Jump to content
The Inquirer-Home

AMD dual socket six-core LV server

First INQpressions Visiting Istanbul on a low power budget
Thursday, 29 October 2009, 13:52

WHILE INTEL HAS the overall performance lead across desktops, workstations and servers alike, AMD had to focus on adding more cores to its dies to make up for the speed shortfall, at least before the long awaited Bulldozer cores appear, maybe in about a year's time.

There are still many types of jobs where the number of cores can be more important than the individual core speed, though. For instance, highly multithreaded tasks or situations where many small tasks that are more I/O than CPU bound - like networking or database search - proceed in parallel. In such cases, the lowest possible CPU power consumption coupled with the availability of many cores to take care of the each thread or task with minimal thread switching overhead can be very helpful.

amdistanbulmobo

AMD now has its ultra low power 40W "Istanbul" - or Opteron 2419EE to be precise - six-core CPUs for exactly such applications. While its quad-core "Shanghai" Opterons reached a maximum frequency of 2.3GHz, the six-core part had to give up some clock speed, down to 1.8GHz, to manage the thermal load for six cores. However, the Istanbul chip keeps the Opteron line's full 1GHz Hyper Transport 2 speed as well as dual-channel DDR2-800 memory per CPU.

We had a look at one of the first test platforms using these chips, the reference Gigabyte server mainboard platform with two 1.8GHz LV Istanbul Opteron 2419EE processors, providing a total of 12 cores. And, to spice things up, we also ran some of the well multithreaded desktop benchmarks to check how far this configuration will go with the codes our readers are more familiar with.

The feature rich GA3-C board from Gigabyte, based on the Nvidia Nforce Professional 3600 MCP system chipset, was set up with 4GB of DDR2-800 registered ECC server memory, one for each of the four channels, as well as a Sapphire ATI HD4870 2GB graphics card fitting the only available PCI-e slot. The usual dual Gigabit Ethernet, SATA plus SAS storage interfaces, as well as Gigabyte's integrated server management circuitry complete the feature set.

We used the plainest possible default aluminum-only heat sinks with small 6cm fans and no heat pipes or such widgets at all. With this in place we installed Windows 7 64-bit and a bunch of benchmarks.

First off was 3Dmark 06 which came up with a pretty decent CPU score for a 1.8GHz CPU.

3dmark06-2560-amd6coreati4870

What's more important is that the nearly four year old benchmark scaled well across eight cores, leaving the other four cores for other system needs. The new 3Dmark Vantage would scale well across up to 32 threads.

Now, look at the Sandra 2009 CPU and memory results below. The Sandra 2009 benchmark test uses all the threads available.

By the way, the heat sinks were barely hot during these benchark runs, a testimony to the low actual power usage and heat generation of the new low power Istanbul CPUs. Keep in mind that many new games and other desktop software can also make use of many threads at once - with a combination of physics, AI, effects and background tasks - to fill the new cores with work.

sanamd6corecpu1800

Overall, We're quite happy with the performance of the low power technology in general benchmarks, and will also take a look at the server-specific point of view in future tests. While the CPU clock rates are down, memory and I/O bandwidth wasn't reduced, allowing fairly balanced performance to be achieved across most applications.

sanamd6coremem1800

Also in the future tests, we'll look more closely at the actual power usage graph of AMD's low power Istanbul chip and how it compares against the low power versions of both the old Intel Harpertown Xeon, the L5430 CPU, as well as the new L5500 series low voltage Nehalem Xeons. µ

 

Share this:

Comments
Can someone please fire Nebojsa Novakovic

This is getting ridiculous. Benching nothing other than 3Dmark or sandra on regular systems already produces some of the most useless reviews on the internet, but running those same benches on a 12 core server is making Nebojsa look like a caricature of a reviewer. Not even Tomshardware produced such nonsensical articles in their worst days.

Whats next, Crysis framerates using the integrated graphics? Benchmarking floppy format performance ? Measuring how fast it falls fron the ground if you drop it from 20m ? Those would all be equally useful and insightful

posted by : NooneYoudKnow, 29 October 2009 Complain about this comment
More, please!

I eagerly await your report on the floppy controller's performance scaling with 12 cores. That's next, right? It has to be!

posted by : Rasem Brsiq, 29 October 2009 Complain about this comment
Wrong?

"quad-core "Shanghai" Opterons reached a maximum frequency of 2.3GHz"

He must be talking about the ULV? I just found a 2.5 on newegg and I am pretty sure I saw a 2.7 on Dell.

posted by : voidlogic, 29 October 2009 Complain about this comment
FLOPPY DISC RAID!!!

Dude not just the floppy controller throughput we need to know how this thing does in a floppy Raid 5 configuration :D

posted by : Mauller07, 29 October 2009 Complain about this comment
Pre 32X Pci-e ???

Thomas Has had Much Worse. point may be that clipping marks of unsoldered slots. Sometimes when NEW Slot is in Next Gen Bin, Those strange solder marks appear.

Could already establish standard of 32X pci-e. WITH Mary Jo Inflection, Ability to take Next Gen dual gpu cards into NEW Ralm.

drashek

posted by : Smartee' Pants...., 29 October 2009 Complain about this comment
@Mauller07

Raid 5 would be way too technical for the author, but perhaps (s)he can benchmark overclocked floppy format performance.

posted by : NooneYoudKnow, 30 October 2009 Complain about this comment
4GB for 12 cores??

How cab you seriously test a 12 core system with just 4GB of ram, especially with a windows OS? This cpu would be great for my 3D rendering stuff but even under linux I'd want at least 1GB per core, and up to 4GB per core for large and complex scenes, if I'm going to avoid a lot of swapping (don't forget you'll want to allocate a correspondingly large amount of swap space too, and with so many cores it would help if you use swap partitions spread across several drives).

I'm afraid that any testing done with the current memory configuration just seems nonsensical to me. I also query the wisdom of using 3DMark or Sandra on such a system.

posted by : LeeE, 30 October 2009 Complain about this comment
Advertisement
Subscribe to the INQ Newsletter
Sign-up for the INQBot weekly newsletter
Click here to sign up Existing user
Advertisement
INQ Poll

Nvidia Fermi

Will graphics cards built with Nvidia's Fermi GPUs be a hit?