This content requires Macromedia Flash Player 7 or higher. Get Flash

Nehalem Architecture

Wednesday, April 15 2009 10:24 AM

David Scott from Intel talks about performance features and improvements of the Nehalem platform. Find out why it is the biggest change for Intel since the original Pentium processors.

Hi, My name is David Scott. I'm with Intel's High Performance computing group.

And today I want to talk to you about the Nehalem Architecture.

Each Nehalem die has 4 cores, each core has a private L1 and L2 cache.

There's a shared 8MB L3 cache.

There are 3 on-die memory controllers and finally there are 2 Quick Path Interconnect (QPI).

Each Nehalem server node consists of two Nehalem sockets and an I/O hub.

As we said before, each Nehalem socket has 3 DDR3 channels.

If you use 1 DIMM per channel, you can run up to speeds of 1333. If you run 2 DIMMs per channel, you can run up to speeds of 1066. If you run 3 DIMMs per channel, you can run up to speeds of 800.

The Quick Path Interconnect is used to connect 3 sockets for sending data and maintaining cache coherency.

Each Interconnect runs with 12.8 GB/s in each direction between each of the socket.

Finally, the PCI express lanes of the I/O hub are used for connectivity to disks, networks, and graphics.

Now I'd like to talk about performance of the Nehalem platform, as compared to our previous generation Stoakley platform.

For peak performance, it's actually about the same. It's the same number of cores, and about the same frequency.

On the other hand, for memory bandwidth, as measured by the streams benchmark, 4x faster on the Nehalem platform compared to the Stoakley platform.

But what really matters is the performance of your application.

The number depends on how memory bound it is on the Stoakley platform.

We have measured a number of ISP codes, comparing Stoakley to Nehalem, and seen speed ups anywhere from 50% to a factor of 2.5.

There are many other performance features of the Nehalem architecture and I'd like to talk about 2 of them.

Hyperthreading, that is two threads resident in the core of the same time to improve throughput, is similar to what was on the Pentium 4 architecture, but wasn't available on the Core 2 microarchitecture.

And turbo boost, this allows the frequency of the processor to be increased when you're not using all of the cores or in fact even if you are using all of the cores but you are not consuming the available power budget.

The Nehalem architecture is a big change for Intel and in fact is the biggest change since the original Pentium processors.

I'm David Scott for Intel Corporation.