Today is desktop Hammer day. It’s been a long wait.
Today will also be the coming-out day for the PIV EE. That has been a very short wait. 🙂
In both cases, though, the conclusions are the same. From the overclocker perspective, both companies are offering products that offer a little more performance than current offerings for a lot more money.
They are both products that represent the last gasp of 130nm technology. They are both products that, for different reasons, will be overtaken (cross your fingers) by newer technologies from these companies within a year. So will the current products, but the newbies won’t last any longer than the old fogies.
High price, little gain, short lifespan. These are hardly recommendations.
Today, we’re going to look at the performance of these uberchips from a different perspective: that of the CPU senior citizens.
We’re going to analyze some benchmarking data a bit to show how much gain you’d get from these budget busters compared to what you can do today with much cheaper, easily overclocked, current CPUs.
What We Have Today
The current overclocking favorites are one of two types of systems:
1) A TBredB/Barton system for most, or
2) A 2.4GHz PIV Springdale/Canterwood system.
Both systems can be significantly overclocked with virtually no effort. It is pretty safe to say that most recent TBred/Barton overclocked systems hit the performance level of an XP3200+ processor running with default settings.
It is also pretty safe to say that most recent overclocked PIV systems hit the performance level of a 3.2GHz PIV running with default settings. In fact, we’re being on the conservative side compared to what most serious overclockers try to do, but that’s OK, for this, we want to establish a baseline for a pretty painless overclock using today’s equipment.
The core parts of a Barton/full nForce2/dual-channel 512Mb system costs about $350 USD. The PIV equivalent system would cost about $500 USD.
In contrast, an Athlon 3200+ system with single-channel memory looks like it will cost about $700 USD, while the Intel PIV EE should cost a bit more than $1,000USD, and an Athlon FX system more like $1,200USD.
What do you get extra for your extra money?
What we did was analyze the benchmarking data from this review (and please, give it a long look) and normalized the performance of the CPU using the performance of the AthlonXP 3200+ as the standard with a score of 100.
We then converted the other scores to that scale to show how much better (or worse) they did than the 3200+. If a CPU gets a score of 110, that means it was 10% faster than the 3200+ at that particular benchmark. If a CPU got a 90, that means it was 10% slower than the 3200+.
The results were color-coded into three categories. The 3200+ and the PIV 3.2GHz were coded purple, so you could see how current overclocking solutions do.
The results from the Athlon 64, Athlon FX and Intel PIV EE were colored light blue so you could easily see how well the new chips did at default speeds.
Any overclocking results from the review from the 64, FX or EE were colored black, so you could see how much overclocking helped. Please note that these chips aren’t going to overclock a lot without dire means being taken, and the results are probably quite representative of you would might expect from the average overclocking effort.
Finally, we made sure all graphs started at zero, so you’d get a clear visual picture of the actual degree of differences (or lack thereof).
A Few Words First On the Uberchips
The AMD chips do OK. Viewed strictly from a performance standard, they’re pretty competitive against the Intel chips, and the picture looks prettier than looked likely even just two months ago. That’s because AMD made some last-minute changes and corrections.
First, they created the Opteron that Isn’t An Opteron at the last minute when they realized single-channel wasn’t going to quite cut it against dual-channel PIV systems. They should have realized that a lot sooner, and this 939/940 pin issue just shoots themselves in the foot, but it was good that they did it.
Second, they downgraded the expected PR on the Athlon64 a few hundred points. Calling a 2GHz Athlon64 the equivalent of a 3.2GHz PIV is at least defensible, calling it the equivalent of a 3.4GHz wouldn’t have been.
Finally, they got the core frequency of at least the FX up to 2.2GHz (with the same for the 64 expected to follow fairly shortly). That helps. Outside of what appears to be a weak SSE2 implementation, there’s nothing wrong with the Hammer design. On the whole, it’s pretty good and should even be better in 90nm.
However, as we’ve said all along, this chip needs x86-64 to jump substantially ahead of Intel. It is no Intel killer or even wounder in 32-bit; it’s roughly equivalent, better in some things, worse in others.
The Intel EE? It’s obviously a last minute stand-in, and gets off a few good lines, but it’s beginning to look like the reason why it’s in there is because the star broke his leg. Prescott has been delayed another month until early December.
Remember, this chip was supposed to be out last spring. Intel doesn’t normally do December launches; they either launch a CPU in September or October to catch Christmas, or they wait until mid-January.
It’s getting harder to believe everything is OK.
In any event, the EE gives little more bang for the buck in 32-bit than the FX, and can’t even wave 64-bit someday in front of people like the FX can.
Let’s see how the CPUs did in SuperPi:
It takes overclocking for even the best of the uberCPUs to beat an XP by more than 20%, or a PIV by more than 10%.
Another calculation-intense benchmarking program is ScienceMark:
The little old XP beats the PIV in this one, and really doesn’t get beaten until you start overclocking $800 CPUs.
I don’t know why people use 3DMark to benchmark CPUs; the only change that really moves the numbers is a better video card. Nonetheless:
The XP is a bit disadvantaged, but the PIV loses by little.
There’s even less of a difference in 3DMark 2003:
If you want to shine on this benchmark, spend your money on a Radeon 9800XT next month. You ought to be able to buy a PIV system AND that little more than the price of the FX alone.
The PIV beats everything. You’d have to overclock the EE or FX to beat a $170 processor.
No crushing blow by the uberprocessors here.
This one’s almost embarrassing.
The PIV meets or beats everything else, and Hammer can’t put his scrappy ancestor away.
This time, the EE puts the PIV (somewhat) in its place, but Grandpa XP is still getting in a few good licks against his kids.
By this point, you ought to be saying “Uggh” at what you’re seeing.
3D Studio Max
This benchmark just doesn’t like AMD. It doesn’t like newcomers of any sort too much, either.
Neither does this one. You have to overclock the new ones to get some margin of victory over the golden oldies.
This one likes AMD, but it respects the elders, too.
Finally, a big difference! Well . . .
Not necessarily that much of a difference.
This game really likes Hammer, but . . . .
. . . maybe not quite as much as you think.
A Weird Omission
We didn’t convert a few of the benchmarks found at x86-secrets, usually because they were generally unknown or just never made much sense to us.
One benchmark we didn’t do was one apparently provided by AMD as a “sample” 64-bit application. It’s called Minigzip, and, well, the benchmarks results were just weird.
As you might expect, in 64-bit mode, it annihilates the PIV, but nowhere nearly as bad as it annihilates itself in 32-bit. In theory, under perfect conditions, a 64-bit processor might be able to do a task in half the time of a 32-bit computer, but not better than that, and that’s what this test shows.
Update 9/24/03: We got an explanation on this one from x86-secrets. Apparently, minigzip source code is modded by AMD to use *ONLY* packets of 64 bits, even
in 32 bits mode. So when the program is in 32-bit mode, the CPU needs to split the integer from 64 bits to
2x 32 bits. Therefore, 32 bit CPUs take more than twice as long not only to run two 32-bit packs rather than one 64-bit pack, but to also split the packet.
You can see it here. Sorry, too hard to believe.
One Simple Thought
If the cheapy processor is pretty (as in the XP) or very (as in the PIV) similiar to the uberchip, and they’re running at the same speed, just how much faster could they be?
Yes, when that Hammer memory controller can do some good, it can do a lot of good. But that doesn’t happen all the time, and even when it does, results vary.
I must confess. While perusing this, I found myself rooting for the XP. For a sub-$100 CPU (and for TBredB owners, maybe more like a $50 processor); it gave one hell of a fight against challengers that cost two, five, even ten times as much.
The PIV was the not-so-under-underdog in this battle, it costs much more than the XP, but it just didn’t lose honorably; it gave a few beatings, too.
On the other hand, the uberchips were . . . underwhelming. Yes, they won (usually), but much like the New York Yankees nudging out the Montreal Expos. Given the Yankees’ payroll, they ought to win.
These aren’t perfect measurements. On the one hand, many overclockers machines probably do somewhat better (either by higher MHz or more FSB bandwidth) than the 3200+/3.2. On the other hand, most of the measurements didn’t include overclocked uberchip results.
So the real differences between current overclocked systems and the ubersystems are probably somewhat smaller if you compare O/Cd cheapies to default ubers, and a bit bigger than that if you compared more O/Cd cheapies to O/Cd ubers.
Nonetheless, the ubers don’t kill the cheapies by any means. Even a 20% difference between the cheapies and the uberchips from the same company is pretty rare (it’s possible when we see more games that they’ll paint a somewhat different picture), and if you factor price into the picture at all, it’s no contest.
It’s not that these are bad processors. If they cost $50 or even $100 more than the oldies, they’d be worth it to at least some. But they don’t cost $50 to $100 more. They cost $250 or $350 or $600 or even $750 more.
These CPUs are God’s way of telling you you have too much money. 🙂
Get Enthused Later
There are good times to get enthused, and there are bad times to get enthused.
A good time to get enthused are those points where you get a lot for your money. Nobody at any website had to say a word to pump you up about $50 XP1700+s. 🙂
This is not a good time to get enthused.
Most of the problems I have with the first generation of Hammers ought to go away in the second generation. The 90nm process should provide a lot more overclocking headroom, it will be clear what CPU goes into what socket, memory controllers should be ready for DDR2, there may be some real x86-64 software out there and prices should be lower. A lot lower.
Get enthused then.
Intel is a bit more doubtful at the moment. We’ll have to see how Prescott progresses, or doesn’t. That will likely have a very big impact even on AMD buyers. If Intel can get cheap Prescotts out in February or shortly thereafter, that ought to do a number on Hammer prices. If they can’t, odds are FX/64 prices will stay high.
But whether you like AMD or Intel, don’t let yourself be wound up and feel compelled to buy these uberchips anytime soon. They’re just not worth it.