Graphics chip flirts keep foundries on their toes

IMAGINE THE SHOCK and surprise when ATI came home one night to TSMC with lipstick on its collar, and not the shade that TSMC usually wears.

Instead of going into hysterics and calling its lawyer for a divorce, TSMC sat ATI down, cooked it a fabulous dinner, and partied into the night. Unlike marriage, the relationships in the semiconductor industry are rarely monogamous, and while companies will have favourite partners they are still willing to sleep around.

The analogy above somewhat shows what is going on between UMC and TSMC, while Chartered is aiming to be the third option in this struggle.

UMC may not have the shiny reputation that TSMC has, but it is a very solid foundry partner for hundreds of companies out there. Typically the UMC processes are not as advanced as TSMC's, and for the most part the transistor performance is lower than their primary competition.

This does not necessarily mean it's a bad thing. UMC is offering many of the same basic production lines as TSMC, but because UMC is not pushing the bleeding edge as TSMC does, it can offer these products for cheaper.

UMC is currently offering full production 90nm on 300mm wafers at two Fabs (Fab 12a and UMCi). It is also planning to begin 80nm production very shortly. UMC's 65nm process looks to be some months behind TSMC and, for now, UMC will really be focusing on selling production space for both 90nm and 80nm.

Having a second option for producing complex GPUs is something that both ATI and Nvidia are quite happy with. This tends to keep the primary production partner honest and competitive (in this case TSMC). It also puts pricing pressure on TSMC to offer more value for what production lines they have.

ATI will be releasing 80nm parts this autumn, from both UMC and TSCM in its RV5x0 series of chips. It is unclear what percentage of production goes to UMC, but apparently it is well less than half. Nvidia is also using UMC as a secondary source, and already have limited 90nm production going on with that company.

Nvidia has also been testing the waters of UMC's 80nm process, but no official decision has been made so far to use it for full-scale production later this year. When it comes to foundries, cutting prices yet keeping the lines moving is much more desirable than underutilizing or shutting down a line when there are no customers.

While companies such as ATI and Nvidia enjoy cheap chips, they also have performance levels they need to meet. With TSMC having the faster process, it is a balance between getting enough high speed parts to satisfy that market and getting cheaper but slower chips to address those particular price points. Let me illuminate my point with an example. Let's say that both TSMC and UMC are producing 7600 series parts for Nvidia.

While the slower-clocked 7600 GS is selling more quickly than the higher performance and higher priced 7600 GT, that does not necessarily mean that because the GS sells more that UMC would get more business by producing the lower spec and lower priced part. In fact, it is quite likely the opposite. Due to things such as yields and speed bins, Nvidia needs to have a guaranteed supply of faster parts for the GT SKUs.

This means that TSMC may in fact get the majority of those orders because there is a greater possibility that these chips will speed bin at a higher level than what UMC can provide. So, while Nvidia knows that the majority of UMC chips will at least bin to GS levels, putting more production to TSMC means that a greater portion of those parts will bin at GT levels. It is always much easier to take a faster part and put it in a slower SKU than vice versa.

UMC is setting itself up as a more price oriented supplier, and is not so keen on massaging every ounce of performance out of their processes. This means less R&D expenses, and in the end makes for a more cost effective product. If pure MHz performance is secondary to a design, then placing orders with UMC is a good business decision.

So where exactly is Chartered in this? Chartered is a bit of a black horse when it comes to the well known TSMC and UMC. This does not mean that their production facilities and technology are inferior. Chartered has been shaking up things a bit as of late though with their strong partnership with AMD (and by extension IBM and Infineon). Currently Chartered is producing 90 nm based AMD Athlon 64 processors under a foundry agreement with AMD, and this agreement will continue to 65nm. This means that Chartered has integrated not only AMD's APM software suite into Fab 7, but it also utilizes advanced process features like SOI and DSL.

Assuming that Chartered's 65nm process will mirror that of AMD and IBM, they will have production space based on that technology available for other interested third parties.

Microsoft has already stated that the Xbox CPU will be ported to 65nm with Chartered in Q1 2007. Nvidia already has limited production with Chartered, so that relationship is already in place. ATI is certainly looking at Chartered as well, because the possibility of producing a complex GPU on Chartered's advanced 65nm process might be too sweet to pass up. Sixty-five nanometre products from Chartered are expected to start arriving later this year.

Third-party semis will always pick the sweetest deal, no matter what the past has held for them. So it doesn't matter if Nvidia or ATI comes home with three different shades of lipstick on their collars, they will always get the red carpet treatment when showing up for dinner with a wallet full of money. µ

L'Inqs
UMC's 65nm Process Sees Widespread Acceptance
AMD tech is saving Xbox 360 CPU
AMD And Chartered Sign Sourcing And Manufacturing Technology Agreements