Notes from the AMD June 1st Tech Demo.
by gzader on Comments
I took a few notes on the webcast. I thought I'd share them here. These are all based off of AMD's statements so expect a heavy pro-AMD lean on this, however as this could be considered guidance and as AMD is a publicly traded company, this should be at least reasonable if not exact.
65nm production will be released in Q4 of 2006. This could be as early as September. If you remember the transition from 130nm to 90nm you'll remember that resellers had lots of 130nm parts left in inventory for some time after the transition this was due not only to stock levels but also in that 130nm parts were still being produced. This may not be quite the case as AMD has been working on streamlining their inventory chain. This has resulted in 35% less inventory being stocked and therefore may result in a similar reduction in cross over times. However AMD will be producing 90nm parts till mid-2007. At which point, all production (FABs et. al) should have transitioned to 65nm.
I had the opinion that 65nm versions of most if not all of the AM2 product line would be released. This would make sense as some chips not meeting specifcation X would undoubtably be viable at lower specifications.
As 90nm closes out, 45nm should be in the last stages of pre-production. AMD's belief is that 45nm will begin production in mid-2008. Note that 65nm should phase out in 2009. AMD has accelerated it's time tables on pretty much everything to keep pressure on intel and the various production technologies are of course a large part of that. In the far future, 32nm is slated for early 2010.
Quad core should not be available until early 2007, it's my opinion that they'll ship quads within the same period as intel does if not earlier. Quad cores will be the next PR coup and AMD isn't about to miss it.
One of the more interesting things was AMD's production abilities. With it's contracting with Charter to produce parts AMD no longer feels the need to build new fabs but rather they believe they can now act on a process of continuous upgrade to meet needs with the additional capacity that Charter offers.
Part of AMD's design philosophy has been to continuosly upgrade it's transitor design. Each production run has featured an evolving transitor design that would seem to allow AMD to make a more natural progression from 90nm to 65nm as the transitor itself is slowly produced in a smaller and smaller space. This would explain AMD's ability to squeeze out faster and faster chips from a core that is basically several years old. It would also explain why many of these chips have the same overclocking headroom as their older and slower siblings. AMD believes this philosophy is what is allowing them to transition easily to 65nm.
It's my impression that AMD is currently producing or very shortly will be producing production quality and production quantity 65nm parts.
The quad core will feature a 2MB L3 cache shared between cores. This is smaller than what intel will ship and AMD states the cache is unnesscary when you have a more efficient design. Regardless, the small cache will allow for a small chip size and therefore more chips per wafer which means lower cost per chip and either better selling price or more likely, higher profit margins.
The quad core will get HyperTransport 3, but the release date is more tenative. I would expect the quad to go with current HT now and then later ship with HT3. As HT3 will work in HT1 systems, there is no reason for the chip to not have a mixed bag approach, with HT3 inside and HT1 talking to the outside world.
My favorite item above all this was AMD's new modular chip design philosophy. By breaking down the standard components that go into a chip (Core, Caches, Interfaces, etc) they have the ability to quickly modify their base level chip to meet various needs. So in theory and Quad with HT3 and 2MB of L3 for the very high end could be re-tooled with little effort to be a quad with HT1, no L3 cache and a meger amount of L2 cache. The design time would be in essence negligible but the ability to have a QuadFX and SempronQuad for various markets could drastically shift what is expected from a processor vendor.
AMD believe it's next generation will see a 3x improvement in floating point performance. This is one of the last few performance crowns that intel undeniably holds, with a 3x improvement via wider and faster data paths intel may have an even tougher fight than it imagined.
Finally, the base level technology on the next AMD core allows for a maximum of 8 sockets with 4 cores each before adding 'glue' chips to bridge multiple busses together. This means that you could have a workstation or server with 32 available cores for processing, each with a direct HT bus to ALL of the others. 1 hop to go from any processor to any processor and 1 or 2 (maybe only 1) to go from any core to any core. This puts the Opteron in the same class as heavy duty unix systems. Add to this some off the hot swap options that HT3 has you can see why Sun is moving more hardware to AMD.
Further, if you think about this design, each processor has it's own dedicated memory and memory controller. On the Xeon side of the house you might have a few FSB's to work with but they all have to hand off data between each other. This would seem to mean many more hops and far less data throughput.
AMD is gaining market share and increasing earning while intel is reporting declining revenue and sales. Intel will get a significant bump with Core2's release but how long will the sales and mind share last? It really depends on if AMD lives up to what it says it can do. AMD has a good track record. It made the switch to DDR2 with an performance increase, Intel made the switch with a performance decrease. This is likely more a matter of timing rather than design, however AMD managed to increase performance design a processor that thrives on ultra low latency, something DDR2 simply doesn't have.
It remains to be seen what will happen, but should be an interesting time watching.
Log in to comment