% fortune -ae paul murphy

Benchmarks: Intel vs AMD

The last few weeks have seen a certain amount of name calling going on between some key AMD executives and people from, or representing, Intel. As I understand it, this all started whan some AMD people got upset because the press was buying into Intel's claim to have delivered a quad core CPU when, in reality, all they've done is package two dual cores together.

Intel then fired back with some very selective benchmarks, AMD ultimately replied in kind, various Intel sites praised Intel's brilliance, technology, and honesty while Wintel apologists yelled at AMD for stooping to Intel's level and, well, here we are.

All of which raises one key question: is anyone right about any of this?

AMD is rather obviously right about Intel's ability to con a credulous PC press into equating packaging with design - but that's largely because these guys are so partisan that they completely miss both the big picture and its short term reflection in the current contest.

Big picture wise neither Intel nor AMD is going to take x86 design anywhere exciting because both are limited by what Microsoft can deliver - meaning that four way parallelism is currently well past the point of diminishing returns to hardware and therefore as far as either Intel or AMD is likely to go in this direction.

The short term version of this is that AMD and Intel have concurrent design and manufacturing contests going on - with AMD ahead in design and Intel ahead in manufacturing, but both already committing to taking x86 evolution back to the megahertz race.

Thus AMD's on-board hypertransport -and the forthcoming "barcelona" evolution - is clearly better than Intel's reliance on bridging hardware, while Intel's lead to both 65nm and 45nm manufacturing gives it clear advantages on both processor cycles and processor power use. Basically, AMD's design lead produces more throughput per cycle while Intel's manufacturing lead allows it to produce more cycles per second for less input power.

So what does this mean in real life? I think it means that, on average, AMD boards are likely to perform closer to their potential in the hands of ordinary users than Intel boards because the defaults are pretty good and you don't need a lot of hardware expertise to tune them for specific uses.

With hypertransport, for example, you don't need to debottleneck the bridge for large shared memory applications by tweaking the BIOS to create a kind of faux NUMA -arranging for large objects to be spread across multiple SIMMS, and therefore channels.

Since Intel has that expertise and the average user hasn't, you'd expect their benchmark results to be unapproachable in ordinary use - but, of course, we can't really test for that effect from published data. A corolary effect we can test for, however, suggests that AMD's benchmarking efforts should show better returns to power input than Intel's -i.e. less loss of processing efficiency per cycle as the number of cycles per second rises.

And, in fact, you do see that effect in published results. Consider, for example, a number of late 2006 results for the 100GB the TPC/H benchmark, all on Dell PCs using Microsoft SQL Server 2005 Enterprise Edt (x64) and Microsoft Windows Server 2003 Enterprise x64 Edition.

The best of the Intel dual core results shows a Dell PE6800 with 64GB and four dual core Intel Xeon MPs at 3.4Ghz scoring 16,320 QphH or 1.66Mhz/QphH, while the top rated Dell 6900 with dual Intel X5355 "quad" core packs at 2.66Ghz achieved 15,724QphH or 1.353Mhz/QphH.

In comparison the result for a Dell PE6950 with 64GB and four AMD dual core Opteron 8220SE at 2.8GHz shows 17,180 QphH or 1.30Mhz/QphH - 27% better per cycle ( and 5% better in absolute terms) than the comparable Xeon and trivially better (4%) per cycle but almost 10% better in absolute terms than the Intel quad pack.

And if all of that leaves you confused, think about it this way: all of these comparisons pit 90nm AMD processors against 65nm Intel products - meaning that AMD should jump to a 25% or greater lead when both are made at 65nm, and Intel will catch up again when they go to 4Ghz and 45nm.

Paul Murphy wrote and published The Unix Guide to Defenestration. Murphy is a 25-year veteran of the I.T. consulting industry, specializing in Unix and Unix-related management issues.