UPDATE: Intel attacked over Xeon 5500’s power-efficiency

Maynard (MS) – Intel’s Xeon 5500 processor, formerly code-named Nehalem-EP, is all about performance and power-efficiency, but its rivals disagree. At least for SiCortex, a company that claims to offer one of the most power-efficient processors for HPC applications, says that Intel simply “ignores” the need for improved energy efficiency. Is SiCortex right?

UPDATED:  March 31, 2009 – 7:15pm CSTIntel has issued an official response to this article. See TG Daily’s publication of Ronak Singhal’s article (Intel’s Digital Enterprise Group):  Intel responds officially to Nehalem power concerns.SiCortex’ claims appear to come down to the way power-efficiency is defined. Is it an overall substantial drop in power consumption or is it simply more processing power that can be squeezed from the same amount of energy? In Intel’s view, Nehalem is vastly more powerful for, according to its current technical documents, 15-20% less power. Let’s look a bit closer.

The company says the chip is about 2.25 times faster than its 5400-series predecessors and offers 3.5x the data bandwidth. At the same time it runs at 50% less idle power. In direct comparison, the 5400 chips ran at a thermal design power of 150, 120, 80, 70, 50 and 40 watts, depending on the version, while the new 5500 CPUs are rated at TDPs of 130, 95, 80, 60 and 38 watts.

At least from that perspective, there may be a clear decrease in power for now. But realistically, there may actually be slight increases, if predecessor chips will be swapped for their direct successors. The 150 watt 5400 “Extreme” chip has yet to become a successor, which will be an 8-core version later this year, which is expected to run at 150 watts or more. Realistically, the 5500 lineup will not see a decrease in power consumption when it is complete. In most direct and fair comparisons, Nehalem may actually be slightly more power hungry – despite the fact that the current 5500 flagship outruns the previous high-end 5400 model.

So, is SiCortex right? They are right in the case of server and workstation manufacturers, and they may be right if large corporations are aiming for a 1:1 swap of server units and not just decrease the number of processors to maintain the same computational power. Common sense suggests that computational power will increase at a level that makes financial sense and no one can predict if the increased power-efficiency of Nehalem-EP will be able to offset the greater power consumption in exchange for the targeted increase in computational performance in each case.

That, in effect, will be at the discretion of each organization running Nehalem-EP servers. That scenario has been the case for some time and both Intel and AMD are stuck in a thermal power envelope that was defined back in 2006 – low power of 65 watts or less, a mid-range of around 80-100 watts and a high-end of more than 120 watts. To Intel’s defense, the new chip provides tools to limit the power consumption of the CPU: For example, there is a new Node Manager, which allows IT staff to cap the power consumption the CPU. 
However, SiCortex also says that Intel has reached the “practical limit of processor clock speed,” and the common expectation that Quick Path Interconnect (QPI) does not scale. As a result, Nehalem-EP servers may need “external I/O interface chips and commodity switches to communicate, impacting performance, energy consumption and cost.”

As revolutionary as the new Xeon 5500 may sound, it appears that it is not the great solution that will solve problems for datacenter designers and all others who have to find the best compromise between processing capability and power consumption. Even this new chip and its improved power efficiency will require a lot of math and testing to figure out which data center will support which and how many processors.

Alternate Take
by Managing Editor, Rick C. Hodgin

I have to take exception to both the Gruener and SiCortex claims that Intel’s newest chips actually consume more power. And I have a chart to prove it.

Gruener’s viewpoint of how the 5400 and 5500 models line up is in the left two columns. My view of how the 5400 and 5500 models line up is in the right two columns. The differing viewpoints on power consumption can be valid you look at a literal model-number to model-number comparison. However, Nehalem delivers such an increase in performance that a relative comparison of abilities seems more appropriate. This is fundamentally the difference in view between Gruener/SiCortex and Hodgin/Intel.

Gruener’s View Hodgin’s View
5400 5500 / Nehalem 5400 5500 / Nehalem
40 38 40 Does not exist,
or possibly 38
by default
50 60 50 38
70 80 70 60
120 130 120 95
150 Future
150 Watt chip –
To be introduced
later this year
150 130
Does not exist Future
150 Watt chip –
To be introduced
later this year

The idea is simple, on Gruener’s side the parts compare by literal number. However, Hodgin believes that because Nehalem performs so much better, that the previous generation’s next-higher model is more on par with Nehalem models. As such, the process begins with the second 5400 part at 50 watts comparing to the first 5500 part at 38 watts.

So what do you think? Is Nehalem in a new class all by itself where a literal model-number to model-number comparison is repugnant? Or do the greater power consumption numbers of corresponding Nehalem models ring Intel’s claims as false advertising?