: News
: Reviews
: Editorials

: Archives
: About
: Contact
: Advertising
: Privacy
: Links
: Forums

AMD Q3-2005

And Why some Analysts May Need to do Their Homework

by Josh Walrath

With the switch to 65 nm, it is now much more cost effective to introduce dual core products to both mobile and mainstream markets.  Once AMD starts to introduce these new products, we will see introductory dual core prices hit in the low $200's US, while sub $200 will still be single core.  AMD will continue to have fast single core processors for their FX series until at least late Summer 2006.  In January we expect the jump to 3.0 GHz on a single core Athlon 64, while all of the other single core Athlon 64 products will be silently replaced by 65 nm parts as needed.  Most likely though, most single core Athlon 64's will still be 90 nm parts as AMD will continue to utilize Fab 30.  AMD has essentially stated that for desktop single core parts, other than the new FX-59, there will be no new products.  This means that AMD will likely focus on dual core for the mainstream and mobile, and leave the Sempron parts as single core for the time being.  Going to 65 nm will also give AMD another good advance in the mobile market by offering very low power Athlon 64 processors in both single and dual core.  AMD will be very competitive with Intel in the mobile sector with these parts, and because the memory controller is already integrated into the processor, not to mention how relatively easy it is to implement the HyperTransport connection, the size and cost of a notebook motherboard should be smaller than a competing Intel based product with similar specifications.  In terms of TDP, both the next generation of Pentium M and Athlon 64 should be on very even footing.

Another thing to consider is that AMD will finally be adopting DDR-2, just as it is starting to move to faster speeds.  The latency penalty that DDR-2 has over DDR will be mainly overcome by the official introduction of DDR-2 667 and DDR-2 800 speeds.  Not only that, but since the memory controller on the Athlon 64 is integrated, it has a greater advantage in terms of single and double data-word access than even the fastest Intel based memory controller located off of the CPU.  AMD will have redesigned their memory controller with the 65 nm chips to handle DDR-2 effectively, and will include features in that controller that will offset whatever latency penalty DDR-2 imposes.  This change will come with the new socket implementations AMD will introduce in the beginning of 2006.  This will allow the Athlon 64 architecture to move to DDR-2 just as it truly is becoming mainstream, which means lower priced and better performing modules.  The new sockets will also allow AMD greater headroom in power and data delivery, not to mention effectively doubling the current bandwidth the Athlon 64 has with PC 3200 DDR.

By the time Intel releases its next generation 65 nm part based on the Conroe core, AMD will have its second generation 65 nm part ready and most likely shipping.  This is where the analyst at UBS is really wrong.  In this now famous article, the analysts at UBS expect Intel to "leapfrog" over AMD with their next generation Conroe based chips (which includes Merom, Conroe, and Woodcrest).  I think they have taken Intel market relations a little too literally here without looking at AMD and their previous track record.  Yes, the Conroe based parts will probably be faster than the current Athlon 64's on a per clock basis, but that is assuming that AMD stands still for the next year.  That is not going to happen.  Conroe is still going to be stuck with two very large limitations: it will still be based on the enhanced GTL+ front side bus, and it will have to rely on an external chip to provide main memory access.  Now, all of this involves latencies as well as complex design considerations.  As memory gets faster and faster, its efficiency will decrease overall because the connection to the processor from the memory controller is not as scalable as one would like to think.  Intel can design a lot of their architecture to hide these latencies and bandwidth limitations, but the very fact of the matter is that it is not nearly as scalable or efficient as an integrated memory controller on the CPU.  To get efficiencies and speed up, Intel needs to push the aging GTL+ bus to where it probably was never designed to go.  To get the FSB to go as fast as it needs to be, motherboards will have to be more complex and electrically sound to make this work.  This will increase manufacturing costs, which is something board makers and OEM's dislike, not to mention retail customers.

AMD has already beat this problem by including the memory controller into their CPU's as well as instituting a scalable HyperTransport link from the CPU to the rest of the system.  16 to 32 lanes of HyperTransport running at 1 GHz effective is much easier to design and implement into a motherboard than a GTL+ FSB running 64 bits at 1066 MHz and above.  Not only that, but scaling the processors from 1P to 2P and 4P is a lot simpler in terms of board design, and has a higher efficiency factor than what Intel can do even with their next generation server products that include two separate FSB's to increase scaling.  Once processors get to 4P and 8P, AMD has a dramatic edge in scaling efficiencies.  With each processor able to address its own memory controller, as well as the huge overall bandwidth that HyperTransport affords with three HT connections per chip currently (not to mention the overall scaling that AMD's NUMA- Non-Uniform Memory Architecture gives for 2P and above).  Once we consider the underlying architecture, as well as the upcoming advancements that AMD will integrate into their processor designs, I highly doubt that Intel will "leapfrog" AMD.  On the desktop I think that Intel may have a slight performance advantage with their new design, but once we get to servers that advantage will go away due to Intel's reliance on an underlying architecture that was first introduced with the Pentium Pro.  When it all washes out, AMD and Intel will have a lot of parity on the desktop, but if AMD does introduce something special with their 65 nm 2nd generation part, then Intel may still be at a distinct disadvantage.

Next: More Intel

Copyright 1999-2005 PenStar Systems, LLC.