Interface Overkill: AMD’s Quad FX System And Strategy

This blog post references my cover story 'Interface Overkill? Is eSATA Necessary For Your Next System Design?' in EDN's May 10, 2007 edition. It's one of a series of web addendums to the print writeup.

Last March, when Intel unveiled its exotically packaged, dual-die, quad-core Kentsfield processor (subsequently and formally named the Core 2 Extreme QX6700 in September), AMD reacted with by now predictable scorn. It was bad enough, AMD felt, that the dual cores in Intel's existing CPUs were sharing a common front-side bus. But now Intel was clogging that common bus with four cores' worth of traffic? No amount of FSB overclocking, in AMD's mind, would get around the fact that not only was Intel's approach inelegant, it was also performance-sapping any time each core's access reach extended beyond its dedicated cache.

The fundamental problem I had with AMD's pitch was that the Core 2 Extreme QX6700 didn't seem particularly performance-sapped when I benchmarked it. Nor, as I predicted at the time, has it seemed particularly performance-sapped to the vast majority of folks who've subsequently used the chip. Granted, it and its family follow-ons are pretty expensive when you evaluate them at the IC level (a function, among other things, of the lack of single-package quad-core competition from AMD at the moment). However, because they drop into a single CPU socket, they're fairly economical on a total system cost basis.

Compare an Intel quad-core CPU-based system to its alternative derived from AMD's consumer-tuned competitive response, the Quad FX architecture (formerly known by its 4×4 marketing moniker) showcased in my hands-on project, and you might walk away even more baffled at AMD's strategy. Instead of squeezing two dual-core die into one package, as Intel did, AMD went with a dual-CPU arrangement. There's nothing fundamentally flawed with this approach, aside from perhaps the incremental motherboard space it consumes; AMD's enterprise-focused Opteron products have been multiprocessor-arranged for years.

However, presumably so that it could deliver the biggest, baddest platform possible from a feature set standpoint, AMD partnered those two CPUs with two Nvidia nForce 680a MCP (media-and-communications-processor) core logic ICs. Page 2 of Anandtech's review includes a good high-level block diagram of the Quad FX topology. Granted, the combination delivers PCI Express connectivity potential in spades, along with beaucoup SATA, USB2 and Gigabit Ethernet ports (only a subset of which, I feel compelled to note, the ASUS L1N64-SLI WS motherboard implements).

However, if you scan the Quad FX reviews highlighted in the two-paragraphs-back Google link, the shortcomings of AMD's approach will quickly become apparent. It's expensive, for one thing, from a bill-of-materials cost standpoint. Since Quad FX is defined as a boutique power user offering, versus Intel's more application segment-flexible single-socket approach, the high BOM cost also couples with low production volumes to translate into high prices at retail. Not to mention the fact that, because Quad FX is a NUMA (non-uniform memory access) architecture from a system topology standpoint, it's hampered in the NUMA-deficient Windows XP world of today versus Intel's more uniform memory access approach, and until the more NUMA-friendly Vista O/S gains sufficient market momentum.

Lots of ICs, running at high clock speeds, also translate into lots of heat. By the time my Quad FX system arrived just before Christmas, AMD and ASUS had largely solved the loud fan noise issue that initial platform reviewers had uniformly complained about, via a combination of more robust BIOS- and driver-based power management and a more efficient active and passive cooling subsystem. However, it'd still be a stretch for me to call my Quad FX box 'quiet', and the combination of dual CPUs, dual MCPs, dual SLI-configured graphics cards, dual internal HDDs and other redundancy, all fed by a 1000W power supply, will still put a noticeable hit on your electric bill.

Will AMD's upcoming 65nm-based and single-die quad-core capable Barcelona microarchitecture be enough to help it steal the performance-per-watt and performance-per-dollar crowns back from Intel? We'll have to see; Intel's not standing still either, with 65 nm CPUs in full production, 45 nm-based Penryn-based systems already being benchmarked by reviewers, and follow-on Nehalem CPUs planned for next year.