NOW THAT WE'VE SEEN
AMD's next-generation Hammer chip (now known as AMD64) up close, with real benchmarks, we have a much clearer sense of how the new AMD chips stack up against the competition from Intel. Judging by the benchmarks, Hammer has loads of potential, and it's easy to see why AMD elected to go ahead and release a server version of the chip. The architecture, with an integrated memory controller on each CPU and high-speed HyperTransport links between system devices, is brilliant for a server platform. However, the Opteron's performance in workstation-class applications seems a little shaky versus the Pentium 4 and Xeon. No doubt an Athlon 64 chip, in the form we expect it, would have trouble matching up against the Pentium 4 3GHz/875P chipset combo we tested recently
. By this fall, Intel will be hawking processors based on the Prescott core, built on Intel's 90-nanometer fab process and rumored to have a longer pipeline and a larger cache.
With those considerations in mind, I offer my own recipe for what the Athlon 64 needs in order to fulfill its potential and supplant the Athlon XP as a tough, direct competitor to the formidable Pentium chips it will likely face when it debuts.
Dual-channel DDR400 (or 533) The Athlon 64 is slated to appear with support for a single channel of DDR333 memory. The Opteron chips' integrated memory controllers support dual-channel DDR333 memory, apparently by treating the two memory channels as one big 128-bit bus. That's essentially the same approach Intel uses with the 875P chipset. It's a little crude, but it worksparticularly in applications like media encoding and gaming where memory bandwidth is typically a bottleneck. The Athlon 64 needs to adopt its big brother's dual-channel memory controller, and it needs to be tweaked to support dual channels of DDR400 or even DDR533 memory, in order to keep up with the Pentium line.
Get SOI right This one is probably more appropriately labeled "higher clock speeds," but right now, getting the chip fabrication process right is key to more MHz. AMD had to delay the Athlon 64 (this latest time) because of difficulties producing chips in volume with silicon-on-insulator technology, and they've partnered up with IBM in order to iron out those wrinkles. Let's hope they can manage.
The Hammer architecture offers stunning clock-for-clock performance, especially with SSE2 and floating-point math, but being a brainiac alone isn't enough. Some things simply require more megahertz, and the deeply pipelined Pentium 4 shines in these sorts of tasks. The Hammer's Athlon roots and its 12-stage pipeline don't add up to stratospheric clock speeds, so AMD's fabrication abilities are doubly important. The Athlon 64 will have to launch at speeds well over 2GHz in order to keep pace with the Intel chips.
Decent AGP (or PCI Express) There is some concern about the AMD64 architecture's ability to deliver high-speed graphics performance. By its very nature, the platform requires AGP (or PCI Express) communications to be packaged up and transmitted over a HyperTransport link. I'm personally optimistic about the AMD64 platform's prospects here, but transferring graphic data over a narrow, serialized link like HyperTransport carries some inherent latency penalties. AMD64 chipsets will have to overcome this potential problem in order to succeed in the workstation and consumer markets.
Frequent updates I'm not referring here to the occasional clock speed boost. Because AMD has integrated the memory controller into the A64, AMD will be obliged to refresh the Athlon 64 chip with some regularity in order to support new memory technologies. Chipset designers like VIA and NVIDIA won't be able to help out here, so the onus is on AMD. Also, from what I hear, there is lots of room left to improve the AMD64 architecture's memory controller. Newer chip revisions with improvements to the memory controller would be helpful. Intel has been moving at a decent clip on chipset enhancements in the past year or so, and AMD will have to devote resources to ongoing development in order to keep up.
A real 64-bit edge The Opteron is already laying the foundation for the Athlon 64, and yesterday's announcements from Microsoft and others offer hope. AMD64 can benefit from 64-bit code because of its extra registers, even if a larger memory address space isn't required. AMD needs to solidify support for its AMD64 instruction set architecture in order to set the Athlon 64 apart from its competition. Doing so will require support not just at the OS level, but in terms of drivers and applications, as well. For the consumer market, graphics and sound drivers will be key, as will the support of high-profile game engines like Unreal and Doom. One can hardly overstate the importance of code optimizations to the Pentium 4's current success. AMD needs to take a page from Intel's book and aggressively pursue optimizations for its new processors.
Keep the TPI promise AMD promised the world a clear, new processor performance metric out of its True Performance Initiative back when the Athlon XP debuted, and we're still waiting to see the results. When I last checked in with AMD about TPI, they anticipated delivering a new metric early in 2003. That time window has already passed, but I expect TPI is more tied to the Athlon 64 than AMD is letting on; the delay of the CPU likely triggered the delay of TPI.
Whatever the case, the Opteron model number system simply won't work for consumers, and the limitations of the Athlon XP rating system are becoming ever-more obvious. (How does one rate a processor that clobbers a Pentium 4 in Business Winstone but takes over 60 seconds longer to process a video clip?) AMD can address this credibility gap only by delivering on its TPI promise.
Of course, if these things don't work out, AMD could always fall back on the Apple option: sell cheap SMP systems with slower individual processors and older memory technology. The AMD64 platform today is much, much better suited in both software and hardware for multiprocessor systems than the Mac was when Apple went to duallies. (Just think: AMD could double theoretical memory bandwidth by dropping in a second CPU.) Still, that is not a pretty fall-back option, and I'd rather not contemplate it any further. Let's hope AMD can get the Athlon 64 recipe right, and such things won't be necessary.