Barcelona: AMD Gambles on an Evolutionary, Not Revolutionary, CPU

Barcelona Issue #4: Will four cores make that much difference?

When AMD announced the development - if not yet the availability - of its Phenom quad-core performance-level desktop processors last May, it began its shift toward a bold, if risky, value proposition: performance differences you can feel even if you can't always measure them.

Feelings are respectable target points for selling processors to consumers. For the server market, feelings don't count. IT departments have more quantifiable performance goals they need to meet, and the more those goals can be simplified and incrementalized, the better.

In a very short period of time, the typical data center server has become an eight-core system, either as a four-way dual-core or a two-way quad-core. One of the perceived side effects of AMD being late to market with quad-core is that Intel doubled the processing standards in IT shops, from four to eight cores. As a result, there are quite a few two-way dual-core Opteron systems out there that CIOs are already beginning to perceive as old.

To overcome that problem, AMD needs to offer a very quick fix. So it's playing up this angle: As an IT customer, you've already invested in a platform, and that investment should be long-term. So why not make the most of it by using Barcelona as a refresh, rather than consider Xeon MP as a complete overhaul?

AMD's John Fruehe tries to characterize the alternative as something customers have already rejected. "The customers have been pretty clear: 'If you're going to continually rev the platform like that, we're not going to buy you.' It's easy enough to just say, 'I'm not going to change my style [or] my purchasing habits, because I don't like the cadence that you're changing things at.' So for our customers, it was very, very clear to them that they don't want a whole lot of change in their platforms. They really want stability, because our OEM and system partner customers see that they can leverage that investment over a much longer period of time, allowing them to provide more cost-effective systems. And our end-user customers say, 'Look, you're cutting down the management nightmare that I have to deal with.'

"The cost of the processor is a drop in the bucket," Fruehe continued. "You take a database server, you're going to spend a few hundred dollars on processors, you're going to spend a few thousand dollars on the whole platform, and then after you've been running that for five years with the software costs and the management costs, you're talking about $50,000 for a database server. When you start looking at the price of the processor, it's a pretty insignificant amount of money, so why should the processor drive so much cost change for the customer?"

In a sense, it's a very dangerous argument. By divorcing hardware architecture from responsibility for innovation on the platform, AMD gives itself fewer tools to work with when crafting a value proposition. (Certainly Microsoft won't mind AMD's point of view.) But in so doing, it may also counteract Intel's efforts to convince customers that hardware is the driving force behind all performance innovation whatsoever, as indicated by its recent promotion of the vPro business platform.

The new value proposition goes like this: Enterprises invest in server platforms not to drive hardware innovation, but to run specific software. It's the applications that drive the business need for servers. Along those lines, multiple cores will get you better performance for now, but that's not a guaranteed fact going into the future. With each new multiple of two in the number of simultaneous cores, the server will be boosted less and less as the law of diminishing returns takes over.

"I think so much of this is becoming software-dependent," AMD's Fruehe told BetaNews. "In the old days, it was all about clock speed, and if you could drive the higher clock speed, you had better performance. Then we got into multiple cores, and as we went from single-core to dual-core, you saw silicon drive really 100% of the performance increase. Then as you go from dual-core to quad-core, half of the performance increase was silicon, and the other half is really going to be dependent on software. As you get beyond four cores, and you start to look at eight cores, it's really going to come down to almost exclusively software, the efficiency, the overall scalability of the software is going to have such a huge factor in this that it's really going to become incumbent upon the silicon providers to do a much better job of working with all of the software manufacturers to ensure the best efficiency."

Barcelona Issue #5: Can AMD hold onto the price/performance proposition?

If AMD continues to hold a lead in price/performance even through the toughest competitive storm it has ever faced from Intel, it is in the server division. But with the new "APC" power metric being applied to AMD processors from here on, it will be a little more difficult for buyers to calculate "performance-per-watt" when they're not exactly sure what "per-watt" means.

The comparative price table between Quad-Core Opteron and Quad-Core Xeon MP presents a very telling story of AMD's new value proposition:

A bit of deciphering first: The Opteron 83xx series refers to processors built for four-way (four-processor) systems; the 23xx series are built for two-way. Xeon MPs have an L2 cache per die; QC Opterons have one L2 per core. On Intel systems, of course, memory frequency is determined by the front-side bus; on AMDs, it's determined by HyperTransport, which is all on the CPU.

On the one hand, the new Opterons don't have the clock speed. That's an extremely important issue, because clock speed is a coefficient of overall performance. Everyone expected AMD to come out with the lower speed Opterons first, and for the low-power HE models to have the lowest speed of all. Not everyone expected the first batch to be capped at 2.0 GHz. The 2.33 GHz "SE" models are coming, we're told, probably before the end of the year.

But how far will that range stretch? John Fruehe: "It's 2.3 GHz plus. We're not making a statement about how high it will go, so a 3 GHz SE might be overstating."

On the other hand, notice the price differences. Based on information Intel has given its partners, its Xeon MP prices are likely to remain stable through the first quarter of 2008. With Opterons' prices noticeably way below comparable Xeons' on the low end, Opteron could actually afford to give up a little in total performance and still command the value end of the scale.

But how much is it giving up? ZDNet blogger George Ou received what appear to be leaked benchmark test results from IBM - which may yet be confirmed today - for a Quad-Core Opteron-based System x3455. If Ou's estimates are right, Intel's Tigerton may have almost a one-third advantage in integer processing speed over Barcelona at 1.9 GHz speeds.

That would not be good news for the architecture that's supposed to demonstrate the efficiency of on-chip memory control.

John Fruehe makes the best of what could be a sticky situation by pointing out to BetaNews that AMD's quad-core performs about 50% better than its dual-core Opterons, while Intel's new quad-core performs only about 35% better than its Xeon MP dual-cores. He blames Intel's greater reliance upon its L2 cache. "In floating point, the performance increase is closer to about 10-12%, he added, "and that's because there's a lot of memory bandwidth required to do floating-point. So what you see is scalability differences between dual- and quad[-core] of 10-35% versus 50%. And that's really what the architectural designs bring to the customer, much better scalability, much better efficiency."

It's a tricky message, requiring its listener to make several mental adjustments - between memory frequencies, types of wattage measurements, and price/performance scales. To help the medicine go down, as it were, AMD adds an element of high drama, though this time it's not the image of the lawman standing alone on the street of Western town, deserted but for the shadow of the challenger on the opposite end. This time, it's the image of the underdog who deserves the benefit of the doubt.

"When we bring out a platform, that platform will have longevity and each generation of processor will get significantly better performance with only a small amount of tweaking that has to happen," John Fruehe told us, "because my competitor has a separate memory controller that isn't integrated into the processor. Ultimately, every single time they go to make a small incremental change that may only be a couple of months from the last change, they're literally going out and changing the platform. And we don't have the luxury of being able to do that. We're the smaller guy, so we have to be a little more in tune with what the customers are looking for."

Once again, AMD opts for the evolution, not the revolution. But this time around, the outcome could still change everything.