Battleground ready for AMD Athlon 64, Intel Prescott spat

WITH THE RELEASE of the first set of Nforce3 based Opteron numbers ( here), AMD has fired the first shot in the computing dogfight that is coming in Q4 of this year, Prescott Vs Athlon64. If AMD sticks to the unofficial rumors of a socket 939/940 dual channel Athlon64, then we are seeing a pretty good preview of the upcoming performance numbers. AMD will release the Athlon64 in two flavors, 3100+ at 1.8GHz, and 3400+ at 2.0GHz.

The Opteron reviewed on AMDZone is running slower, higher latency memory, and is based on an unreleased motherboard. Even so, it is fairly close to the P4/3200 that is reviewed along side it. Bumping the performance up a few percent for the new mobo, and a few more percent for the faster ram, DDR400, non-ECC, non-Registered, and you have something that will have no problem living up to it's 3100+ moniker. The 2GHz version should also be able to pull off a credible 3400+ rating, and move AMD into the desktop lead for a few months until Prescott hits at 3.4GHz or more.

There is a slight problem with the numbering scheme, Intel isn't playing by the same rules, and the ballgame has changed since the PR scheme was first announced. Intel rates the CPUs in MHz, and the Prescott 3.4 is much more than a mere .2GHz jump from the current P4 3.2GHz. Double the cache, more instructions, more just about everything, and you get one of those big jumps in the performance graph that one expects from a new core. This should put it at least a bit ahead of the Athlon64 2.0GHz, probably a lot ahead. The problem lies in the numbering scheme.

AMD measures its performance by performance, not MHz, so if a new core comes out, and it is 10% faster, AMD has no problem calling it the same PR rating as a 10% faster clocked older core. So far, so good, it is an internally consistent rating that lets you pay for what you get. One look at the controversy surrounding the ratings of the Barton cored Athlons, and the furor that erupted when some benchmarks ran slower on the "faster" chip shows that AMD has a potential PR problem on their hands.

Adding to that problem is the fact that the tests AMD runs ( here) to benchmark the CPUs are seriously outdated. Consistency is good to a point, but updating on a regular basis in a transparent manner is better. So, you have processors from one company that are measured by one yardstick, and processors from another measured by a completely different yardstick. The problem is that both give similar looking numbers, and they are becoming less and less comparable over time. Prescott will only make this trend more obvious, something must be done.

While AMD has been accused of sitting on its hands here, that is quite far from the case. There is, or at least was, hopefully it is settled now, a lively internal debate at AMD on what to do about the problem. Everyone I know there will readily admit that it is time for a change, and just about everyone has a different idea about what that change should be. I can see a lot of hair being lost over this one.

There are four major paths for AMD to take in regard to marketing the Athlon64, all of which have good and bad points. First, there is the "do nothing, it is working camp", followed by the closely allied "update the tests" camp. More radical people have been calling for an abstract numbering scheme, like the Opteron is using. The last idea, and I have not heard anyone at AMD seriously say this, is to use simple MHz numbers like Intel. Lets look at these in greater depth.

The first is the "do nothing" camp. While I have heard one AMD engineer say this, I don't think even he took it very seriously. Apps in the benchmark suite are far from current, and the only game in the list that still sells is QuakeIII, and that is about to be supplanted by Doom3 about the time the Athlon64 comes out. Overall, this is not a good way to test anymore. If you add in the fact that the apps listed are not SSE2 aware, you have the curious situation of AMD's own benchmarks not taking advantage of the performance their latest chips offer. All is not hopelessly lost for this scheme, it does maintain consistent ratings going back to the Thunderbird/Palamino era. Then again, so what? The current x86 compatible chips more or less maintain backwards compatibility to a chip that no self respecting dishwasher would dream of using, so this may not be a good thing in the long run. How much call is there for AdLib compatibility in sound cards any more?

Next up is the "update the benchmarks" scheme. This is a good option, and one of the two taken seriously by the AMDers I talked to. An updated suite of programs, kept consistent for the life of the K8 core could potentially be just as relevant as the current scheme. If people bought the first scheme, and by all accounts, even the most jaded geeks did, this one should be an easy sell. AMD obviously knows how to construct a good suite of programs for the ratings.

As option 2.5, AMD could have a list of programs, and update it on an infrequent, but regular basis. About every 2 years seems to be a fair tradeoff between stability and outdatedness. If AMD publishes benchmarks, and shows the relative gain/loss of the older chips on the newer suite, it would allow enthusiasts to keep relative performance straight, and the general public, as usual wouldn't care.

The third scenario is the Opteron type nomenclature. This one is the second most talked about plan, and has some merit, but much less merit in the consumer market than the corporate market. If you are paying $50K for a server, or more likely $50K per server for a large number of them, you are going to make damn sure you know how it is going to perform. People run benchmarks before spending hundreds of thousands of dollars, you would be a fool not to. To this crowd, 3.0GHz vs a model 244 is largely irrelevant, how fast it runs the code I want it to however is not. The buyers here are more than smart enough, and educated enough to make rational decisions. The Athlon64 crowd largely is not. Model numbers like this will only make the sales job harder for the average Best Buy employee, and lead to less sales.

While it may be a good idea to make a strong brand identity with performance numbers as a large part of it, I would think the downsides outweigh the gains in the short term. AMD can't afford a short term loss right now, so I would think this one is at best going to be shelved for a later date, or used along side the old PR scheme. Do you think an Athlon64 Model 146 PR3400+ is a bit cumbersome? Me too. Some names are meant to be shorter rather than longer ( here).

The last numbering scheme is not one, it is the simple, the good old fashioned method, give the people the speed. Because of the radically different architectures of the two chip families, they are in the curious situation of rough performance parity at hugely different speeds. The average American, think knuckle dragging, cheap beer swilling consumer drone here, if faces with two choices, one 50% bigger than the other. For the same money, a buyer would obviously go for the bigger number, and most likely the case with the fake chrome, side window, and neon light.. This is why those miserable excuses for a chip, the Celeron, still sells. I would rate this strategy somewhere between voluntary gum surgery and slow suicide for AMD, they are not, by design, able to come close to parity in the MHz race. IPC is a hard concept for most borderline geeks to grasp, and more or less impossible for the scraped knuckle crowd.

Where is this all going then? AMD released a bunch of clues the other day in a study that basically said people are confused by acronyms, and can't be bothered to learn (See here). Oh what a shock. Reading between the lines, I would think AMD is going to go for option 2, the simply updated benchmark suite. It isn't a bad choice, and can be done with credibility and consistency, that they have proved. If they don't let this scheme overage, it will work out well. Part of me says that they will go for the extended model scheme, but I doubt it at first. Anyone know for sure? µ