Barcelona: AMD Unveils New Power Metric for CPUs

Throughout the past five years, AMD has stated Intel has been unfair in the way it measures its CPUs' power consumption, and presents those numbers to the public. But it hasn't come up with an alternative method...until today. Beginning with the Barcelona generation of quad-core CPUs, AMD will utilize what it calls average CPU power (ACP), which unlike Thermal Design Point (TDP) is not based on how much power is required to cool a chip, but rather to run it.

"One of the biggest problems that our customers have come to us with is, 'I want more performance but I'm really tapped out. I'm at the limit for the amount of power I can pull into the data center,"' relates John Fruehe, AMD's worldwide marketing development manager for server and workstation products, in an interview with BetaNews. "They may have empty space left in their racks, but their facilities guys have told them, 'No more. You aren't getting any more power we can't provide any more.' So in order to put in new, more powerful servers, they have to find somewhere else in the data center they can go find something to unplug in order to make room for it. And we don't think you should have to do that."

AMD describes its ACP metric as based on tests of the average power draw on processors that are given a specific, measured high-utilization workload, such as a SPEC benchmark that can be balanced across cores.

AMD provided BetaNews with a picture of one of the ordinary power meters it actually uses in its ACP testing. The device measures power at the processor in volts, amps, and watts, for the duration of the given workload.

"In a perfect world, if I had the best opportunity, I would tell the customer, don't ever look at a manufacturer's spec sheet, don't ever look on the Web to find out what the power consumption is," remarked Fruehe. "Plug in a power meter at the wall - it costs about thirty bucks at Home Depot. That is absolutely the best way for a customer to determine the power draw: Get your configuration, plug it into a power meter, run your application, it'll tell you what the actual power draw is.

"Now, only about a quarter of the customers actually do that," he continued. "The other three quarters rely on the manufacturer's specifications to determine what the power consumption is. So what we heard from customers - and you probably heard it yourself - is, TDP is really a terrible way of measuring power. But, for instance, our processors that have a 95 W TDP never even get close to that. It's probably more like about the 50 to 60 range, that you're going to see an average power consumption."

Presumably, Intel's processors would also run cooler than the TDP. But don't take our word for it, Fruehe says - go buy a power meter and find out for yourself.

AMD's new Quad-Core Opterons are rated on the new ACP scale at 75 W at the standard performance series, 55 W for the power-saving HE series. This is in comparison to 95 W and 68 W TDP for both series, respectively. The high-performance SE series will be available in the fourth quarter, though we don't know yet what its ACP rating will be.

Fruehe said AMD came up with these numbers by taking samples from the "hottest bins" from its fabrication facility - the units which are likely to slide into SE territory, then load those samples with high utilization benchmark workloads. While the CPUs were busy, their power draw was sampled twice per second using a common power meter. "It's really still an overly conservative number," he believes, "but it's going to be much closer to what customers would see in real-life scenarios. It's probably still going to be a little high, but that's okay. I'd rather have it be a little high than a little low. It gives customers a better way of measuring what the actual power consumption will be.

"The complaint I used to get from customers all the time is, I go do the math on how many servers I can fit in a rack, and then I go and I pull way too much power to the rack, or I'm not filling the rack all the way up to the top," Fruehe continued, "so I'm either wasting valuable space in the data center or I'm over-planning on power capacity. Neither one of those things is a good option."

At some point, someone (outside of Intel) will want to try AMD's ACP testing methodology on Intel's quad-core CPUs. Will they be able to replicate AMD's methods to yield a comparable result?

"We are going to publish a white paper that goes through the exact methodology, so if somebody else wanted to follow it, they'd be more than able to," Fruehe said. And then he paused for a bit and realized, no, not quite. "If a publication were going to actually do a comparison, I would recommend that they don't even try to figure out how they replicate ACP," he made clear. "You've got to take into consideration that [Intel has] integrated memory controllers. So if you're really going to do the true apples-to-apples comparison, you'd have to not only look at Intel's processors but also their memory controllers."

Once you take into account all those other Intel parts that drain power, he said, AMD's comparable processors should outperform them by 10% in a worst-case scenario, 70% in the best-case. We may be interested in taking that $30 trip to Home Depot to find out for ourselves.