This is basically a part II to the blog post I put up earlier today regarding how Moore's Law could finally be failing us. The premise of that theory, put forth this morning by Uptime Institute chief analyst Bruce Taylor, is basically that, even though the performance of chips is doubling while the parts costs are halving, the parts cost no longer tells the complete story. When Moore's Law first went into effect, running a chip (or all your chips in all your systems) at full tilt was a non-issue. Power, cooling capacity, and space were neither scare nor that expensive. Now they are and so, the real cost of owning those chips isn't necessarily halving.
During the second part of this morning's panel, Taylor (whose credibility went out the window for me towards the end of the event..... more on that coming) asked each of the participants to talk a little bit about what each of them are doing in their technologies to improve the power efficiency of a datacenter. Here is what I heard:
AMD: Senior Fellow Samuel Naffziger said AMD has a huge funnel of technology going into its chips but that one of the big challenges is in making sure all that technology gets used. For example, he said, "As much sophistication that we put in, a lot of it goes untapped. We have 8 temperature sensors per core and their information is available digitally, but no one is really taking full advantage of it."
Naffiziger also said that "multicore technology is a great lever for power efficiency but that it comes with downsides too. For example, the software community has to learn how to use [those multiple cores] better." He noted how, in the new Barcelona chips, the ability to control each of the four cores is right on the die. With the right automation tools, you can turn on portions of the chip to support certain computations as those computations come up, and then turn them off. Additionally, AMD sees its approach to chip design where it has been pulling system componentry right into the processor -- dating back to floating point processors in the 80s (even Intel does this), then (more recently) memory controllers, and now graphics (with the acquisition of ATI) -- as one that will lead to greater power efficiency. Naffiziger made it clear that when it comes to that sort of integration, AMD isn't done yet (my guess, networking is next).
AMD's Opteron product marketing manager Brent Kerby added that what AMD is doing outside the processor matters as well. For example, with the increased integration of componentry into processors, there's more emphasis on memory (another reason networking is probably next. Like graphics, it needs memory too). Said Kerby, "Since we integrate memory controllers, we get to pick what memory we use and can focus on low power memory, thereby keeping power down within actual logic design."
EMC: Senior marketing manager Dick Sullivan talked about the practical things we could be doing now to save energy. For example, he described a scenario (might have been a customer) where there were a number of older systems with 18 GB drives that could be consolidated into fewer systems with fewer but larger drives. Fewer spindles equates to less energy. Sullivan said that tiering can also yield a savings. "Look at Tier 1 storage that hasn't been touched in 90 days" said Sullivan. "If it doesn't need to be there, move it to a lower tier of storage that isn't just more economical, but uses less energy."
Regarding the "other" form of consolidation that EMC espouses (virtualization, VMware is a subsidiary), Sullivan said that "virtualization will not only eliminate some energy [consumption] and space [requirements], we will improve the utilization within virtual machines so that power management becomes part of that story."
HP: Going back to the temperature sensors in Opterons, HP blade system infrastructure technologist Ken Baker said that his company's solutions do in fact leverage that information. He talked about how, with older gear, you had processors that ran at one speed and fans that ran flat out but that now, HP modulates that all the time. He said that "We modulate that all the time..... intelligenty monitoring the the fan speed to give it just enough air to keep the processor healthy." Baker was the first of the bunch to draw a connection between the sensor data available from chips like Opterons and overall modulation of a datacenter's air conditioning systems.