Low wattage Harpertown - more details tip up

TODAY, INTEL is going to round out their 45nm server offerings with two new low wattage CPUs, the L5420 and the L5410. Both are 50W quad cores, and in our testing, delivered measurable power savings over their non-L brethren.

The new products are all quite similar to the Penryn CPUs you already know and love, they are simply low wattage versions of the existing Harpertown CPUs. The technical changes are basically nil, they are just binned for power consumption.

The older 65nm Clovertowns mostly sat in the 80W power consumption range and topped out at 2.33GHz in that envelope, and you could get 120W quads that hit 3.0GHz. With 45nm process tech, the Harpertown CPUs reached 3.0GHz in the same 80W envelope, but they were able to do it on a 1600FSB, one up from Clovertown's 1333. Harpertowns currently max at 3.20GHz/1600FSB while sucking 150W.

TDP does not tell the whole story though, far from it. Few servers spend more than a fraction of their lifespans running flat out, HPC being the notable exception. Many more putter along at low utilization with occasional spikes up to 100%.

To this end, leakage power and sleep states will be the overriding factors in power consumption. Harpertown scores big here, 45nm takes a huge chunk out of leakage power, and the new C6 sleep state completes the process. Harpertown can idle much lower than Clovertown, and one look at the TDPs shows it is more efficient at full throttle too.

This is what makes the two new CPUs so interesting, the L5420 runs at 2.5GHz while the older L5335 ran at 2.0GHz. Both take the same 50W, as does the 2.33GHz L5410. The new process allowed Intel to ramp up the clock by 25% without touching power used, upping the cache 50%, and in general, putting out a better chip.

Looking at TDPs is one thing, but as you know, a CPU on it's own doesn't do very much, even with 50W flowing through it. To make it be useful, you need a server, memory and HDs as well, all of which take various amounts of power.

To measure how well the new chips work, we started with a Supermicro 2U server, the 6025W+NTR+B with an X7DBE+ mobo. To that, we added 8 sticks of Nanya 2GB FBD-800 memory, an Adaptec 5805 SAS card and 2 146GB Seagate Cheetah 15K.5 drives in a striped config.

Four sets of CPUs were used, with two per run seeing as it was a two socket machine. We first used a 3.0GHz Harpertown engineering sample (corresponds roughly to an E5472), a Clovertown 2.33GHz (E5345), a low wattage Clovertown 2.0GHz (L5335) and the new low wattage 2.5 GHz Harpertown (L5420).

Because of the thrice damned Windows activation, triggered by hardware swapping, we decided to save the week of scratchy phone calls to India begging for use of our paid for licenses and just used Ubuntu 8.04 Beta x64 Desktop. It installed without a hitch, detected all hardware correctly, and just worked perfectly.

The test was designed to push the wattage as high as we could, a trickier thing than you might expect. If you push the CPUs hard, the memory and drives don't get worked out, and pushing either of those can starve the CPUs. In the end, the most CPU intensive task we could find that had notable effects on power use was video decoding. It didn't get us to 100% load, but it was more than enough to show realish world differences.

8 HD (1080p/MPEG2/24K, two different videos) streams were run, four per screen, and just for fun, a single SD MPEG2 video was played on a third. Additionally, we used 7Zip to compress the install CD (copied to the HD) and letting task manager run. A little unscientific, but quite repeatable and it gave the same results within a watt over 5 test runs.

The idle power is just that, sitting on the desktop doing nothing. Load power is decoding all the streams, zipping the install CD, and watching the task manager graph go by. It is peak power used for the entire run, not average.

One last word of note, there is a column titled "Load%", and it wasn't part of the original testing plan. During the run, with 8 cores fluctuating wildly as tasks are passed around, the average load was guestimated, and it was just a rough guess. It should be taken as no more than a ballpark figure about what the CPUs were doing, all were far from 100%.

That said, here are is the raw data:

You can see several things, from the graph above, both from an efficiency and sheer speed point of view. If you consider that the tasks were the same, the lower CPU usages were the faster CPUs, and they fall into line with the E5472 fastest and the L5335 slowest. This is exactly what you would expect, clock speed wins.

When you look at idle power, the picture changes radically. The obvious conclusion is that low wattage chips will idle lower than the normal ones, and that is the case, but just barely. The first interesting part is the 3.0 Harpertown takes a piddling 3W, just over 1%, more than then low power 2.0 Clovertown. The 2.33 Clovertown idles at 16W more than the high speed Harpertown.

In the idle power game, the low wattage Harpertown takes 12W less than the low wattage Clovertown, and a huge 28W, almost 10%, less than the 2.33 Clovertown. These differences are not trivial, if you have a low utilization server, these numbers can add up to pretty huge savings.

Loaded wattage is a little closer for the energy efficient duo, and this is to be expected. Both are binned to have the same rough loaded wattage, and this shows with a 3W difference between them. If they were vastly different, then something would be odd. Much more interesting is the 17W difference between the 2.33 Clovertown and the 3.0 Harpertown. That is around 5%, which can probably be explained away by the lower load on the Harpertown.

What this means is that between the low voltage CPUs, they both take about the same wattage, but given that constraint, the L5335 does much more work. Remember, everything in this test is equal, nothing was changed except the CPUs.

On the high wattage side, the 3.0 Harpertown does the same work for notably less power and with more headroom to spare. It is a clean kill. Probably the most important difference is that the low wattage Harpertown destroys the normal wattage Clovertown in every way. It uses more than 10% less energy for the entire server and has lower loading, so more headroom.

In the end, the choice is clear, the new low wattage Harpertowns are vastly better than any of the Clovertowns that came before them, low or normal wattage. From an performance per watt standpoint, even with the overhead of the server taken into account, the L5420 is a clean kill over any of the Clovertown series, both at idle and at load.

From the introduction of the Harpertown, it was clearly better than it's predecessor. The last of the Harpertown series is quite possibly the best, it may lack a little in raw performance, but it more than makes up for it in efficiency. Even with the relatively inefficient server setup that they were tested in, the new low wattage parts differentiated themselves.

Both the L5410 and the 5420 should be available immediately. The L5410 retails for $320 and the L5420 runs $380. The L5410 is also going to be slotted into the embedded line meaning it will have a 7 year availability.

If your servers will take them, the new parts are a great upgrade, they deliver something quite rare, speed and efficiency. If you are shopping for a new machine, it is well worth evaluating the new low wattage parts, something that was not necessarily the case for their predecessors. µ