INTEL RELEASED FIVE new Xeons yesterday, collectively known as the E-step parts, all of which promise greater efficiency than their predecessors. There are four quads and a dual core, Harpertowns and Wolfdales respectively.

The overarching theme here is a mid-life kicker for the server line, nothing groundbreaking. The E-step parts slot in above the D-step variants, with the X5470 being a 3.33GHz/1333/120W part supplanting the X5460 which runs at 3.16GHz, although that should be obvious from the easy-to-read naming scheme. Same for the X5492 which runs at 3.4/1600/150W, it comes in above the old X5482 @ 3.2GHz. To make matters a bit more confusing, the third E-step is also an X5482, but it pulls at 120W instead of 150W.

To clarify the naming scheme, on a 1333MHz bus, quad core Harpertowns are prefixed with an X if they are 120W, E if they are 80W. They all end in 0 unless you count the 2.0GHz model which is the E5405 instead of the E5400 for some reason.

The Harpertowns on a 1600 FSB all end in a 2, and they are prefixed with an E at 80W. The 120W parts are also X- series, unless they are 150W, which are also X-. Thus, the X5482 going from 150W D-step to a 120W E-step doesn't change the name.

To make matters funnier, the dual-core parts use X- to denote 80W, not E like the quads. There seems to be no 'normal' wattage dual cores, just eXXXtreme and low-voltage L- parts, but all seem to use the last digit, either 0 or 2 rather willy-nilly. Still following? The dual core Wolfdale released today is the X5270, a 3.50/1333/80W part. This could quite possibly be the best gaming CPU out there right now, FPS addicts take note.

The last new one is the heretofore-hinted-at, low-voltage part. It is a Harpertown L5430, 2.66/1333/50W chip, and it slots in above the 2.50GHz L5420. We told you about most of the predecessor parts here.

The new parts may seem to be nothing new under the sun, and that is sort of the intention. They are just tweaks to the older models to save a bit of power and to clock higher. Voltages didn't change, die size didn't either, nor did performance per clock. It is just a little better, and allows for another bin or two on the top end. This translates to a bit more efficiency on the lower-speed parts.

That all means better yields, better parts for the end user, and in general, more profit for Intel because the top-bin parts command top- end prices. That said, the X5492 costs $1493, X5470 comes in at $1386, and the L5430 is only $562. The dual core X5270 slots in at $1172 while the newly 120W X5482 doesn't change. Even without a price drop you probably feel quite good that you are saving almost $9.27 in electricity every 4.3 years with the new X5482, and that may be something to be proud of.

The new CPUs do have one really nice touch, they are all halogen free and lead free. In adition to being compliant when the 2010 RoHS exemption on lead goes away, they also avoid possible 'oopsie' like the one Nvidia just suffered, see here, here and here. It also saves polar bears, egrets, emus and little kids with a penchant for chewing on green squares with metal chunks on them.

This is all fine and dandy, but how do they actually work in the real world? We took five sets of Xeons and plugged them into the same rig we always use, a Stoakley-based development platform. In this case, it was a Supermicro X7DWN+ with 4 2G DDR2/800 FB-DIMMs and 2 Seagate Cheetah 15K.5 drives striped.

The CPUs were an X5472/3.0, X5470/3.33, L5335/2.0, L5420/2.5 and an L5430/2.66. All are 1333FSB models except the X5472, it is based on a 1600FSB. The memory correctly stepped down to 667MHz for all but the X5472. All parts are 45nm Harpertowns except the L5335 which is a 65nm chip. Power was measured at the wall with an Extech 380801 True RMS Power Analyzer.

To beat on the CPUs, we took used the same Valve Map Compilation benchmark that we have previously used. It will max out 16 cores easily, and hardly touches memory or disks, so that is exactly what we want to test CPU power with.

The raw numbers, W=Watts, time in seconds

Looking at the numbers, the X parts are right on top of each other at idle, as are the 45nm Ls, the 65nm chips showing a bit more leakage while doing lots of noting at millions of cycles a second. Maybe that Hafnium stuff works after all.

Under load, the X parts are on top of each other, and all the Ls are in a different but very close cluster as well. This tells us two things, the different classes are really different under load, far more than at idle, and that Intel can bin parts very precisely. If you do a quick check of Time * GHz, they all end up at roughly the same point, so there are no oddities skewing the results a lot.

If you look at power used, load watts - idle watts, you have a good idea of how much energy is used to do useful work. Multiply that by the time it took to do the work, and you have a good estimation of efficiency. By this count, it should be no surprise that the X series are far less efficient than the L series, and the faster chips in the same power envelope are the thriftier ones.

As you might have guessed though, CPUs don't work all that well on their own, they need HDs, memory, motherboards and inefficient PSUs to make them spit out bits in the correct order. Once you add in system power instead of simply excess power used over idle, you end up with a measurement that has real-world relevance. In this case, it is of paramount importance - it is how much energy you use to complete a task, and how many dollars you will have to spend completing that task.

Load power * Time is exactly that, how much energy you used to complete the Valve map compilation. By this measure, the two X series parts are marginally more efficient than the 45nm L series. Both sets kill the 65nm version, mainly because it took notably more time. Since load power is 3+ times 'used' power, this is no surprise.

In the end, what it comes down to is that if you are at high CPU loads, the faster the CPU you can get, the better off you are. The lower wattage CPUs are overshadowed by system overhead and the slight clock deficit they are at. It would be interesting to see how an 80W part does under the same circumstances. If you are at low load, buy the cheapest part you can.

In the end, the E-step Xeons do exactly what they say they do. No performance gain per clock, but a lot more efficient. They clock higher for the same, or a little less energy used than their D-step predecessors. As far as we can see, there is no down side to them, just better for less power, and that is what progress is all about. µ