AMD now has a webpage which shows Opteron MTOPP scores.
Some explanation of this is in order. The MTOPP is not a benchmark. It is a calculation of theoretical performance the U.S. Commerce Department
makes to figure out whether or not certain countries can get certain computers. It was created ten years ago and is meant as a rough-and-ready guide more than anything else.
For instance, it makes no difference to the formula whether you have a PIV Willamette or a hyper-threaded Northwood. As you’ll see, it holds up pretty well given its age, but it’s hardly a precise measurement.
The last time this calculation was (ab)used as a benchmark; it was used to justify Apple’s claim a few years ago that the Mac was a “supercomputer” (which was solely due to an old definition of the term by the Commerce Department which was updated a few months after the claim).
It would hardly be fair to look upon these scores without seeing Intel’s scores, too, which are on this page:
Puts matters in a bit different light, doesn’t it?
Perhaps more to the point, look at how much these processors cost (or will cost):
(PIV prices are based on Newegg’s prices for retail processors; Opteron price is estimated based on current official’
AMD pricing for 1.8GHz and AMD’s declared official price on Opteron 246. Opteron 1XX series CPU are being priced about
$250 less than Opteron 2XX processors at the high end).
I know AMD needs more money, but this hardly looks like a bargain, even by Intel standards.
Predicting the Future
There’s one good thing about this formula; it’s pretty easy to calculate the scores of future processors in the same line.
If you look at the numbers on the Intel page, you’ll see that every extra 200MHz speed increase on the PIV gets you another 533.33 MTOPPS.
If you look at the numbers on the AMD page, you’ll see that every 200MHz speed increase for the Opteron gets you another 717 MTOPPS.
So, if you apply the formula to AMD and Intel processors expected in the next year, you get this:
Again, the formula seems pretty insensitive to internal improvements. Adding another 256K cache to the Willamette didn’t budge the MTOPP calculation one bit.
Let’s Get Realistic
The Commerce Department formula basically says that an Opteron MHz is worth about 1.34 Intel MHz.
As a measurement of raw CPU power, that’s not too bad a calculation at all, and it’s good enough for what the Commerce Department, uses it for, but they aren’t interested in minor differences.
Based on the benchmarking we already have on Opteron, and factoring in the difference between Opteron and the “mainstream” Athlon64 (and assuming a bit more tweaking in the meantime), a ballpark estimate of 1 Athlon64 MHz = 1.45-1.50 Intel MHz will probably be pretty close. Let’s take the middle of that range and see how the Athlon64 fares:
That looks better than the MTOPP scores. AMD is still trailing, but not by much.
Unfortunately, if you use the expected PR ratings; the results aren’t pretty:
Can anything make this ugly duckling a swan?
The 64-Bit Fairy Godmother
AMD estimates that running 64-bit apps will increase general performance 15-20%. Let’s presume a 15% increase, modify the Athlon64 numbers accordingly, and see what we get:
It’s fun being green, isn’t it?
Substitute PR ratings, and PR starts looking reasonable, even conservative at higher levels:
Fairy Godmothers Have One Fault
They’re not always around when you need them.
For AMD, the fairy godmothers are the software writers. If they help Cinderhammer out, and sooner rather than later, AMD has a nice coach.
If they don’t, unless the backup godmother, IBM, adds magic to the SOI sauce, AMD will have a pumpkin on its hands.
Like the Commerce Department’s formula, this are rough estimates. They obviously aren’t perfect, but they don’t have to be. They’re good enough to illustrate the moral of the story.
Hammer needs x86-64 to compete against Intel, to justify its own PR ratings.
It would take a last minute miracle for Athlon 64 numbers to jump up the PR levels being suggested for it.
On the other hand, presuming that 15% increase in speed is reasonable, x86-64 makes PR looks OK.
Will the fairy godmothers show up in time?