Intel or AMD - Which is better for HPC

If you were asking this question 6 months ago, there would only have been one possible answer. AMD had the HPC marketplace in the palm of their hand, with their Dual Core Opteron solution. At that time, Intel's Dual Core offering, was expensive and didn't offer the same performance level as AMD. What a difference a few months make. In the last few months, not only have Intel introduced a Dual Core technology which rivals AMD in both pricing and performance, but they now have Quad Core solutions shipping a couple of months ahead of AMD's expected first shipment. Now that we have a choice, it means that we need to take a closer look at our HPC requirement. No longer can we just take the fastest AMD processor that fits within our budget, now we have to consider which processor is likely to work best with our code. So what are the differentiators which make one processor better than the other?

AMD have HyperTransport - a 1000MHz channel dedicated to communication between processors and memory. This gives the advantage that memory transfers are much faster than via the FSB technology implemented by Intel. However, the downside is that each CPU is effectively responsible for its own bank of memory, meaning that for optimum performance at least one pair of DIMMs is required for each CPU installed. There is also a HyperTransport Slot on AMD boards to support connection of devices directly onto the HyperTransport Link. Recently network cards supporting the HT connection have been released, offering ultra-low latency for inter-node communication. Intel have larger cache sizes, and a higher raw performance than the AMD equivalents.

For CPU intensive tasks which require little interaction with memory, the Intel CPUs offer a better performance than AMD. For network interconnects, the PCI-X bus needs to be used, which will give a higher latency than similar cards attached to the HT connection on AMD units. One other advantage with using Intel processors, is that the Intel Compilers are optimised for the Intel Processors - that's not to say that code compiled with Intel compilers won't work with AMD, it just works more efficiently with Intel CPUs.If you're using open source compilers, they don't generally tend to favour either CPU. With the introduction of Intel Quad Core, another choice is thrown into the mix - should I go for 2 twin Dual Core nodes or a single twin Quad Core node? The answer again depends upon your requirements. Running a twin Quad Core node will give you a similar performance to running 2 twin Dual Core nodes (Quad Core is marginally slower than two Dual Cores), however, you need to bear in mind that all of the system resources will now be shared amongst twice as many cores (So if you have 16GB RAM, you're now running 2GB per Core, instead of 4GB per Core - You now have 2Gb/s network bandwidth shared between 8 cores, instead of 4 etc.). If you have an application which is particularly resource intensive, then running two Dual Core nodes, would be more beneficial than a single Quad Core node. Quad Core, does bring a couple of advantages to HPC though - Cost and power draw.

Since using Quad core, means that you take 1 chassis to run 8-cores, instead of 2 chassis, the node cost for Quad Core is lower than Dual Core. Although Quad Core draws more power than Dual Core (Dual core draws 65W, Quad Core draws 80W), if you compare on a performance basis, the Quad Core consumes less power (i.e. a single Quad core is the same performance at 80W as two Dual Cores running at 2x65W=130W) - a reduction of over 35%. Of course, this reduction in power requirement means that you can save money on cooling your HPC solution - 35% less power used, means 35% less heat generated meaning you can run more cores before you need to go to the expense of additional air-conditioning.

If you're still unsure as to which processor type best fits your code, then Viglen offer the opportunity to log-in to our cluster and benchmark your code on the actual platforms you're considering using - See below for more information on accessing our HPC Cluster.


Viglen's HPC Cluster
The Viglen HPC Test Laboratory has been configured to allow users remotely connect (via ssh) to our in-house cluster. This facilitates users and developers access to various hardware platforms and software tools, which can assist with the selection of the most appropriate cluster solution to suit their requirements. Our cluster is constantly being added to and upgraded with the latest technologies. If you have any doubt as to whether AMD or Intel is best suited to your code, this is a perfect opportunity to make that decision for yourself! Test the scalability of your code across multiple cores and multiple nodes, determine what memory configuration is best suited to your code, test out various compilers and libraries to maximise the performance, all form the comfort of your own office. As both Intel and AMD claim their architectures are superior to one another, it can be quite confusing as to which architecture to select. Online benchmark figures don't always reveal an overall performance for a system. This facility provides users with an excellent opportunity to gain some hands on experience with our hardware/software and obtain REAL performance figures for YOUR applications or code.