|AMD Q2 2004|
|by Josh Walrath|
Further Clarifications on AMD's Recent Announcements
Damon from AMD was kind enough to give me a call concerning a few questions I had. There were several things left unanswered by the dual core press release, as well as the Sempron release, so I thought I would post a few confirmations here.
AMD is on schedule to release 90 nm Athlon 64's in a variety of products by the second half of this year. The first product we will see are the Opteron and Athlon FX variants, followed by notebook and then mainstream desktop. The Opteron/Athlon FX core will be approximately 114 mm square (I am assuming this has the 1 MB of L2 cache). The lower power Oakwood (mobile) will be around 80 mm square. This chip will most likely have 512 K of L2, and possibly a single memory controller. These are the only two confirmed sizes for the 90 nm lineup as of yet. Going to the 90 nm process will give AMD a hefty boost when it comes to producing larger quantities of its Athlon 64 processor. On 200 mm wafers, going to 90 nm will give them approximately a 75% boost in die count. Current Athlon 64 processors with 1 MB of cache are around 193 mm square, and the Newcastle core with 512 K L2 is somewhere around 150 mm square.
Fab 36 is well on its way towards completion, and it looks as though AMD will start 65 nm test wafers in mid-2005. As you may know, AMD recently held a ceremony at Fab 36 celebrating the fact that it is now fully enclosed and they can start work on the inside (placing machinery, setting up the lines, etc.). The jump to 90 nm was a great one for the entire industry, but I think that AMD may be overly optimistic here in saying that they will have 65 nm working by mid-2005. However, AMD did confirm that it will have 65 nm products in 2006 (though no more exact dates within that time frame). Initial full-scale production is not scheduled until mid-2006 though, so this can give them some time to work out the kinks in the process. Also, Fab 36 is based on 300 mm wafer production, which more than doubles the surface area for die production.
On to dual core! These dual core processors will be socket compatible with current Opterons. The die size for these dual core chips will be slightly larger than current Athlon 64's on the 90 nm process (approximately 210 mm square vs. 193 mm square). These will first be introduced as an Opteron class processor, but when AMD gets closer to moving to 65 nm production, AMD will transition the dual core architecture to the high end and upper midrange desktop space. Each core will most likely have its own memory controller, and most likely will rely on the current NUMA (Non-Unified Memory Architecture) specification that currently powers dual and quad Opteron processors today. Of course the interconnect between the two cores will be significantly faster than the current HyperTransport connection that MP Opterons use today. This is a true answer to Intel's HyperThreading technology, and in fact will show greater levels of performance than the H-T technology (because it truly is two cores as compared to the intelligent scheduling featured with H-T).
Heat and power have been the trademarks of Intel's Prescott based Pentium 4, and AMD is trying to sidestep this situation with their Athlon 64 design. Prescott was designed to achieve high clock speeds, while still retaining nearly the same IPC as the Northwood based P4. To do this though, Intel had to lengthen the pipeline to 31 stages, but to cover the increased latencies and possible pipeline stalls, Intel had to add twice as much cache (both Trace and L2), as well as greatly improve the Branch Prediction Unit. This has led to a 125 million transistor processor that occupies around 110 mm square of die on their 90 nm process. There are two problems with the design though: heat and power. To get it to run at 3.6 GHz and above, the processor eats up more than 103 watts. Unfortunately, the die size is now much smaller, so it cannot spread the heat out across a larger area. This has made cooling a major problem (not to mention power supply).
AMD on the other hand has gone with a more balanced approach with the Athlon 64. With the integrated dual memory controller, memory bandwidth is not a problem. Add to that the severely decreased latencies involved with main memory accesses, the improved branch predictor, larger TLB's, and the large L1 and L2 caches, the Athlon 64's pipeline is kept busy most of the time. There are very few cycles wasted due to branch misprediction or data access latencies. In the face of this architecture (and other reasons) Intel has abandoned future P4 cores and will eventually transition its server and desktop processors to the Pentium M architecture (a significantly upgraded Pentium !!!). This is not a bad move on Intel's part, as the P4 architecture (while very robust with good performance) just brings too many negatives to the table. AMD's design decisions with the Athlon 64 architecture has allowed it to migrate to the 90 nm node and increase clock speeds while still staying within the thermal and power envelope of the current 130 nm Athlon 64's. Further experience and advancements with the 90 nm SOI process at AMD will lead to better power and thermal properties for these processors.
Still, the most impressive thing about AMD's announcement is that it has actually taped out this dual core processor. Further redesigns will be done, but AMD expects first samples to be ready by Q1 of 2005, with full production in Q2 2005. By summer of 2005 AMD will have the first x86 dual core processor. Intel is not a dumb company, and they are also designing a dual core processor, but it is nowhere close to being taped out. Prescott will be their main processor for both desktop and server, but many more resources will be placed into the Pentium M design teams, as that appears to be Intel's future (and a good one considering how well Centrino has done in the mobile market).
AMD will most likely support DDR-II, but not until it is well established in the marketplace. DDR-II does add a good amount of latency to memory transactions, but with the current dual channel memory DDR-I controller in the Athlon 64's, there is no real need at this time to push DDR-II introduction. DDR-II 400 is in fact slower than DDR-I 400. It is not until DDR-II reaches 667 and 800 will it really begin to eclipse current 2:2:2:6 PC3200. There is also some talk about a next generation DDR based technology that could see introduction around mid-2005. DDR-II could be a technology that AMD can afford to miss, but this is a story in the very early stages of development.
AMD and PCI-Express
In the past AMD has released chipsets that pave the way towards next generation technologies. They have done this with the original AMD 750, which gave the first support for the Athlon processor. Later they did this with the AMD-760 and its support for DDR memory (and for a long time was the dominant Athlon/DDR chipset). This was followed by the AMD-760 MP, and then the current AMD-8000 series of chipsets that support the Opteron. At the moment AMD does not see the need to release a AMD-8000 based chip that supports PCI-Express. Currently many of the 3rd party chipset manufacturers have PCI-E/Athlon 64 chipsets nearing completion, and AMD expects to see production based motherboards supporting PCI-E by Q3 of this year (if not closer to the end of this Summer). So, while PCI-E will take a while to become mainstream, AMD will only lag Intel by about 3 months in terms of shipping products. AMD expects to have PCI-E boards by the time PCI-E video cards hit the market in force.
AMD's announcement of the Sempron is really nothing more than solidifying the market segmentation for their products. The high end will be Opteron, the mid-range will be Athlon-64, and the budget level of performance will be showcased by Sempron. There will be several lines of Semprons based on both the Athlon XP core with the Socket A infrastructure, and then moving onto 256 K L2 Athlon 64's on the Socket 754 infrastructure. This will replace the Duron name as the budget offering from AMD. Nothing terribly exciting, but a good way to segment and define the different markets for AMD.
Thanks again to Damon for getting back to me in such a prompt manner!
Copyright 1999-2004 PenStar Systems, LLC.