AMD versus Intel: CPU wars roadmap

Intel announced the next revolution in Microprocessor fabrication by replacing Silicon Dioxide gates with High-K dielectric metal gates.  Microprocessor industry analyst David Kanter who contributed a lot of the information in this piece had this in-depth analysis on what this means for Intel and AMD.  I spoke with David Kanter of Real World Technologies and came up with the following roadmap of the two chip giants for the next two years.

Intel made some large gains with its new Core 2 architecture introduced in 2006 and reversed its downward spiral with the aging NetBurst architecture.  At the end of 2006 we saw the release of the first quad-core CPUs from Intel for the desktop (QX line of products) and server market (Clovertown) which gave Intel a huge advantage in both markets.  Intel was able to gain key design wins such as Sun and Google while AMD was forced to cut prices and profits to the bone to maintain market share.

According to Kanter, the big reason for AMD's problem was Intel's manufacturing lead.  While Intel was producing 65 nm processors all of last year, AMD was outputting 90 nm parts and barely got their first 65 nm product out by the end of the year.  The significance of this is that 90 nm manufacturing with small 200 mm wafers produces less than half the number of wafers compared to a 65 nm process on 300 mm wafers.  That means that while the price wars of 2006 put a damper on Intel's profits, it tore in to AMD financials because of higher chip fabrication costs.

Next: AMD hits 65 nm, honeymoon in Barcelona ...

AMD hits 65 nm, honeymoon in Barcelona

Good news came for AMD at the end of 2006 when the company announced its new main stream 65 nm products using its new Fab 36 facility which uses large 300 mm wafers.  This took a huge load off AMD's manufacturing expenses and it also means that AMD will be able to keep up with the new demand from its new partners like Dell in the high-volume market.  But this didn't solve their performance deficit since the new 65 nm process was only a die shrink and not a redesign or clock increase.  Even the desperate attempt for a symbolic victory in the PC gaming enthusiast market with the AMD Quad FX failed to deliver a victory.  Server competitiveness in the 1 to 4 socket server market was another area AMD lost ground in.  Not only was Intel able to produce faster solutions for the 1 to 2 socket market, Intel could offer a quad-core solution in a 2 socket configuration that almost rivaled AMD's 4 socket dual-core servers at half the price.  Server licensing is an even bigger problem for AMD since many enterprise software vendors charge by the number of sockets and it's not uncommon for them to charge $40,000 per socket.

Barcelona will be a big sigh of relief for AMD since it will be not only bring AMD to the quad-core era but it will also be a nice boost in performance primarily because of the move to a 128-bit floating point processor.  There has been widespread confusion in media reports as to the amount of improvement in Barcelona because some of the key details were left out and quotations of AMD Barcelona clobbering Intel by 40% have run wild on the Internet.  According to David Kanter's in-depth report, here's what was left out:

David Kanter:  AMD has claimed an advantage based on performance models, which are extremely accurate but may not account for faster speed grades from Intel, of around 10-15% for TPC-C and 40% for SPECfp_rate.

So we're primarily talking about a significant boost in floating point performance which is important for scientific work loads based on a future 2007 Q3 AMD product against a 2006 Q4 product from Intel.  Of course no major upgrades are expected from Intel until the end of the year, but that doesn't mean Intel can't squeeze out a small clock boost.  With the TPC-C numbers AMD is boasting, it is possible that AMD Barcelona can take a small but solid lead in most cases in the quad-core server market.  This should give AMD a great sigh of relief as it becomes competitive in the server space again, at least until Intel's Penryn comes out at the end of the year.  If both AMD and Intel execute on time with their next generation of products, AMD's Barcelona honeymoon should last 5 to 6 months.

The high-end desktop market on the other hand won't be so easy to conquer because it's dominated by higher-clocked dual-core processors and Barcelona is actually a reduced-clock quad-core processor.  The performance gap between Intel and AMD on desktop performance is also much greater.  The desktop enthusiast market is also one that likes to overclock processors to squeeze as much as performance out of their CPUs as possible and that's one area AMD can't compete in right now.  The Intel Core 2 Duo products have an additional 40 to 50 percent to overclock while the AMD Athlons X2 and FX products have a peak of 10 to 15 overclocking potential and they're significantly slower to begin with.  The Barcelona class processors may at least neutralize the clock-for-clock advantage Intel held with Core 2 but it's doubtful AMD will make up the overclocking potential if the current batch of 65 nm Athlons which exhibit no additional room to boost clock speed are any indicator.

Next: Intel strikes back ...

Intel strikes back

As we approach the end of this year, Intel will make a huge leap to the High-K metal gate 45 nm "Penryn" processor.  The hafnium-based dielectric and mystery metal electrode allows Intel to drastically reduce gate power leakage.  This translates to lower power consumption at the same frequency or same power consumption at higher frequencies.  The estimated clock speed boost could be 20% or more depending on how Intel executes its transition to 45 nm.  Penryn will also include a 50% boost in L2 cache and roughly 50 more instructions in SSE4 for high performance computing and media applications.  Intel will also continue to use a dual-die process for quad-core Penryn chips because it doesn't really negatively impact power consumption or performance but it does improve their yield because of the ability to pair off good dies.  Penryn if executed properly will likely mean that Intel will retake the lead in servers again and open up a wider gap in Desktops which could spell trouble for AMD.  It could also open up another round of price wars where Intel can better afford to cut prices with a 45 nm process while AMD gets squeezed hard financially with their 65 nm process. 

AMD of course is well aware of this which is why they're so aggressively targeting mid-2008 for their own 45 nm process launch.  IBM, Toshiba, Sony, and AMD are all co-developing a 45 nm process.  IBM indicated over the weekend that they're also planning to incorporate High-K dielectrics as if to say "me too and we're not too far behind" after Intel's big Friday night High-K announcement so it's reasonable to assume that AMD will also be using High-K in their 45 nm process as well.  I asked David Kanter what he thought of AMD's aggressive mid-2008 time frame and here's what he had to say:

David Kanter:  Mid-2008 is a very aggressive target for AMD, since they will be shifting to immersion lithography. It would be a real victory for AMD, but a bit of healthy skepticism is due.  The other thing is, that it's possible they hit mid-08 but in really really low volumes.

Kanter also added that immersion lithography is a brand new chip fabrication process that no one has ever done before and AMD has had major delay issues in the past.  Intel on the other hand is deciding to stick with existing dry lithography techniques and won't attempt immersion lithography until they go to their 32 nm process.  The fact that they were already showing off 45 nm prototypes last Friday would seem to indicate they made the right decision.

If getting to 45 nm wasn't a big enough challenge for AMD, Intel isn't in to mood to let AMD catch up.  Intel will release a whole new microarchitecture called Nehalem some time in the second half of 2008 which will likely be around the same time as AMD hitting 45 nm.  Nehalem will be Intel's second generation of 45 nm chips and it will not only feature a brand new microarchitecture, it will also replace Intel's aging FSB (Front Side Bus) technology with CSI (Common System Interface) which will compete with AMD's HyperTransport architecture.  While Nehalem will be a brand new microarchitecture, AMD's first 45 nm part will be a refresh of the Barcelona architecture.  At this point there isn't enough information to speculate on performance in Q4 2008 but the fact that Intel is leaping so quickly after Penryn has to be giving AMD some headaches.