In true role reversal, the past six months have seen AMD do just about everything right and Intel just about everything wrong. You aren’t doing well when you have built every kind of motherboard for your best CPUs except a good one and are suing the only people who have.
The Athlon is a truly fine processor. AMD has done a great job so far, and looks to do the same in 2000. However, what is best for AMD or even computing is not necessarily best for overclockers. Even though the Intel DeathStar has looked like a Rube Goldberg contraption lately (did Vader ever have to hit the planet-killer button twice?), it looks like they’ll have their turn in the sun for a while.
The key to overclocking is “What speed can you hit at what price?” Right now, advantage, AMD, but will this continue?
What is the current Athlon overclocking situation?
1) The 500MHz CPUs are cheap (about $200).
2) The CPU cores can run much faster than that, indeed, the average 500MHz Athlon CPU nowadays is really a remarked core.
3) Like PIIs and Katmai PIIIs; the Athlon uses outside L2 cache chips. Unlike those Intel chips, though, you can adjust the Athlons’s cache/CPU ratio. While adjusting the ratio downward to increase the CPU speed causes a performance hit, it usually is a net gain. While adjustable cache ratios gives you more overclocking flexibility than you had with PII/Katmais, they still limit top speed.
1) The Athlon is much more difficult/expensive to overclock than Intel CPUs. You either get out the soldering iron or you pay someone to do it for you. The first is beyond the (rational) capacity of most overclockers. The second is fairly pricey. One can buy a preset chip guaranteed to run at 700/750 MHz and pay a premium of about $200 over the cost of the processor. Or one can take the risk of the chip not overclocking enough and buy a pre-soldered processor with a series of dip switches to control processor settings. A simplified set will cost $75, but unless your motherboard lets you control the cache ratio; you’d better get the $125 model (see below).
2) AMD first made a whole lot of .25 micron K7s, then went on to the .18s. These processors have a number of differences, both in processor core and cache. Most notably, you have considerable variation in cache speeds. If you can adjust the cache ratio and do not have insane O/Cing fantasies, this isn’t too bad at all. It is bad if you can’t change that ratio. If you can’t and you get a processor with 4.0ns cache, you aren’t overclocking very much.
3) The current motherboards are flaky. While a good deal of this no doubt are initial BIOS bugs and Darwinian selection of overclocking wannabes, the number and types of problems faced by users are more than people from the BX board era have come to expect. There are too many memory problems; too many spontaneous reboots from people with approved power supplies.
4) These are .25 micron power-mongers that strain the capacity of power supply and heatsinks fans as you run them faster and faster. Sure, you can get 700-750MHz, and that’s pretty amazing; but you are pushing the envelope expecting anymore from them.
Now let’s see what the near-future holds for the Athlon:
.18 micron migration appears to have gone well for AMD, but two overclocking barriers become quite apparent.
First, cost. It looks like AMD has decided the 500-700 MHz range is .25 micron territory, and when they are gone, that will be it. 750 MHz appears to be the bottom floor for .18 micron, and they are not going to be cheap for a good while. Add to that the (by then cheaper) cost of “golden fingers” and premiums for Slot A boards. You will have high-performance, high-cost chips for the first half of 2000, and overclockers want high-performance, low-cost chips.
Second, AMD has limited the overclockability of these chips compared to the .25 micron chips by making the cache chips run at 40% of clock speed. If you have a 500 MHz processor with 250 MHz cache, you can set that cache to 1/3 and potentially still get 750 MHz.
A 750 MHz processor will have 300 MHz cache chips. If you lower the cache/CPU ratio to 1/3; that gives you only 900 MHz. Assuming normal leeway in cache capability, 950 MHz-1GHz.
While that is not bad, and certainly is much better than the historic overclockability of AMD chips, what will the Dark Side will have to offer with Coppermines in the near future, and at what price?
1) Coppermines will have ondie cache running at the speed of the processor. This, combined with better cache latency at least temporarily eliminates or reduces most of Athlon’s initial design advantages. Early indications are that Intel followed the same path they took with Celeron and designed cache to run at the speed of the family of processors. Despite the problems Intel has had with Coppermines, the few who have tried to overclock high-end Coppermines are getting over 800MHz without much problem so far.
2) Unlike AMD, Intel has brought .18 micron CPUs fairly far down the food chain. They aren’t cheap or even very available yet, but they will be. If Intel follows recent patterns and slightly accelerates its planned price cuts, by February, slot 1 600 MHz Coppermines should cost about $250. By that time or a bit later, Intel will probably have a new stepping for the Coppermines which should get them consistently operational up to around 900 MHz or a little more.
3) Unlike Athlons, Coppermines will be able to take advantage of a simple increase in FSB. No soldering required. A 150 MHz FSB is feasible with the best of current equipment; 133 MHz is doable with much current equipment. PC133 SDRAM costs little more than PC100. Given the existence of 100 MHz FSB Coppermines; a 600 MHz Coppermine should be able to run at 800 MHz at 133 MHz with ease, and could get up to 900 MHz in February. AMD will have nothing like that available for a few months at a comparable price.
In a word, motherboards. No overclocker is going to buy RAMBUS any time soon, the 810e is a joke, and the converted 820 is not much better. The BX boards can run the processor speeds, but the AGP ratio becomes excessive.
The Apollo Pros have been unpopular because of their performance hit compared to BXs, but the new Apollo Pro 133As tweak performance up to within a smidgen of BX levels. They should be fairly common by February, so even if Intel’s Solano is delayed, or ends up being another flop, there will be a reasonable and affordable platform on which to overclock these Coppermines. With an AGP 4X at a ratio of 1/2 and a PCI divisor of 4, a 150MHz FSB operation will cause no more strain than a 75/112 MHz overclock.
While I’m sure there will be some motherboard problems along the lines of those faced by the Athlon pioneers, the Apollo Pro has been around a while and not had too many problems
Athlon supporters will point to 1 GHz processors, copper interconnects, 266 MHz busses, socket As and on-die cache. They are quite right to do so. They will greatly benefit those who will wait, but they won’t be available and/or be cheap for the next six months, and the Coppermines will.