Athlon 64 and Prescott Get Ready To Rumble
After a busy spring, desktop PC processor buffs have nothing to watch but summer reruns — Intel’s Pentium 4 and AMD’s Athlon XP received speed bumps and faster front-side buses in April and May, respectively, but things have been pretty quiet since. That’s going to change in a big way, however, as both firms launch truly new CPUs this fall. Here’s an updated preview of what to expect — and how the archrivals’ offerings show very different views about where the desktop market will be in late 2003 and early 2004.
Opterons for Everyone: AMD’s Athlon 64
Sometime this autumn — the latest guesses point to the second half of September — AMD will take a quantum leap from its venerable K7/Athlon/Socket A architecture, and traditional follow-the-Intel role, with the introduction of the Athlon 64. The desktop sibling of the Opteron server and workstation CPU introduced last April, the Athlon 64 represents AMD’s bold bid to move mainstream computing from 32- to 64-bit technology, without leaving existing 32-bit Windows software behind.
Like the Opteron, the Athlon 64 offers a colossal increase in memory-addressing headroom compared to the 4GB limit of 32-bit processors; the chip can handle up to a terabyte of physical memory (using 40-bit addressing) or 256 terabytes of virtual memory (using 48-bit addressing).
While PCs with more than 1GB of RAM are still scarce today, AMD is betting that cinematic-quality 3D games and other applications will be bumping up against the 4GB ceiling on the desktop — as enterprise databases and simulations are already doing on the server — before we know it.
Meanwhile, to meet the never-ending demand for faster performance, AMD has made several improvements — such as a longer pipeline, improved branch prediction, larger look-aside buffers, and support for SSE2 multimedia instructions — to insure the Athlon 64 will outrun an Athlon XP at the same clock speed.
The most important piece of the Athlon 64 performance puzzle is the CPU’s integrated memory controller, which allows extremely low-latency accesses to DDR memory — while Intel’s latest Pentium 4 raised the speed limit of the front-side bus from 533MHz to 800MHz, AMD has eliminated the front-side bus altogether.
This requires a new way of looking at memory and CPU bandwidth requirements, as the chipset and system bus speed have been taken out of the equation. It also brings up the option of boosting memory performance through a CPU design change, and AMD may be considering this path already — while the initial Athlon 64s will incorporate a single-channel memory controller, AMD may be readying a dual-channel design for 2004.
Using its 0.13-micron SOI (silicon on insulator) fabrication process, AMD has found room on the Athlon 64 die for a hefty 1MB of Level 2 cache. This is an important design win, since Intel will move to a 90-nanometer process before reaching the 1MB mark for its desktop line.
AMD has decided to identify different Athlon 64 chips with model numbers indicating overall performance, akin to the Athlon XP’s, rather than the generic Model 240/242/244 scheme of the Opteron. In clock-for-clock performance comparisons, the enhanced core, larger cache, and 64-bit architecture should all help the new CPU outrun its competition.
The only potential issue is that the Athlon 64 will ship at significantly lower clock speeds than its Intel rivals or even than the Athlon XP: The lowest-speed model at release, the Athlon 64 2800+, runs at 1.6GHz (like the Opteron 142 and 242), while the Athlon XP 2800+ has a core speed of 2.08GHz.
This disparity in clock speeds brings its own set of challenges. Many benchmarks demonstrate overall system performance, while others concentrate on video- or CPU-specific functions. Excelling in all of these areas is key to winning the benchmark race, and a 1.6GHz processor is going to be hard-pressed to match a 3.2GHz or higher competitor in pure CPU tasks. However, AMD’s prospects should get steadily better, as Athlon 64 3100+, 3400+, and 3700+ models are all scheduled for 2003 release, with a 4000+ to follow early in 2004.
A Duron 64?
All of the above refers to the primary, top-of-the-desktop-market Athlon 64, but AMD is reportedly readying a lower-cost model — sort of an AMD64 equivalent of its old Duron CPU — to replace the Athlon XP as an affordable mainstream offering. This Athlon 64 “Paris” will feature only 256MB of Level 2 cache, and could duplicate the Socket A Duron and Athlon families’ same-pinout approach, much beloved of system upgraders and IT managers, for the 64-bit generation.
In the meantime, it might get a little crowded with all the AMD desktop CPUs to choose from — not only two different Athlon 64s, but multiple bus-speed variations of the Athlon XP “Barton” and older “Thoroughbred,” or perhaps the rumored “Thorton” (the up-to-date Barton core with a Thoroughbred-level 256K instead of 512K of L2 cache). AMD even still has a couple of Duron models in production.
Though the Athlon 64, like the Opteron, is built to be blazing fast with existing 32-bit operating systems and applications, software support will be a critical factor in the new CPUs’ success. While Microsoft has pledged to join today’s handful of Opteron Linux distributions with 64-bit versions of Windows XP and Windows 2003, time is growing short: the Opteron is already in the channel, and for the Athlon 64 to make a big splash in the 2003 holiday-season buying market, AMD64 Windows XP must be ready upon launch.
Intel Shrinks and Speeds Up the P4 with “Prescott”
In October or November, Intel is expected to ship its next-generation “Prescott” desktop CPU — not as radical a change as AMD’s Athlon 64, but an evolution of the Pentium 4 to a smaller, faster, more powerful design. Expected enhancements include microarchitecture and cache changes, not to mention higher clock speeds — all made possible by the biggest advance of all, the move from 0.13-micron to 90-nanometer (0.09-micron) process architecture.
Prescott’s anticipated 1MB of Level 2 cache is double that of current Pentium 4 processors. As with the 0.13-micron Athlon 64, squeezing all that SRAM onto the die requires a lot of transistors, which might negatively affect chip yields, but the higher density of 90-nanometer design makes it easier for Intel to add more cache and still get a cost-effective number of chips per wafer. Prescott also doubles the Pentium 4’s Level 1 cache from 8K to 16K, and even though the numbers are small, the performance impact could be significant.
It will all add up to nearly 100 million transistors per chip, obliging Intel to introduce various internal improvements to boost efficiencies and reduce core voltages — to a rumored 1.1 to 1.25 volts, versus today’s Pentium 4 parts’ 1.5-plus.
Unheard-Of Clock Speeds
These are important considerations, especially if Prescott is to hit the clock speeds projected for it — while the CPU may debut at little more than the current Pentium 4’s 3.2GHz, analysts expect 4GHz if not 5GHz in the relatively near future. Considering that the original, 0.18-micron-process P4 core reached 2.0GHz, double that would seem to be an easy target for the 90-nanometer Prescott, but only time will tell, as die size is just one facet of the quest for clock speed. The front-side bus speed is expected to stay at 800MHz, though as Intel has shown with the P4, this may be only temporary.
Hyper-Threading Technology is a top-priority initiative for Intel, and Prescott will be the first processor line to include this mock-multiprocessing, better-multitasking feature from top to bottom. Intel has also alluded to new x86 instructions and improvements in Hyper-Threading efficiencies, which — if the chip giant gets its wish and software developers optimize all their wares for the technology — could turn Hyper-Threading from today’s “nice to have” status or 5- to 15-percent performance boost into a formidable weapon against AMD.
There are several open questions regarding Intel’s Prescott, related mostly to its release date and projected name. Currently, as mentioned, October is the optimistic and November the cautious prediction, but this could be subject to change, especially given the shift to a new 90-nanometer process. Nvidia learned this lesson with its late-to-market GeForce FX, and introducing unproven technology is always risky, even for Intel.
The brand-name betting is turning into an Internet side show, with most predicting that tradition — and the several iterations that have already put different Pentium 4 models’ comparative clock speeds out of sync with real-world performance results — will dictate a fresh start with the label Pentium 5. On the other hand, the name Sexium definitely has its fan base.
When Giants Collide
In short, Intel and AMD are gearing up for another round in their head-to-head battle for desktop CPU supremacy, and each has adjusted its strategy or tuned its technology to adapt to the competition. AMD can’t compete with the insanely high clock speeds that Intel can reach, so the Athlon 64 design stresses architectural efficiencies and 64-bit compatibility. Its integrated memory controller is one of AMD’s key advances, and the prospect of a dual-channel successor has likely caught the attention of Intel’s engineering department.
Likewise, the Intel Prescott and accompanying chipsets will need an answer on Day 1 for the faster memory accesses of the Athlon 64’s on-die memory controller, though doubling the L2 cache to 1MB is an obviously good gambit. Look for this type of punch and counterpunch to continue into 2004 — or until Microsoft and its level of AMD64 support get the last word in the eventual success of either CPU.