AMD's K Series

AMD's K5 was the company's first independently created x86 processor; previously, AMD made 286, 386, and 486 CPUs that were directly derived from Intel's designs. Because AMD underestimated the difficulty of creating an independent design and made some bad design choices, the K5 was late to market and has been limited to entry-level systems. By the end of this year, the K5 is likely to disappear.

AMD's future lies with the K6, which began shipping this spring. The K6 started life as NexGen's 686; when AMD ran into trouble with the K5, it bought NexGen to get its 686 design and the team that created it. The K6 has MMX, more on-chip cache (32K each for instructions and data), is able to process more instructions in parallel, and runs at higher clock speeds than the K5 processor.

For the K5, which delivers better-than-Pentium performance at the same clock speed, AMD used the PR rating system. A 100-MHz K5, for example, delivers Winstone performance comparable to that of a 133-MHz Pentium, so AMD calls it the K5-PR133. With the K6, however, AMD has gone back to a MHz rating. This change was driven by the shift in the reference from Pentium to Pentium MMX and Pentium II. No longer is there a single clear scale for comparison--and the raw MHz numbers turn out to describe AMD's positioning adequately.

Based on Business Winstone running under Windows 95, AMD positions the 166- and 200-MHz K6 chips as being more powerful than the Pentium MMX at those speeds--a position supported by our tests. Using the same criteria, AMD positions the K6/233 as comparable to the Pentium II/233; according our tests, it comes close, but falls a few percentage points short. It is easily faster than a Pentium MMX/233, however, delivering a Business Winstone 97 score 7 to 9 percent higher than that of the Intel chip. The K6's larger L1 cache enables it to deliver more of a performance boost than the Intel chip as frequency increases.

Where the K6 falls short is on applications that use MMX or floating-point (FP) instructions. AMD did not devote as much silicon to these functions as Intel did. AMD's MMX unit can process only one instruction at a time, for example, while Intel's can handle two.

FP and MMX performance depends on two measures: latency and throughput. Latency is the time from the start of an operation to its completion. It is the defining factor for a single, isolated computation whose results are needed before processing can continue. Throughput is the rate at which new operations can be started; in a pipelined multiplier or FP unit, two or more operations can be overlapped, allowing a faster throughput rate but increasing the latency. All of Intel's CPUs have fully pipelined MMX and FP units, so a new operation can be started with each clock cycle--even though the results may not be available until several cycles later. When performing a long series of computations, as is typical in multimedia applications, throughput is more important than latency.

AMD's K6 has better latency than Intel's processors on some MMX operations, but the throughput for individual operations is the same, and the shorter latency doesn't compensate for the inability to handle two MMX instructions at once. (Intel's MMX unit has only one multiplier and one shifter, though, so it cannot perform both of these critical operations at one time. Also, only one MMX instruction can access memory or the integer registers at a time.) On the FP side, the K6 again has shorter latencies than Intel's processors for some instructions, but it can only start one operation every two cycles, while Intel's chips can start one operation every cycle. As a result, its throughput on many FP operations is half that of the Intel units.

These weaknesses show up as poorer performance on our ZD 3D WinMark 97, synthetic floating point, Pro/Engineer, AutoCAD, and some Photoshop tests. On most of these tests, the K6/233 is much slower than a Pentium II/233 and sometimes even slower than a Pentium MMX/233. On 3D WinMark, for example, the K6/233 delivers only about two-thirds the performance of a Pentium II/233 when using software emulation for all 3-D tasks. With a good 3-D graphics card, this gap shrinks to 18 percent, which is still considerable. Compared with a Pentium MMX/233, the K6/233 is 18 percent short on emulated 3-D, and is still 7 percent short even with a good graphics card.

From the September 23, 1997 issue of PC Magazine

ZDNet Products: DesktopUser
Company Finder: Advanced Micro Devices Inc.