On Friday we posted a cautiously optimistic quick look at Athlon 64 thermals, but stated we were going to return for a more in-depth appraisal.  The delay (and split article) was caused by hardware compatibility issues that had us swapping motherboards three times in order to find a board that supported both chips in BIOS and didn?t have any voltage issues (our A8V beta BIOS supported both, but insisted on running the 90nm Winchester at 1.60v.  As you can imagine, this is a less-than-healthy voltage for our 90nm part. We’ve spent more time looking into the chip across a variety of speeds using the MSI Neo2, and would like to share the results. 

Before we discuss our findings, lets talk a bit more about the 90nm problem in general.  Because my own background in electricity and physics is lacking, I?ve kept the discussion high-level?if you feel you?ve got something to add, feel free to email me. 

Intel?s Prescott Headache first sign of 90nm Trouble.

As we stated in our earlier article, general public expectation was that Prescott would offer improvement to Northwood in the same manner that Northwood improved on Willamette.  Even though Prescott was known to have architectural features that made it less-efficient than Northwood, those features were explained as being compromises that would, in the long run, allow for a higher-scaling CPU. 

When Prescott arrived and proved to be both slower than Northwood in most tests and much hotter we hypothesized that Intel might re-spin the silicon and release a cooler-running part, similar to how AMD did with the Thoroughbred-core Athlon. Nine months later, this is looking less-and-less likely.  LGA775 offered no cooling benefit beyond the better heatsink / fan combo, and there?s no reason to think Intel has a new die revision coming that will drastically change Prescott?s thermal signature.

Its still not clear how much of Prescott?s thermal trouble is unique to the processor and how much is caused by Intel?s 90nm process, but the months since Prescott?s launch have proven the problem to be much bigger than Intel.  IBM?s much-touted G5 shipped months late on 90nm and forced Apple to simultaneously delay the iMac and use water-cooling to push the CPU up to 2.5 GHz. 

Using dual cores to dodge high-frequency problems, Conclusion.

If you look at our graph and then consider what we know about dual core technology, the reason for going with such an approach is starkly apparent.  Because CPU's are extremely poor at translating heat laterally, its possible to include two cores in a single package without overheating any particular area of the chip. 

In relation to this, the current set of heatsinks offered today are typically limited by the amount of thermal energy able to radiate through the die, not by their own ability to absorb and dissipate the heat.  Not only does this mean that current heatsink designs shouldn't need drastic revision to support dual cores, it means that, thanks to lower temperatures at lower clocks relative to 130nm, AMD and Intel may be able to return to the old paradigm of a die shrink = lower operating temperatures and higher clockspeeds.

How much this becomes a factor will depend on how good yields look at particular clockspeeds, but don't be surprised to see dual core chips launching at 2.4 GHz for an "effective" 4.8, while the single core versions are at 3 GHz or more.  Whether or not software will be optimized to take advantage of the situation is another issue entirely; one that will be resolved only with time and a great deal of effort. 

The bottom line is that while AMD's 90nm process doesn't carry the tremendous thermal baggage of Prescott, its clearly on the same general path if our thermal diode was reporting properly.  Does this mean AMD's ahead?  On the one hand, yes, compared to Prescott--but its not the kind of lead that just puts them head-and-shoulders above everyone else--its more like the kind of lead that means Sunnyvale might be able to make their transition smoothly and quickly enough before 90nm single-core thermals become a major issue.

Remember kids, in a race between you and physics...physics always wins, in the long run.

So should you buy a 90nm chip?  We want to do more research to confirm temperatures, but if you want a Socket 939 board that won?t break the bank, we?d say yes.  Its also possible that, even if Athlon 64 is running hotter at the moment on Socket 939, AMD could always re-spin the silicon to cool things down, as they did with Thoroughbred. 

At the moment we?d say that 90nm Athlon 64 is running hotter than 130nm, even though this reverses our Friday position.  We?re not taking a firm stance yet; there?s more data to be gone through.  Either way, the CPU is still a good deal--and for the first time, an affordable price. 

< previous - (5 of 5)

---

• Anatomy of a Die Shrink. (1 of 5)
• Motherboard Thermal Measuring (2 of 5)
• The Athlon 64 3200+: Courtesey of Monarch, Idle Temps. (3 of 5)
• Athlon 64 Load Temperatures: (4 of 5)
• Using dual cores to dodge high-frequency problems, Conclusion. (5 of 5)