Opteron Analysis

All You Need To Know

Overclockers, this is all you need to know about the Opteron for the moment.

Want a machine that would beat the Opteron in most of the workstation tests conducted so far? Buy one of those $60-70 1700+ JIUHBs, get an nForce2 board and a good stick of RAM, and crank it up to a reasonably high degree.

That’s all you have to do.

You can buy all that and a Radeon 9800 Pro for about the same price as an Opteron 1.8GHz all by itself.

Come back in six-nine months and see if things have gotten any better, and realistically expect to wait another six months before you can render a final verdict. In the meantime, hang on to your socket A system.

What You Ought To Know

The benchmarks of the Opteron over at Tom’s Hardware continue to show that current Hammers are light on raw computational firepower.

This time, though, people can’t yell, “Beta, beta, beta.” This is a product that AMD sent over to be tested. The CPU was made the fourth week of 2003, not much later than the XBitLabs sample. After adjusting for the increased processor speed (1.8 vs. 1.6), the synthetic processor scores are pretty much the same (a few percentage points higher in Sandra).

It does some things well. If you need a Linux database server, and/or you really need more than 4Gb address space, the Opteron does well, but even there, it takes at least one and usually two of them running 64-bit mode for it to outclass a high-speed Athlon at 32-bit.

It just doesn’t do very well as a workstation chip, even against its predecessors, much less the newest Intel chips and chipsets. Also remember that this processor has SSE2, so you can’t blame losses to Intel in certain apps on that anymore.

The core problem is that AMD can’t spot Intel 70% more CPU cycles and expect to win with this chip, and that’s what they’re basically claiming. If you adapt Athlon64 PR methodology to this chip, a 1.8 is a 3100+. It just isn’t except when quick, frequent, short memory accesses (like database requests) are of prime importance, or for those apps in which Athlons already had a considerable advantage over PIVs. A more accurate PR rating (at least in 32-bit) would probably be 2600-2700+.

Had this chip come out a year ago, it would have looked great, but it seems like AMD aimed at a fixed target, and the target has moved.

A dual channel here, an 800MHz bus there, a little memory acceleration and an occasional hand from hyperthreading, Intel has made 3GHz more than it was a year ago. None of these are revolutionary changes, or add more than a few percentage points to performance, but they add up.

Cut the gap between comparable processors to 40-50%, and then Hammer has a fighting chance against the new Northwoods. There’s nothing wrong with the chip, just its clock speed and the bloated PR rating it has. It needs a major speed boost (at least 20%, and really more) for it to become really competitive against Intel’s latest offerings.

Or, it needs 64-bit to keep up to justify the PR.

Servers Aren’t Everything

Sure, there’s much to be said for the Opteron as a server CPU.

But just how big is the market for server chips?

Here’s a hard, cold number: 9 million a year. That’s how many server chips get sold a year.

What’s the market for desktop chips? Right now, about 130 million a year.

You make good money from server chips, but you pay your bills with desktop chips. Hammer can’t just do well as a server chip.

Just like we never liked the PIV at 180nm, we don’t think it makes much sense to buy a 130nm Hammer.

Unless AMD can get 130nm Hammers to run at clock speeds higher than what Athlons can do today, we think that for typical use, they just won’t be price/performance competitive against either Athlons or PIVs (presuming pricing somewhat similiar to PIVs). We don’t believe that enough of a performance gap will open for fairly current socket A users to shift to a new platform until 90nm.

We hope that the 90nm version will be able to clock higher relative to Intel chips than what we see now (or expect to see the rest of the year). Maybe this processor can spot Intel 40%-50% in clock cycles and win, but not 70%.

More likely, though, this processor is going to need a 64-bit OS to be seriously competitive with a 70% clock handicap, and we won’t see Windows x86-64 (at least finalized) until early 2004.

Someone Ahead Of The Curve

I got this email this morning, which should provide some food for thought:

Let’s forget about the server performance of the Opteron chips for a moment. Sure, it is exceptional, but to the majority of ‘us’ enthusiasts server performance would have little relevance to our everyday computing. Credit where credit is due, AMD did a great job, but that is not why I’m writing to you.

Workstation apps (3D, video/music encoding, etc) does, and that is what worries me about Opteron. Take a look at the single 1.8GHz Opteron score against that of a 3GHz P4. It gets thoroughly trounced in basically every benchmark.

Now, correct me if I’m wrong, but wouldn’t a 1.8GHz A64 be identical to a 1.8GHz Opteron except without SMP and DC-DDR support?

So a 1.8GHz A64 certainly won’t be any *faster* than a 1.8GHz Opteron, correct?

If that is so, how exactly is a 2GHz 3400+ A64 supposed to compete with a 3.2GHz P4 (lets not even talk about Prescott yet…) in the desktop market, which is obviously a lot closer to the workstation rather than server side of things as far as programs/applications are concerned. x86-64 might help a *bit*, but it’s not gonna be a huge quantum leap. And besides, desktop x86-64 apps will be rare at launch and stay that way for quite some time…

Many people criticised Xbit for posting the A64 preview because it was using beta silicon, but it seems, when comparing THGs to Xbits results, that A64 really is struggling when it comes to desktop/workstation applications.

Your opinion on this would be greatly appreciated.

See, I’m not the only crazy one. 🙂

Right now, the best answer I can give is “AMD is basing its PR on use of a 64-bit OS.”

Does anyone have a better one?

Does AMD?

Email Ed