Fast times for industry-standard system hardware

If 2004 was the year that Advanced Micro Devices’ 64-bit x86 technology caught the attention of IT, then 2005 was the year that AMD began reaping the benefits of being noticed. In April 2005, AMD transitioned its Opteron server CPUs from single to dual cores, clearing the path for eight-way servers with the same footprint and power consumption of four-way machines. Budget-sensitive buyers were thrilled with affordable quad servers that occupied a single rack unit, and AMD’s share of x86 server sales passed the 10 percent mark.

This triumph of technology over marketing seems uncharacteristic of an IT market that is often slow to embrace new technology. A key enabler in AMD’s growth was the market’s acceptance that CPU clock speed is useless as a core criterion for selecting x86 servers. Instead, factors such as acquisition and operating costs, expandability, and throughput — areas in which AMD leads — have been elevated to the top of the list.

Also key to AMD’s surge has been the company’s refusal to give ground to Intel in any area of technology. Intel expended considerable money and energy touting the fast, scalable PCI Express expansion bus as a differentiator, but it took no time at all for AMD servers to take up the PCI Express bus as a standard. Intel’s vaunted advances in front-side bus clock speed impress only when compared with prior Intel architectures; AMD servers integrate the features of the front-side bus into the CPU, leaving Intel’s technology at a disadvantage.

In 2006, the arrival of four cores, hardware-assisted virtualization, and Intel’s stop-gap fixes to its squeezed shared bus are sure things, but trench warfare between AMD and Intel will mean flexible road maps and, we think, the vendors holding back the most interesting bits as trade secrets. The first one or two quarters will play out as planned, but third quarter 2006 through first quarter 2007 will be a hootenanny.

Sun’s low-power, low-clock-speed UltraSPARC T1 (code-named Niagara) — an eight-core, 32-thread CPU — foreshadows some of Intel’s approach in 2006. Multiple cores and threads running simultaneously equates to more work per clock tick, and slower clocks mean cooler chips, an Intel objective that makes us very happy. Intel would also do well to address its trailing math performance, but its traditional passion for large cache — and for getting ahead of AMD on cores per chip — will probably win out as they’re sure attention-getters in competitive ads.

Finally, AMD will have the ability to put a 32-way Opteron server in a 3U chassis, with the power consumption of an eight-way server, as soon as AMD and one of its key OEMs see a market for it. Intel and AMD are pulling in their cards and are leaving us all guessing, but what’s sure is that 2006 and 2007 will leave us with a choice of x86 technologies born of brutal competition in engineering and price.