Perhaps it seems all too long ago, but a mere 6-months prior we were looking at Oil prices above $140 and some speculating that it would top $200 by the end of 2008.Thankfully, the rampant speculators were abated and the pendulum has swung in the other direction, but now there are other primary concerns in the frontal lobes of most people – general economic issues.However, regardless of what is keeping you up at nights – energy prices, economic woes, or even just the environment, if you are making high performance computing applications that need to survive is less than pleasant conditions or last for a very long time, you should be considering the Cranberry Lake Platform from Intel.
The key distinction of Cranberry Lake is that it combines those wonderful Intel® Xeon® Processors that you know and love with the Intel® 5100 Chipset – and that is where the magic begins.It still supports the Dual-Independent Bus (DIB) that maximizes FSB bandwidth up to 1333 MHz (or 10.6 GB/s each) like the Intel® 5000P chipset, but the big difference is that is uses the more power-efficient DDR2 memory technology that lowers the total system power consumption while not significantly changing the performance of the solution.Another important distinction is it is tied to the Intel® ICH9R south bridge instead of the Intel® 6321ESB I/O Controller.When you put the impact of both of these changes together, the total chipset drop in Watts consumed is 12.4W, not including the bonus drop you get by switching from Fully-Buffered DIMMs (FB-DIMMs) to DDR2 DIMMs, which can be 5W to 8W per DIMM (perhaps slightly lower) or 20W to 32W in a 4-DIMM system.Add that all up, and a 4-DIMM DP server (which are very common in blades) could consume up to 44.4W LESS!
However, like how you would not pair a really good Cabernet, say the Monsanto Nemo, that with a Whopper® (but they are GOOD!), you’re going to want to pick the right processor to go with the wonderfully efficient Intel® 5100 Chipset.That’s where ARK comes in!If you go to the home page and pick “Quad-Core Intel® Xeon® Processor,” you’ll get a long list of great processors.However, when you are looking for the lowest possible energy consumption, you can click on the column header “Max TDP” it will automatically sort it by lowest to highest.From there, you will see there are two options that consume a measly 40W (which would be just 10W per core!),but we can see that the Quad-Core Intel® Xeon® Processor L5408 (12M Cache, 2.13 GHz, 1066 MHz FSB) runs faster and has a larger cache than the Quad-Core Intel® Xeon® Processor L5318 (8M Cache, 1.60 GHz, 1066 MHz FSB), so let’s pick that (plus it also has the Intel® 45-nm goodness & supports SSE4).
If we add up all of that so far, we get an outstanding 110W for both Intel® Processors & the server-class chipset.Think about that for a minute.It wasn’t but a few years ago, we were producing server processors based on the code name Irwindale that consumed 110W *per processor* and those were single-core processors to boot.Not you get a whopping 8-cores and the chipset.But, what can you get for that?In my opinion, the best way to answer that is to go to benchmarks because no one really buys a processor for how it makes them feel but rather for what it can DO!
Conveniently, there are 2 great platforms was can look at:
- IBM BladeCenter HS21 XM (Intel Xeon L5408)
- Quad-Core Intel® Xeon® Processor L5408
- Intel® 5100 Chipset with ICH9R
- 16 GB (8 x 2 GB DDR2-5300F ECC)
- SuSE Linux Enterprise Server 10 (x86_64)
- Intel C++ Compiler 10.1 for Linux
- SPECint_rate_base2006 = 82.4
- Sun Netra X4250 (Intel Xeon L5408 2.13GHz)
- Quad-Core Intel® Xeon® Processor L5408
- Intel® 5100 Chipset with ICH9R
- 64 GB (16x4 GB PC2-5300F, 2 rank, CL5-5-5, ECC)
- SUSE Linux Enterprise Server 10 (x86_64)
- Intel C++ and Fortran Compiler for Linux32 and Linux64
- SPECint_rate_base2006 = 82.0
Clearly, both of the above are prime of examples of the Cranberry Lake Platform.For sake of argument, we’ll split the difference and call the performance 82.2 (on average).We take just the processor TDPs (because, outside of Intel, it is incredibly difficult to find the TDP for chipsets), we see that we get ≈ 1.0275 SPECint_rate_base2006 (SIRB6) per Watt.
What about competing platforms?In the world of long-life embedded processors, there are not a lot of options that compete with the Intel® Xeon® brand, but 2 notable ones are AMD Opteron* and Sun UltraSPARC*.Let’s start with AMD and take the best processor on the current embedded offerings found here:
If anyone knows a better way to determine AMD’s embedded offerings, please add it to the comments.To me, the best dual-socket offering would be the “2214 HE” – you can find the complete specs here:
If we do a search for SIRB6 results for that product, here’s the best I can find (feel free to add a better on in the comments if you can find it):
It got a SIRB6 score of 40.5, which is just under half of what we got.On the surface, this makes a lot of sense because their solutions is 4 cores and ours is 8 cores, but that is before you compare them in energy efficiency.The “2214 HE” has a TDP of 68Ws according to this:
So, if we take their SIRB6 and divide by 136W (2 x 68W), we get: 0.2978 SIRB6/W – or less than 1/3 of what we got.
How does the UltraSPARC T-2 CPU, which is touted to be “…the gold standard for green computing and efficiency,” stack up?Well, thanks to this Blog Entry by Sun, I have found a URL for ALL SIRB6 Results!If you scroll far enough down, you will come to the highest performing system based on the UltraSPARC T2: Sun Blade T6320.It get’s a SIRB6 of 76.4, but the processor has a TDP of 123 Watts.Therefore the SIRB6/W is 0.6211 – which is still over double what AMD got, but less then 2/3rds of our results, given that their performance is close to hours, but the power consumption is over 50\% higher.
So, clearly, when compared to the other long-life options on the market, Cranberry Lake is a leader, but what if you are looking to upgrade what you got already?Well, we’ve got 2 examples:
The first one is based on Irwindale, so it is a bit older, circa 2005, where the second one is based on Woodcrest, so it was released just 2 years ago.Now, in the interest of full disclosure, the Irwindale-based part is not any “efficiency-optimized” part – but it is the oldest part that has been publicly run on SIRB6.However, the Woodcrest-based part is a highly efficient option, so we’ll do that comparison first.
Here is the highest results for the Dual-Core Intel® Xeon® processor LV 5148: Dell Inc. PowerEdge 1950 III, which got a score of 51.1.However, given that the TDP is exactly the same as our Cranberry Lake-based systems above, but the performance is 38\% lower, it gives us a SIRB6/W of 0.6388, or just a pinch above what Sun was able to obtain.
Now, as you can imagine, the Irwindale-based product will be much worse.Its results are based on the HP ProLiant DL380 G4, which got a score of 20.9.However, given that each processor was 110W (or 220W combined), then the SIRB6/W was a sad 0.0950.While not a great example, it does serve well to demonstrate exactly how much progress has been made in just under 4 years.However, it we neither long-life nor efficiency optimized, so it is just an interesting sidebar.
Where it all really get’s interesting is when you put all these results together into a Table and compare the results:
And I think that is its own conclusion – when you take these processors, in real-world systems, and put them through the exact same benchmark, then if you care about your performance per Watt – the choice is clear.
But, to close on the opening: Being able to do more while using less - is good for everyone.First, I think the pendulum on Oil has swung too far and I think we will soon be back into the $60 to $80 range (my guess is mid-2009), so the “energy monkey” will soon be back on everyone’s back, just not quite as painful as before.In the meantime, using the extreme corner case of going from servers using Irwindale-based processor to those using Harpertown would be a huge 4x jump in performance, which could mean you only need 25\% of the servers you have today and each one of them would also consume less energy then the older generation, as well as be more cost effective to manage – especially with advanced technologies like Intel® VT.So, while there would be some initial CapEx spending, this would drive a large reduction in OpEx, and, perhaps crucially, could prevent the need to have to build a whole new datacenter (which would be so painful in this economic environment).And then, lastly, as I think everyone is aware of by now, most countries get most of their energy by burning one fossil fuel or another.So, every Watt of power you save (both in servers that consume less energy and also put off less heat which then needs to be removed), is that much less Carbon that gets put into the environment. To be fair, Data Centers are only estimated to account for 0.5\% of the overall Carbon Dioxide emissions, but every little bit helps.