Walking Into A Mirror

The most significant development this year in the competition between Intel and AMD seems to be that the days of moving in different directions are over. K6 was not so different from PIII. However, when AMD moved to K7 and Intel moved to P4, the architectures diverged dramatically. As AMD continued with K8 and Intel moved to Prescott, the architectures became even more different. By 2005, in terms of both architecture and corporate structure, there was little overlap between Intel and AMD. This year, with the introduction of Core 2 Duo Intel has pulled its architecture back in the direction of K8. Likewise, AMD is moving in Intel's direction. What we will see through 2007 and 2008 will be the convergence of the two companies both architecturally and structurally.

In 2005, AMD and Intel were as different as they had ever been since AMD started making X86 processors. AMD had only a single 200mm FAB compared to Intel's many 300mm FABs. And, having divested of Spansion, AMD was strictly processors with a few server chipsets in contrast to Intel's diverse interests including its large chipset and embedded graphics business. In terms of cpu architecture, Intel had a long pipeline, virtual HyperThreading, high clock with low IPC design. AMD's short pipeline, dual core, low clock with high IPC design was almost completely the opposite. Intel's MCM Smithfield procesors were a small step in AMD's direction. However, in 2006 with C2D Intel moved to a short pipeline, high IPC, true dual core design and dropped HyperThreading. The architectures are still different but not nearly as different as they were last year and they are still converging.

Just as Intel moved closer to AMD by shortening the pipeline and increasing IPC while extending the architecture to 64 bits so too AMD nows moves in Intel's direction. Core 2 Duo has a dedicated stack unit and AMD moves closer to this with its sideband stack optimizer. AMD improves instruction reordering and out of order loads moving it closer to Intel. AMD now has data dependent latency on the divide instruction as Intel had already done. AMD stretches SSE width to a full 128 bits doubling the speed of SSE operations. It doubles the bandwidth of Prefetch and the L1 bus. Again, these are changes that Intel had made from Yonah (Core Duo) to C2D. AMD might have a small advantage here due to twin L1 buses rather than just a wider L1 bus.

Intel's architecture has been better with DDR2 because of the larger cache. AMD is now moving to a split memory controller where each half can act independently which is what Intel has now with two memory channels. AMD also increases cache size by adding shared L3.This along with improved memory scheduling, a better write burst mode, and out of order loads more than makes up for the higher latency of DDR2. We should see a big improvement for AMD with DDR2 and into DDR3. This again, moves AMD up to where Intel is now.

AMD also moves in Intel's direction by adding some inclusive functionality to its cache while also adding shared cache as Intel has now with C2D. AMD is expanding the number of FastPath instructions. This is important because the decoders can do 3 FastPath instructions or 1 complex instruction per cycle. The more instructions that are fastpath the better. This again makes up for some of the improvement that Intel gained by adding a fourth decoder and macro-ops fusion.

Overall, AMD moves much closer to Intel's C2D with the changes in K8L. AMD will add to its already greater memory access speed and match or beat C2D in terms of SSE. However, the changes don't quite add up to C2D's 4 instruction issue and Macro-ops fusion. I would guess that AMD will remove about 3/4 of the current IPC gap. This would mean that a K8L of 3.1 Ghz would match a C2D of 3.0 Ghz. This is much better than today where it would take a 3.5 Ghz K8 to match the same 3.0 Ghz C2D.

Both AMD and Intel use independent L1. AMD's independent L2 plus shared L3 should be pretty close to Intel's large shared L2. The main differences today are that Intel's C2D still does not allow I/O access above 32 bits and I'm not certain that C2D supports addressing beyond 36 bits. AMD has the advantage with an onboard memory controller and point to point interface. It looks though like Intel will defintely move in this direction with its own onboard memory controller and CSI point to point interface on a new socket in early 2009. AMD has also given indications that it is trying to shorten Intel's smaller process lead.

Structurally, AMD has taken big steps in Intel's direction by creating a 300mm FAB and purchasing ATI. ATI gives AMD a directly competitive position with Intel in terms of chipsets and embedded graphics. AMD will now be able to satisfy the needs of customers who want a complete solution from one manufacture as Intel has provided for some time. AMD will now compete with Intel directly in both embedded graphics for notebooks and embedded graphics for corporate customers. At the same time, Intel has begun divesting of underperforming sections of the company and trimming down toward the main cpu business closer to where AMD already is. AMD's single 300mm FAB greatly boosts its production capacity while reducing cost. Upgrading FAB 30 to 300mm will continue to move AMD in Intel's direction.

There are probably AMD fans who are hoping that K8L will allow AMD to leap ahead of Intel's C2D in 2007. Similarly, I imagine there are Intel fans who are hoping that Intel's move to 45nm will allow Intel to stay ahead of AMD. However, neither of these scenarios seems likely. K8L should be pretty even with C2D architecturally rather than a leap ahead. For process technology the rumors abound with stories of 4.0Ghz clock speeds for Intel while decreasing power draw. These same rumors often limit AMD to as little as 2.9 Ghz putting AMD one full Ghz behind Intel. This would put Intel significantly ahead of AMD if true but it doesn't seem likely. AMD shares process technology with IBM. It is certainly in IBM's interest to be competitive in terms of process. So, it seems likely that if something were lacking that would allow Intel to gain significant advantage IBM would find another company with additional process technology that could be licensed. I'm reminded of the many rumors that Intel would have 3.2Ghz C2D chips at launch but these are still unreleased. I'm also reminded that after Revision F of K8 was released there were claims that it was actually slower than Revision E. It seems that optimism tends to follow Intel while pessimism follows AMD regardless of what each company actually does.

I expect AMD to continue to slowly gain share at the rate of about 0.66% per quarter. I expect AMD and Intel to be roughly equal on the desktop in Q3 07. Likely Intel will have to address AMD's new dual socket FX once AMD has quad core chips for these. If Intel did not address this area they would have nothing competitive. It also looks like Intel will finally have a 4-way server C2D offering in the same timeframe. However, just as AMD's quad FX won't reach its full potential until AMD has quad core FX's, Intel's 4-way server chipset won't really pull Intel even with AMD. Clearly, this will be a stopgap until Intel can release a new socket with integrated memory controller in 2009. However, Intel has no choice but to offer a stopgap chipset since the P4 Xeon will be completely outclassed by K8L. Intel's processor will be fine; it will be the chipset that is stopgap. Intel's quad FSB northbridge will be slower and more expensive than the equivalent Opteron 4-way but it should be close enough to stop Intel's 4-way server share loss. Intel will probably get a nice boost with 45nm technology that will give it an edge on the desktop late in 2007 but AMD will follow with its own 45nm technology by mid 2008. However, any advantage that Intel might have will tend to be countered by AMD's increased competition with embedded graphics. But this won't be the only change. Increased revenue gives AMD more money for R&D and AMD's advantage with APM will come into its own once AMD has two 300mm FABs. My guess is that Intel will also be taking steps to try to match APM and will probably have to do a second reorganization in 2008. AMD for its part may begin construction of a third FAB, possibly after mid 2007.

We know that Intel will move to an Integrated Memory Controller moving it towards AMD. It would seem reasonable that AMD must be looking into Intel's 4 instruction issue and macro-ops fusion to gain future speed. We also know that Intel will eventually use SOI as AMD has been doing for years now. By 2009 AMD and Intel should look much more similar than they did in 2005. I imagine the two will continue to converge as well go into 2010. For example, Intel is moving to a 2 year core design cycle while AMD is moving to a modular core. Intel's approach is less efficient than AMD's however Intel has more engineers to throw at the problem so I would imagine the results would roughly match. If 4 instruction issue really is a good idea then I would expect AMD to add this perhaps in late 2009. There will be a slight difference in process with Intel using high K dielectrics. I don't see any problem with this. Intel only needs to find a high K material with a suitable bandgap for this to work. The only area that I'm not so sure about is Intel's tri-Gate design. From what each company has said it looks like the tri-Gate will not be able to be shrunk while AMD's finFet will be. Still, one would assume that if the tri-Gate can't be shrunk that Intel would follow AMD and use a 3D dual gate like AMD's finFet. If AMD does build a third FAB the X86 processor market should turn into a true duopoly with much more stable prices and production. The big question is whether the market will reach duopoly status with just two 300mm FABs at AMD. However, it seems a certainty that into the foreseeable future AMD and Intel will continue to converge.