What to Make of Anomalous Thoroughbred Overclocking Results


By Van Smith

Date: June 18, 2002

Last week we reported the storm of controversy surrounding an enthusiast sites' claim of successfully overclocking the new Athlon XP2200+ Thoroughbred core by 360MHz.  This deviation was so great from attempts announced by all other sites that there was widespread skepticism of the anomalous report's  legitimacy.

Responding to the clouds of doubt surrounding their numbers, AMDMB produced a follow-up article.  In it the site reported obtaining slightly slower but essentially identical results for the same chip in another platform.  Moreover, their chip seems to come from the same batch as all of the other review parts.

We have no reason to suspect that AMDMB is making fraudulent claims.  Accepting their test results raises interesting questions.  A quick summary of the possible explanation for their "Golden Chip" follows:

  1. The sampling size was simply too small to draw any conclusions about overclocking potential for the distributed batch of Thoroughbreds.  However, the odds are against this unless AMD deliberately chose to distribute to reviewers poorly overclocking Thoroughbreds while intentionally cherry-picking a part for AMDMB.  This seems unlikely unless AMD was trying to demonstrate the overclocking headroom of the new design through a single source.
  2. AMD is binning out higher speed chips with the process mix used in the review parts.  This explanation would account for the faster CPU -- in AMD's speed tests, this one microprocessor might have simply failed a single qualification test for 1.9 or 2GHz and was therefore binned at 1.8GHz.  However, it is unlikely that AMD is producing chips faster than 2GHz from the process mix used in the review chips given the voltage requirements of the new 2200+ Thoroughbreds.
  3. AMDB received a "Golden Chip," a freak of process technology.   If so, this level of deviation suggests a process issue.  For instance reticles are sometimes calibrated only within the very center of the "field of view."  However reticle imperfections can lead to deviations in dimension uniformity away from the calibration area.  In other words, all chips outside of the center of the reticle are susceptible to imperfections that can slow them down -- only the dies where the reticle is most nearly perfect will reach the process's potential.  If this is the case, then depending on the coverage area of the reticle, perhaps one in ten chips would consistently bin faster than all of the rest.  In the worst case that AMD has a defective reticle that must be replaced, this could serve as an expensive setback. Even though their chips are capable of much faster speeds when produced with faster transistors, AMD simply might not be able to produce faster transistors right now because the same deviations with shortened gates could cause yields to nosedive.  However, this is a worst case scenario and it might be possible to correct such a problem inexpensively and quickly.
  4. There is a design-limited speed path in the Thoroughbred.  This is highly unlikely given that the design is very similar to Palomino's.

What is most important to understand is that right now the market is soft and both AMD and Intel are suffering through a chip glut.  AMD’s Thoroughbred makes an extremely attractive alternative to the Pentium 4 in the mobile space which is still showing some life especially for AMD.  In fact, while a mobile Thoroughbred XP1600+ has a maximum power dissipation rate of only 16W, the Pentium 4 can use as much as 20-27W simply idling.  Since mobile chips have higher ASPs and higher margins, it makes sense for AMD to secure this space first especially in today’s market.

The current batches of Thoroughbreds seem to demonstrate very good “leakage” characteristics (the power consumed by the chip when it is doing nothing) indicating that AMD is using conservative, mobile friendly process skews with slower transistors.  In fact, the current Athlon XP2200+ can almost be viewed as an overclocked mobile part.  Pronouncements that AMD is “slipping” with the Thoroughbred based on early overclocking attempts are simply wrongheaded or worse - perhaps even FUD-mongering.

Our sources tell us that AMD has been diverting most of its resources to Hammer because of the design’s great promise.  Although recently published benchmark results for an 800MHz-locked Clawhammer were impressive in themselves, we have been told that there have been 1600MHz chips available for quite a while as well.  With its longer pipeline and cool-and-fast SOI process, the Hammer should scale to significantly higher speeds than the Thoroughbred.

Most exciting is that we have very good reason to believe that AMD has accelerated a design that we weren’t expecting until 90nm Hammers.  In fact, samples of this innovation should be available by now or within the next week.  We will elaborate on this development shortly.

Lastly, Intel has stepped up its FUD War.  Expect the disinformation campaign that ratcheted up prior to Computex to continue to escalate perhaps through next year, and especially in the server space.  

An interesting related development is that Intel’s long-used kickbacks (exclusive discounts or incentives) delivered to marshal OEMs into Intel-only contracts may have found the end of the line.  Apparently Intel has been so aggressive recently in threatening to pull these incentives from OEMs that support AMD products that at least one pressured company might now be thinking of reporting the chip giant to the government for anti-competitive behavior. 

Inside the industry, Intel’s bombastic monopolistic practices are infamous and many have wondered when one of the beaten down parties would stand up to the Santa Clara-based chipmaker.  It will be interesting to watch as this story develops over the next year after Hammer storms out.


Pssst!  We've updated our Shopping Page.