TechPulse 360

An Intel Nehalem server system cost almost twice as much as an equivalent (2 memory channels) AMD Opteron server

AMD finally strikes back at Intel’s competitive server claims.

It took indeed several days for AMD to respond to some of the claims that Intel made when it launched its next-generation server chip Xeon 5500 (Nehalem) last week.

And reading through the counter-claims, Intel’s Nehalem looks more of a jumbo-jet than a supersonic jet fighter!

Here’s a summary of my conversations with AMD’s server Chief Pat Patla and manager John Fruehe.

So is the slowest Nehalem chip really faster than the fastest Opteron chip, including AMD’s Istanbul server chip coming out at the end of the year?

How can that be? The slowest Nehalem is a dual-core chip. How can their dual-core chip be faster than our quad-core? They [Intel] just say those things without anything backing up their statement. The only benchmarks Intel published is on their top end parts. Nothing on the lower end. Intel has done a great job in marketing. I don’t necessarily agree that they have done a great job in driving value for the customers.

But Intel has added the super fast QuickPath Interconnect (QPI) with Nehalem?

QPI is actually a copy of HyperTransport. As a matter of fact, they could have licensed HyperTransport which is an open standard but instead they decided to go with their own proprietary infrastructure.

Now, Intel will rave about the performance of QPI but it’s only if you buy their top end parts. If you buy their mid-range parts, the QPI speed drops down, and if you buy their lower end parts the QPI speeds drop down even more.

Meaning that if you have an application that rely on high I/O and high memory throughput but doesn’t need a lot of compute power, like a Web server, a file server or network infrastructure – which are the real backbone of today’s data centers - you would have to buy the fastest Nehalem processor to get the fastest QPI! Instead, we offer the same HyperTransport speed on all of our Opteron chips.

And hyperthreading?

Real men use real cores. We’ve got real cores across our products. Hyperthreading is basically designed to act like a core except that it only gives 10 to 15 percent performance bump for real applications workload. That’s because hyperthreading requires the core logic to maintain 2 pipelines: its normal pipeline and its hyperthreaded pipeline. A management overhead that doesn’t give you a clear throughput.

You’re saying that Nehalem chips are overpriced. Why?

Yes. A Dell server with the Nehalem 2.93 GHz chip is 104 percent more expensive (~$6.100) than the same configured server equipped with a Shanghai processor at 2.7 GHz (~$3,000). At this price, I sure hope so that they are faster. So if you’re in a tough economy and you’re trying to make your budget dollars as far as you can, you’re probably not going to buy half as many Nehalem servers but more cost effective Opteron servers.

It’s somewhat disingenuous to layout all the benchmarks and say “we’ve got a better platform” and completely ignore the pricing aspect of it.

Why are Intel-based servers more expensive than AMDs?

1. First off, the price of the Nehalem chip itself is more expensive than the Opteron chip;
2. Then, they use DDR3 memory which is more expensive, draws more power and has higher latency. So DDR3 is not a good choice for 2009. But in 2010, the tables will turn on DDR3 with lower prices, lower latency and lower power;
3. The Nehalem servers have 3 channels of memory, versus 2 for the Opteron. So where we would put 2 DIMMs, they would put 3 DIMMs in, which makes it 50 percent more expensive in DIMMs and it’s going to consume 50 percent more power from the memory perspective;
4. Because of the size of the socket and because of the 3 memory channels, Intel needs to have more layers on the board, plus special VRMs, etc… making the whole infrastructure more expensive to build.

What about Intel claim that a customer can consolidate 9 single core servers on one single Nehalem server ?

We also support all the virtualization platforms (VMware,Microsoft HyperV, Xen…) which let one dual-socket server support on average 5 to 10 virtual machines. So what Intel is really talking about is virtualization and we do that as well! There’s no reason that you could not support the work of 10 single core servers on an Opteron. They are making that sound as something unique that only Intel can do, but they’re not the only platform that runs virtualization.

Intel also claims that in some cases, Nehalem servers have an ROI of only 8 months!

Again, it’s disingenuous to talk about ROI to the IT world as a hardware vendor. Because people look at a complete solution: it’s hardware, software, lifecycle management, licensing, power, security… And if you look at any TCO models – which is what you’d use to do an ROI analysis – it will say that acquisition costs (hardware and software) is about 25 percent. And the software is a lot more expensive than the hardware. So if your hardware is about 10 percent of the cost of the total solution, how are they coming up with an ROI of 8 months? I’m sure they are doing the math thinking “if you’re buying the server today and you unplug 10 single core servers, the amount of power that you’d save would payoff this server.”

And Nehalem servers being a cash machine after 8 months?

Maybe after 8 months, it starts to print off enough money to pay for the 104 percent price premium that you pay at the beginning! You could virtualize 10 servers on an Opteron platform using virtualization and unplug them. And because we are half as much in cost, our ROI should technically be 4 months, shouldn’t it? If they can do it in 8 months, and we cost half as they do, we should do it in 4 months, right? But I wouldn’t make that statement to customers because I’ll be laughed out of their office, because it’s not how they measure ROI.

Is Nehalem really all that bad?

Intel have done a lot of great work to bring down the idle power, which is great on a desktop but less an issue in servers. So, while Nehalem has a very low idle power, in a data center you have to set all of your parameters around the highest amount of power the platform can draw. And by design Nehalem servers draw more power than Opteron servers. Which means that you can put less of them in a data center than AMD servers