First, big kudos to Tecchannel.de for getting their hands on a prototype Clawhammer system and doing a little testing.
While results at this very early stage should be interpreted cautiously, they do confirm a few general points that have been made about the processor.
The memory controller is the weight in Hammer’s head The bulk of improved performance comes from the integrated memory controller. It takes less time getting data out of main memory than more conventional approaches.
This is very good, provided what you’re using has to go to main memory all the time. However, CPUs (and many applications) are designed so that you don’t have to go to main memory all the time. That’s why you see those funny words attached to CPUs like prefetching and trace caching.
Ideally, a CPU should never have to make a request to main memory for something. Ideally, anything it wants should already be waiting for it inside the L1/L2 caches, which are much, much faster than anything in memory, and are right there, rather than be a horrible couple hundred nanoseconds away.
When that’s possible, programmers try to do that: write to keep most of the action in the caches. Where they’ve succeeded, the memory controller in Hammer won’t be much help, since it won’t be used so much.
There are situations where it’s impossible to keep all the action in the caches; you have to go to main memory all the time. This is where the memory controller will really shine.
So when we see benchmarking of the final product, vis-a-vis the Athlon XP, you’ll probably see rather lopsided results. Benchmarks of items like office applications probably won’t improve much at all, while audio and video editing should show stellar improvements (improved yet more by the inclusion of SSE2 capability).
Games will likely be a mixed bag. Quake happens to be one of those main memory junkies, so it does very well with Hammer. However, Quake (and those apps based on the same game engine) tends to do better than the average game with this improvement. Just remember that all games are not alike, and what’s true for Quake may not be true at all for your favorite game.
For the moment, as a very rough guide, if your game does considerably better with a PIV RDRAM system than an Athlon XP, it will probably be wonderful with Hammer. If it doesn’t, then it probably won’t.
The PowerPC Comes to x86 The Apple MacIntoshes do not use x86 processors from Intel, AMD, or Via. Instead, they use CPUs from Motorola called PowerPCs.
PowerPCs are by no means bad processors. They just march to the tune of a different design drummer. The pipelines are much shorter than those for x86 processors: seven for PowerPC G4 as opposed to ten for the Athlon, twelve for Hammer, and twenty for the PIV.
Generally, the shorter the pipeline is in a CPU, the more work that CPU can do in a clock cycle. However, the more work you do in a clock cycle, the fewer clock cycles you can have in a given period.
So a PowerPC G4 certainly does a lot more work per clock cycle than a PIV, but the fastest G4 runs at only 1GHz, while the fastest PIV runs at 2.5GHz (and will end up over 3GHz).
The problem with those who takes the G4 design route is that those who take the PIV route can ramp the speed of their chips to a degree that more than compensates for the relatively little work they do per cycle.
In short, 3 X 1 is more than 1 X 2.
This is why GHz all by itself can be misleading. However, that doesn’t mean they don’t matter at all. So long as it’s possible to get a better result from doing a little work very quickly as opposed to a lot of work just quickly, GHz (properly interpreted) will continue to matter.
The Athlon (and Hammer) falls inbetween the extremes of the G4 and PIV. The Athlon does more work per cycle than the PIV, but can’t run as quickly.
This is why Athlon XP (and their Thoroughbred descendants) have held up against the PIV despite spotting them hundreds of MHz. It’s also why the AthlonXP (and Thoroughbreds) are spotting them hundreds of MHz). Finally, it’s why the PIV has finally been able to overcome the XP through sheer GHz-age.
How Fast Can It Run Vs. How Fast Will AMD Says It Will Run?
If you take the Clawhammer Quake results at 800MHz and scale them out based on how well the MP scales, at 1600MHz, it will certainly blow away a PIV at 1600MHz, and probably beat a PIV at 2400-2500MHz.
That’s nice, except by then, the PIV will be running at 3000MHz or more. While you can’t put a ton of faith in these extrapolations; it’s hard to see how a 1600MHz can blow away a 3000MHz+ PIV.
At 2000MHz, this is somewhat likelier (though it still looks like “blowing away” rather than just doing somewhat better would require using x86-64).
This is why seeing a mobo manufacturer posting 1600Mhz as the Clawhammer speed on their products is disturbing; it looks to be the difference between winning and losing against the PIV.
Also remember that a Quake score will show better than typical Hammer results.
Even if Clawhammer does reach 2GHz or a bit more, given what is likely to be the lopsided benchmarking results, AMD may be tempted to play with PR a bit to lean it towards those application/games that get really helped by the memory controller.
The urge to tinker may become overwhelming should 1.6GHz end up being more like it.
The real war between AMD and Intel the next six months is more likely to be in the offices of those creating next year’s benchmarks than anywhere else.
Not saying AMD’s going to do it. I am saying this is something to keep an eye on down the road.
One More Time: What Do the Joes Think?
Intel has a bit less than 80% of the CPU market. AMD has a bit less than 20% of the CPU market. The PowerPC has 2-3%.
While there’s certainly other, bigger factors involved, Apple’s arguments that its processor is “really” faster than the x86 processors haven’t done too well outside of Macdom. Of course, Apple is notorious for making outlandish claims, but I bet most x86ers don’t know that a 1GHz PowerPC really is faster than a 1GHz Athlon or PIII.
Nonetheless, AMD is treading down the PowerPC route with Hammer, despite Apple’s obvious lack of success with it. It, too, will claim that its processor is “really” much faster than it looks due to (what for the average person is) inexplicable reasons.
Not only somewhat faster, but a minimum 70% faster.
The issue is not whether it actually is or not, but whether the Joe Sixpacks and Suits will believe that, true or not.
I think this is going to decide Hammer’s ultimate fate far, far more than any technical consideration. If you leave the average person to his or her own devices, I think knowing or not knowing the actual GHz of the machine will have a big impact on whether they buy AMD or Intel.
The AMD PR strategy counts on the average person not knowing. Some may say PR has worked fine up to now, but I think we have a different ballgame when we’re talking 70% rather than 15-25%.
I think it’s pretty risky to continue counting on Intel pretty much ignoring you under these circumstances and not pointing this little fact out. Especially when your product is so exposed and vulnerable by bucking “conventional wisdom” so much.