AMD Aims To Take Over x86 Leadership

Pages: 1 2 3

Who is AMD Trying to Sledgehammer?

In his presentation at Microprocessor Forum, Fred Weber (VP of engineering, in AMD’s Computational Products Group) made it quite clear that AMD intends to take on existing high end RISC processors and Intel’s new IA-64 product line with x86-64. They have addressed the two biggest disadvantages the current x86 ISA has – lack of 64-bit addressing, and disadvantageous floating point performance architecture. Through conveniently selective comparisons he tried to show that AMD’s K7 family had caught up to the fastest RISCs in integer performance. Mr. Weber claimed that the new architecture will allow future AMD processors to close the gap in floating performance.

Apparently the extension to 64 bits will come without too much pain either. AMD claims that stretching 32 bit x86 to x86-64 only adds 5% to processor die size, and since the same processor core executes 32 and 64-bit code the performance will be equivalent. This might be a bit of a stretch since research conducted at DEC indicates that in practice compiling an application with 64-bit addressing instead of 32-bit addressing causes a performance loss of about a 4 – 5% on average. This results from address data taking up more room in the data cache and main memory causing the cache and TLB miss rates to increase.

Conclusion: You Can Teach an Old Dog New Tricks

There is no doubt about it, AMD is taking on a risky and ambitious course of action extending the ancient x86 architecture once more, this time to 64 bits. Even if they can deliver the new x86-64 instruction set in future products without impact to schedule or performance their destiny is in someone else’s hands, namely Bill Gates. If Microsoft refuses to support x86-64 with tools, applications, and operating systems then x86-64 will simply fail.

In today’s computing world, however, a 64 bit version of x86 might just be the magic bullet that Microsoft needs to offset Intel’s growing influence and rebalance the power relationship in the Wintel oligarchy. It potentially provides another platform to support a 64-bit version of Windows. It also represents an insurance policy in case Intel has bet on the wrong technological horse in the way it designed IA-64.

From AMD’s point of view they had no choice but to develop x86-64 if they wanted to play in the high-end server and workstation marketplace. Any attempt to clone IA-64 would land AMD in the middle of a legal minefield stretching as far as the eye can see in all directions. The only other alternative would be to develop an analog to the IA-64. That is, a split personality MPU that supports both the 32-bit x86 ISA and a 64-bit RISC ISA. It is likely that Compaq would have licensed the Alpha ISA to AMD for use in a hypothetical bilingual processor if asked. But this approach is both more complicated than simply building a 64 bit x86 extension and would likely run afoul of Intel patents related to dual instruction set capability in IA-64.

Based upon the successful development and roll out of the K7 Athlon processor, it is obvious that AMD is in the top tier of processor designers (although the damage from the recent departures of key personnel could spell trouble down the road). Depending on the details, the new ISA should allow AMD to get closer to the FP performance levels of competing IA-64 and RISC processors. They are unlikely to close the gap completely for either integer or FP performance as they claim – the x86 baggage is just too heavy. However, as history has shown us, x86 doesn’t have to be as fast as RISC processors in order to retain the PC desktop market. It just has to stay within a factor of two or so and the software legacy does the rest.

With its x86-64 extension solving the 64 bit addressing issue, AMD should be able to ride the x86 market for at least another 4 or 5 years. Since Intel is unlikely to respond with a 64-bit x86 design of its own (and risk sabotaging market acceptance of IA-64) it seems that AMD is destined to take over the x86 leadership role.


Pages: « Prev  1 2 3  

Be the first to discuss this article!