the microprocessor manufacturer has had a successful yield, then all the processors should be capable of reaching the highest speed in the line, plus a performance buffer above that speed. And the overclocker seeks to exploit the manufacturing process in this respect, purchasing the low speed model (least expensive) and running it or overclocking it to its innate level of performance. This is why many overclockers attain such high speeds without having to raise the Vcore beyond default of the lowest speed processor in the line. Overclocking in this respect is not so simple as to be formulaic, but there is certainly a formula involved.

CPU makers manufacture based upon one core architecture, one set of photo-masks, consistent lithography, and wafer yield. Based from this singular core design, a maximum speed is achieved and maintained throughout, such that every core is capable of running this speed. The processors are then binned by a specific set of criteria which are conducive to their specific attributes. For example, if a certain number of cores are able to run at the lowest voltages, these will be binned as the lower speed models for this reason. If they need more voltage, they may be binned for packaging into higher end CPU's. Irregardless of their place in the model line, every CPU should have the ability to attain the clock speeds, of the fastest.

One advantage Low End Tbred owners had, was their cores were binned to run from a lower voltage of 1.5V. From the lower voltage, multiplier setting (although in TBred's case it's a moot point) they had a more stable base from which to overclock from, and a higher percentage, based upon it's low default speed. Those such as myself who purchased the supposedly High End Tbred-A's experienced the worst overclocking performance, and literally paid the price for it. I only wonder how many followed suit, hoping to AMD would put their best foot forward, and they placed it firmly in our mouths. At this stage I'd have to recommend against purchasing any future Barton's based on the .13 micron die. I'd also advice against purchasing latter models from the original .13 micron TBread masks, including 512K cache architecture. The only recommendation for Athlon XP's I can earnestly make is the "Low End" or rather Low Speed Tbred's, preferably of the B0 stepping. Although the 2100 A-stepping have been showing favorable results, at $42 for 1700's your almost guaranteed 2.4GHz, and from the 2100, perhaps 2.6GHz is feasible in some cases.

I would not expect anymore surprises from this core architecture. The initial Barton's also showed some promise, however; having the semblance of being a great overclocker, was (as in the case of the low speed Tbred's) merely a percentage in overclock from it's base speed. In other words, it had no where else to go but up in its overclocking ability. AMD unfortunately seems to have struck the 248nm wall in this respect, exhausting the wavelength's tune-ability. If you recall, just a few years ago the prediction was .13 micron die, would be achieved utilizing a 193nm wavelength lithography, here we’ve made due with 248nm, and in AMD's case, this has not produced the best overclocker's for the line's culmination, although in the lowerspeed models some of the best silicon ever, came off the line.

It depends upon how one views success. In my opinion the Tbred was entirely successful among overclockers, if one purchased the three lowest speeds. I cannot in good faith say AMD "pushed" the line too far, because the architecture did technically change with the expansion of the cache, and in this expansion perhaps more attention should have been paid to the higher PR's inability to overclock. They did meet, and surpass spec. however; and in this respect they were very successful. Only the end-numbers will tell. Of course up against Intel, they had their work CUT out for them, (pun intended). It was a statement made in a recent review which showed the desperation on the minds of those dedicated AMD fan's, (this from PCStats review):

Unlike some previous launches, the AthlonXP 3200+ is readily available which shows that AMD no longer has any manufacturing problems. What they need though is for the AMD fans out there to "put up" and buy the higher end processors instead of "getting the $50 processor and overclocking".

You know when the reviewers are asking the end-users to buy the higher priced models, and stop overclocking, times are tough.

Intel on the other hand, has done quite well with their .13 micron die, and Northwood-C seems to be the one of their best overclockers in recent years. Of course if the line is pushed too far past 3GHz, the ramifications will be the same. One reason the 2.4C overclocks as well as it does, is similar to the AMD XP1700 binned for low voltage, yet produced from the same wafer purity, photo masks, and Lithographic process as the 3.2C. Therefore if the performance ceiling of the 2.4C is 3.6GHz on air, one might expect the same from the high P4 models. The stepping identification process with Intel, however; is much more complex due to the number of Fabs, and especially the number of facilities where Lithography and binning occurs.

I'm somewhat ignorant to the Intel manufacturing method or rather assembly method, as I'm unsure to the extent they process their wafers. I know there are less Fab's then assemblies, however; where the final Lithographic and burn-in processes occur I'm not certain. Regardless, I do not foresee any overclocking behemoths in the latter Northwood-C's.

Will it always be thus? Cores binned for the lower speed models shall be the overclocking kings from that particular architecture? I do not think we can assign rigid designation to this theory. There are far too many variables affecting a microchips performance ceiling, such as fluctuations in wafer purity, differences among Fab's etc. And given the current metamorphosis in semiconductor manufacture as a whole, the process may be further complicated. We have certainly reached the end of the 248nm Lithographic life-cycle. And even common silicon itself is rapidly approaching obsolescence, or at least critical layers of it.

This is where SOI (Silicon On Insulator) manufacture technology will be critical for the speed, and low voltage requirements of ever smaller more populated circuits:

SOI refers to placing a thin layer of silicon on top of an insulator such as silicon oxide or glass. The transistors would then be built on top of this thin layer of SOI. The basic idea is that the SOI layer will reduce the capacitance of the switch, so it will operate faster.

In Part II of the Tabula Rasa theory, I will go more in-depth concerning the ramifications of SOI and Intel's take on this "insulative" process, SSOI. Stay tuned.

Madshrimps (c)
Click here

Comments, questions and suggestions can be placed in this thread @ our forums (no registration)
  • prev
  • Go to mainpage