Slowing Down the Process Migration

A refocus on engineering

by Josh Walrath

Paving the Way

            To get a good idea of where all of this is going, we need to take a quick look into the past to see where the graphics companies have been before.  NVIDIA has always traditionally worked very closely with their foundry partner TSMC, and NVIDIA spares no dime when it comes to foundry work.  NVIDIA was one of the first to develop a 250 nm part with the TnT 2 (alongside 3dfx and their Voodoo 3), and they followed suite at the 220 nm, 180 nm, and 150 nm nodes.  Each change gave significant die size, transistor performance, and heat optimizations.  These jumps were relatively painless for NVIDIA (though quite a bit of work was still done, and it was not quite as easy as I make it sound), and they thought to continue this trend with the 130 nm node.

            It became evident fairly early that the 130 nm node would be a very large technological jump.  Intel, AMD, and IBM were all having significant problems getting their 130 nm based products to market.  Intel was the first to market with the Tualatin based Pentium !!!’s, but it wasn’t until several months later that it could reasonably produce the Northwood core Pentium 4's.  AMD was even farther behind, and they didn’t release their first 130 nm product until Summer of 2002.  Even then these Thoroughbred based Athlons ran hot and were not significantly faster than the Palomino Athlons made on the 180 nm process.  With these two heavy hitters having problems at the 130 nm node, it only followed that companies like TSMC and UMC would have to take a longer time to develop their processes.

            In the case of NVIDIA and TSMC, this turned out to be true.  Not withstanding some of the significant design problems that NVIDIA had with their NV30 core, the main hurdle for NVIDIA was getting adequate yields and speed bins from TSMC’s process with this complex design.  Multiple tapeouts had to be made for the NV30, and many of the initial chips came back non-functional.  Or in even more frustrating cases the chips came back functional, but after 40 hours of testing the chips suddenly died.  Problems with void migration displacing transistors from the substrate turned out to be one of the biggest problems, and to work around this problem entire design libraries for the TSMC 130 nm process had to be revised.  This of course led to further redesigns of the NV30 core to reflect these changes.

            While the NV30 project at NVIDIA proved to be very costly, it did have some positive effects on NVIDIA.  While the design of the NV30 left quite a bit to be desired, the follow up designs did learn from the hard lessons NVIDIA experienced.  The NV35 and NV36 products were both very robust and much more competitive with ATI’s offerings at the time.  Furthermore, this experience has made NVIDIA much more leery of new process nodes, and I believe that we will see a greater emphasis on overall design rather than early migration to the next node.

            ATI avoided all of this by being very smart about their designs.  ATI has typically let NVIDIA produce parts on the smaller processes for about 6 to 8 months before releasing a product on that process.  In the past this of course kept them behind NVIDIA in terms of competitive products.  The Rage 128 was significantly behind the TnT, and the Rage 128 Pro was very far behind the TnT 2.  ATI did somewhat catch up a bit with the original Radeon, but the GeForce 2 was already in the market for some months.  This again followed that NVIDIA had the GeForce 3 out seven months before ATI released the Radeon 8500 (and of course NVIDIA trumped this release with the original Detonator drivers and the Ti series of GeForce 3 products).

            These products kept ATI in the game, and they continued to have very good OEM and Mac support, even though their retail presence was small compared to NVIDIA.  ATI did know that being on the cutting edge of process development was inherently risky.  Their next step showed exactly how risky it could be.  When ATI acquired ArtX, they gained very valuable engineering experience, as well as a new philosophy on graphics design.  The R300 project was a huge step away from the R200 (Radeon 8500) in terms of quality, functionality, and speed.  When design work started on the R300, TSMC was advertising that its 130 nm process would be up and running in January 2002.  For one reason or another, ATI thought that it would be risky to try to design the R300 on the 130 nm process.  Instead the ArtX engineers decided to utilize TSMC’s 150 nm process.  In hindsight this decision proved to be brilliant.  Upon the release of the Radeon 9700 Pro, ATI proved that a large and fast chip could still be produced on the 150 nm process.

            ATI again reduced their overall risk with new processes by creating smaller designs to be used at the new nodes.  The first example of this was the RV350 (Radeon 9600 Pro) which was the first ATI chip to use the 130 nm process at TSMC.  By using a smaller, less complex design, ATI was able to produce a very fast core that had favorable yields with TSMC.  The knowledge gained by using this type of transition was then applied to more complex and massive designs.

Next: Getting the Most Out of a Process

Copyright 1999-2004 PenStar Systems, LLC.