The Inquirer-Home

AMD Quad FX breaks cover

Part 1: What it is
Thu Nov 30 2006, 07:16
SO, WHAT IS AMD 4x4, or Quad FX with Dual Socket Direct Connect Architecture and AMD Athlon(tm) 64 FX-70 Series Processors? Well, it is a marketing name from hell, but there are two parts to it, a CPU and a mobo.


The CPU is the heart of it all, and there are three sets, the FX-70/72/74 running at 2.6/2.8/3.0GHz respectively. The CPUs are only sold in pairs and cost $599, $799 and $999 for a set as you get faster. This is not a bad deal at all, if you think about it, it essentially drops the price of an FX-62 to $400 as long as you buy two.

All are 1207 pin, IE Opteron Socket F, parts that do not require registered memory. Other than that they are standard Opteron 22xx series parts, 90nm, 125W TDP and 2MB L2, 1MB per core. It is not a bad deal at all, you get a lot of power for the cash, and prices about where the top Kentsfield and Conroe XE systems are. Imagine that coincidence.

The other half of the equation is the mobo, at this time it is exclusively the Asus L1N64-SLI motherboard based on the Nvidia 680a chipset. The chipset is the only thing that matters, I would assume Asus is the exclusive supplier because of the tight time tables involved in getting Quad FX out. You can have speed or you can have variety, looks like they went for speed, and other will follow in time.

This chipset is the first that makes use of the Opteron's inherent scaling on the consumer side. There have been some attempts with the 2200/2050 on the server or workstation front, but none of these set the world on fire. Since the QFX is basically a workstation with enthusiast features, it makes sense that there would be workstation class architectures under the hood.


What you get is two of the 680a chipsets directly connected to a single FX-7x CPU, each on it's own HT link, not chained off the last 680a. This gives massive bandwidth and latency gains, and allows them to put a total of 50 PCIe lanes on the board.

Those lanes are arranged with each 680a having an x16, x16 with x8 electricals, and a lonely x1 slot. Theoretically you can do quad card SLI, but due to the miserable state of NV drivers around SLI, I wouldn't hold your breath on this ever working, much less working well. It took them the better part of a year to get functionality on the GX2 parts, this one will probably be replaced before the hinted at functionality appears.

The list of features goes on. While not strictly a chipset function, one of the main points behind the AMD architecture is memory bandwidth that scales with socket count. In this case, the second CPU adds another set of memory channels almost doubling real world bandwidth at a cost of latency. Still, it is a solid tradeoff.

The biggest and really only drawback to this system design is a mobo layout problem. There are only 4 DIMM slots, all of which have to be occupied to get the performance claimed. Filling the sockets with high speed ram is not a problem for a high end system, but the expandability is all of zero. This is a glaring flaw that will end up biting people in the ass. The CPUs will support 4 per channel, and this mobo flat out should have found a way to put that on.


Back to the features, we have a silly set of capabilities, mainly because they chose to wire it all in instead of saving cash and engineering effort. There is 2x GbE, optical/coax SPDIF, 10 USB 2.0 ports, 2 1394 ports, and 12, yes 12 SATA outs. This shows what AMD can do with HT if the chipset partners step up, and portends good things for the PCIe integration in 2008. Scaling IO with socket count is a good thing.

That brings us to the point of cost, and yes it costs a lot, a huge amount really, $300 and up. Nvidia is milking this one hard, you do get a lot, but you pay for it. Call me a skeptic, but I think you would be much better off with a 'normal' mobo until NV gets the driver issues sorted, $150 extra for a few SATA and USB ports is not worth it.

What you end up with is a system that has two sockets for the price of one, or at least two chips for the price of one. The mobo has two socket for more than the price of two, and that destroys a lot of the value proposition for the systems if you can remotely call $2000+++ systems value. For the enthusiast, it has a lot of promise.

As I said earlier, the price lies on top of Kentsfield, it will use the same memory, HDs and video cards, so the main differentiator between Quad FX and Kentsfield is the mobo. An Intel Badaxe2 is not that expensive in comparison, but does not have 2x x16 PCIe slots and won't do SLI but will do Crossfire.

If you want SLI, you have to buy an Nvidia 680i based mobo, and guess what, they are milking you there too. What ends up happening is a Quad FX system will only be marginally more expensive than a similarly kitted out Kentsfield.

The question on everyone's mind is not the what part, the specs are all above, but more about why and how. Those are fairly thorny problems with no simple answers. They all hinge around the word workstation, the idea of usage models, and how you plan to do things. Those are looked at in part 2 later today. µ


Share this:

blog comments powered by Disqus
Subscribe to INQ newsletters

Sign up for INQbot – a weekly roundup of the best from the INQ

INQ Poll

Heartbleed bug discovered in OpenSSL

Have you reacted to Heartbleed?