It has been demonstrated that at least one model of the forthcoming nForce 250 chipset has a working AGP/PCI lock, at least in the socket 754 version.
Then again, it is being shown that at least some models using the earlier nForce 150 chipset are capable of the same if the mobo manufacturer is cooperative about it (don’t assume they are).
Overclocking the Hammer
There’s three core precepts to overclocking a Hammer with a PCI/AGP lock:
Don’t Have Unrealistic Expectations Removing one bottleneck doesn’t mean the sky is the limit. It just means you run into another bottleneck a bit further down the road.
If you expect 50% overclocks from an Athlon 64 3000, or 3GHz from the current generation of Hammers with no more than a fancy fan and a chunk of copper, do not drive or operate heavy financial equipment (like credit cards) until your high wears off.
Sorry, but those are unrealistic expectations. These are 130nm chips with rated speeds towards the top end of the spectrum and approaching the limits of what 130nm technology can do.
Our overclockers’ database is indicating that under moderate overclocking conditions, a speed of 2.4GHz-ish is typical, perhaps a bit less than that for the Athlon 64 3000.
The next stepping will probably do a few percentage points better than that, so we’re probably looking at 2.5GHz-ish, maybe closer to 2.6GHz.
That’s not bad at all under the cirumstances, but if that’s not good enough for you, wait until 90nm, which should eventually make 3GHz possible.
An Old Bottleneck: Memory A working PCI/AGP lock doesn’t do a thing for memory. Unless you have the latest super-rated with relaxed settings RAM, odds are it won’t run at 250MHz or better. At the least, be aware that that may be the reason why you can’t get past 220MHz or 230MHz.
Fortunately, for most Hammer boards, this is a very avoidable bottleneck. All you need to do is set memory to run slower than FSB.
Unfortunately, there’s more than a few memory fetishists out there who would suggest, well, just about anything before doing that. Like people with other kinds of fetishes, they can not get satisfaction from anything but their fetish, so you find them rather often sacrificing performance for a high memory benchmark score.
This is not good.
Increased memory speed has a low correlation to increased performance; it’s a very small fraction of any increase in bandwidth. On the other hand, increasing CPU speed has a fairly high correlation to increased performance, outside of bottlenecked video apps, usually well more than half the percentage increase in speed.
If you have pretty good RAM already, the best approach is to see how far you get running the computer at a 1:1 ratio, then see how far you get running it at a slower ratio, and finally compare the two to see which is better in REAL apps/games.
By running the memory as slowly as possible, you’ll get a very good idea how far the CPU will go without memory limitations getting in the way. If your CPU will only go to 240MHz FSB when the memory is running at 160MHz, that CPU isn’t going to run any faster if you buy 250MHz memory.
Perhaps more importantly, if you find yourself in a situation where the CPU does 250MHz and the memory you already have does 200MHz at fast settings, you can then ask yourself, “Is it worth $150 or $200 on newer RAM to get maybe an average 2% increase in performance?”
The answer will probably be “No” most of the time.
A New Bottleneck: HyperTransport When you overclock your system, you end up overclocking HyperTransport, and past a certain point, it doesn’t much like that. Again, you should be aware that this could be the reason why your system won’t go past a certain point.
This is another avoidable bottleneck. Most Hammer motherboards let you run HyperTransport at less than default speed (i.e., 3X rather 4X, or, shortly, 4X rather than 5X).
Again, you should test both ways to see what works best for your particular configuration.
The nForce 250 chipset will become the overclockers’s chipset of choice. It will be built to handle higher speeds than the earlier 150, and its PCI/AGP lock gives it a killer edge over the lockless Via boards.
When it comes out, it will come out in both socket 754 and 939 versions. As we’ve said many times before, we do not recommend buying a socket 754 system, mostly because AMD’s current intentions are to make it the equivalent of the Celeron line within the next six months.
Socket 939 solutions using this chipset ought to be available in late May/early June. These will initially come in an AGP-only configuration. A few months later, we’ll see PCI Express for the video slot.
It is not conclusively known at this point whether socket 939 CPUs will initially be strictly high-end or not. If they are, that will preclude the vast majority of overclockers from buying until they are.
If affordable models do come out, or you’re hellbent on buying one anyway, when you buy will probably turn on what your video situation is. If you’ve bought a video card fairly recently and plan on it sticking around for another year to two, you’ll probably want the AGP model.
If you plan on replacing your video card soon and expect to hang on to it for a while, you’ll probably be better off waiting for the PCI Express motherboards to come out and buying a PCI Express version of the video card.
If you’re looking for a DDR2 system, that probably won’t show up on the Hammer side until 12-18 months from now. It probably will require a new socket, so don’t expect any Hammers you can buy in 2004 to work in the more desirable 2005 models. We don’t think DDR2 is a must-have feature for anyone with any urgency to upgrade, and alone isn’t a reason to wait an extra six-twelve months.
If you haven’t upgraded by early 2005, then only at that point might it become a reason to wait a bit more.
Is there any particularly compelling reason to buy such a system right away? Not unless you’re replacing a fairly old system (i.e. pre-Thunderbird, Thunderbird-class, maybe low-end Palomino class). They won’t be bad, but won’t offer any huge performance increase over relatively recent socket A systems. These systems will start to become compelling to most Thoroughbred-class socket A users only when 90nm CPUs become available, and maybe not even then right away.
Obviously, different people have different boiling points. If you already think you won’t upgrade for another year or more, don’t let this change your mind.
On the other hand, if you have a Barton running at 2700MHz and you’re foaming at the mouth for a new system NOW, maybe this is a good time to work on your patience, especially if your ambitions give your wallet ruptures.
We say all this knowing full well that the average AMDer is not going to Hammer until 2005, no matter what AMD is likely to do with its positioning and prices. Right now, Hammer sales are a little trickle of activity. If AMD makes the right moves, we’ll have a rather bigger trickle, if not, a slightly bigger trickle will continue.