AMD Opteron 4P Server -- The First Look
The time is near...
This content was originally featured on Amdmb.com and has been converted to PC Perspective's website. Some color changes and flaws may appear.This is one of the more exciting times to be working closely with AMD and their processors. The Opteron processor officially launched in April ( see our article ) with a bang. The market saturation has been slower than expected from what I have been hearing from vendors and resellers, but it is still a giant leap forward over the initial buzz the Athlon MP received; and for good reason.
What we have here today is a quick first look at the first 4-way Opteron system that I have had access to. Here is the setup for this system:
4 x AMD Opteron 844 Processors @ 1.8 GHz
AMD “Quartet” 4P Motherboard with 8131 (PCI-X Tunnel) and 8111 (I/O Hub) HyperTransport
8 GB Samsung PC2700 ECC Registered Memory
Microsoft Windows Server 2003 (32-bit)
As you can see this is quite the powerful system. And, yes, for these initial tests we are running a 32-bit version of Windows as it was all that AMD would allow us to publish numbers from; they are against me using an Alpha version of the AMD64 version of Windows that has recently become available.
These first benchmarks you’ll see (all from a AMD64 compatible SiSoft Sandra) are running in 32-bit mode on 32-bit OS. Before we publish our full review on this system, we are going to be converting the OS on it to SuSE Linux AMD64. That will give us our best look at how the system will perform as a server under those specific kinds of environments. However, this 32-bit performance (and the added tests we will be doing later) are important to see how the Opteron is performing in 32-bit applications, and maybe even just a smidge of a preview of how we might see the Athlon 64 perform.
First, here’s a screenshot of the system properties window from the Operating System we are using, just for verification.
Here is the SiSoft Sandra information screen for the processor itself (one of the four). While I wouldn’t take a lot of the information to the bank yet, it is neat to see the “Supported Speeds” at 3.0 GHz+. These particular Opteron’s have 1 MB of L2 cache and are indeed running on the 200 MHz front side bus. Edit (6/9/03): This is FSB is a bit misleading. The Opteron processors connect to the other chipsets at a speed determined by the maximum speed of the individual chipset, through Hyper Transport buses. So, the AMD-8111 runs at 200 MHz (400 MHz DDR) while the AMD-8131 (PCI-X) runs at 600 MHz (1.2 GHz) and the processors can handle these seperate speeds individually, all while the Opterons continue to talk to each other at full 800 MHz (1.6 GHz DDR) speeds. Pretty neat, huh? The "bus" as its reading in the screenshot above is probablty how the processor is working internally and doesn't actually pertain to external memory performance.
This is a similar image of the motherboard specifications. You can obviously see this is an AMD reference board by the fake serial number and ID code. The interesting part is that there are four “System Memory Controllers” (only showing 2 on this shot) that correspond to each of the four Opteron processors. According to this, each processor can support 4 GB of system memory, bringing the total to 16 GB for the entire machine.
Here, we are actually going to get into some benchmark numbers. First up from SiSoft, we have the CPU Arithmetic test. As you can tell immediately, these numbers are simply astounding! 28,700 pts on the ALU test and 11,800 points on the FPU test are beyond anything we have seen before. The comparison table shows us what a common 4P Xeon system does. Note the non-SSE2 scores for the 4P Xeon and 4P Opteron – there is nearly a 100% gain on the side of the Opteron.
Next, we’ll have a look at the memory benchmark from SiSoft. Without a doubt, these are the wildest numbers we have yet to see. With the on-board memory controller for each of the four processors, we are able to see the theoretical maximums in memory bandwidth for the Opteron processors. Over 9 GB/s on the ALU test and the FPU test blow away anything we have seen on the market today.
The Sandra Multimedia test in this new revision of SiSoft is a bit more complicated and thorough, actually drawing items on the screen to benchmark different aspects. Here we see the 4P Xeon system actually comes out ahead by a small margin.
Finally, this latest benchmark in the SiSoft suite is a new one to me: the Cache and Memory Benchmark. It tests cache and memory bandwidth based on a number of different sizes of test blocks. The 4P Xeon and 4P Opteron switch back and forth their leads a couple of times with Xeon’s having a large lead in the opening stages (small test block sizes). What interesting is that all of the other test configurations bottom out at the 4 MB block size while the Opteron system is able to continue on to the 16 MB size and higher.
Click to Enlarge
A line from one of my favorite movies goes: "Yes, yes, but what does it all MEAN, Basil?" In all truth, these numbers here don't show us the full picture of the Opteron -- we'll need many more days with this system to get that. But, through these initial SiSoft tests we can see that the Hammer core processors (both Opteron and Athlon 64) are going to be real players in the game. With memory bandwidth of that caliber, Opteron servers have the ability to overtake any other system available. HPC (high performance computing) groups will no doubt see these numbers and drool -- that is why you are seeing groups like Cray quickly gobble up Opteron processors and platforms. We'll have more benchmarks to give a more rounded look at this 4-way Opteron server soon, so stay tuned!
There you have it, a very quick preview of our review on a 4-way Opteron server. You can expect a lot more Windows based benchmarking soon, as well as a considerable amount of Linux based benchmarks and tests in the near future. I want to hear your thoughts on this information in this thread and you can check out the AMDForums for more updates or for any questions you might have.