TSMC: 45nm ready in September

TSMC has announced that their low-power 45nm process will be entering production in September, with general purpose and high performance variants following sooner rather than later. Traditionally, handheld chipsets will be the first to use a new process as they have much lower yield requirements. FPGAs, as per their highy symmetric nature and their redundancy mechanisms, are also among the first adopters of a given process node. Desktop and laptop GPUs, on the other and, tend to lag behind by 18 to 24 months.

For example, TSMC 65nm production started on the low-power variant in May 2006, but no PC GPU manufactured with 65nm is being sold yet, as far as we are aware - with AMD's RV610 and RV630 likely to be the first, unless NVIDIA finally decides to release the 65nm G72. Such a chip would primarily be aimed at OEMs, though. As for handheld GPUs, we aren't sure about the process of AMD's latest announced products but NVIDIA's only current 65nm chip is their GoForce 5300 with eDRAM, for which the process only became available in late 2006.

It should still be noted that TSMC is definitely closing the gap with Intel, and one very simple way to notice it is that their 45nm process is being made available only 16 months after the 65nm process, which is obviously faster than is traditionally the case. Furthermore, Intel released their first 65nm processors in January 2006, while TSMC only began production in May - this time around, TSMC's production will begin before Intel even ships anything.

While the production ramps are not really comparable, because Intel will ramp on a 107mm² Penryn core with more 40mm² of logic while TSMC's clients will initially be producing much smaller chips, it remains true that TSMC is doing the best it can to close the gap with Intel. On the other hand, Intel will likely be the only logic-focused company in the world which will begin their 45nm process ramp on a high-k/metal gate process. TSMC will only follow on the 32nm node while AMD/IBM are likely to be introducing the technology halfway through the process' lifetime, unless they plan on delaying the ramp to later in H2 2008.

While TSMC's 45nm process is not as advanced in terms of materials (on the positive side of things, that might reduce costs slightly...), they are arguably leading the pack in terms of lithography this time around by adopting immersion litography very aggressively, which is also fairly unsurprising as Burn Lin --TSMC's litography guru-- is actually the father of immersion. AMD/IBM are set to follow and adopt the technology in their own process, while Intel is skipping it for the 45nm node and the foreseeable future.

In the end, it's apparently a design choice, and not much else. Immersion should allow for better scaling (in terms of gate/transistor density and SRAM cell size) than using dry litography like Intel does, but this can partially be compensated by using more complex (and more expensive) masks. On the 32nm node, TSMC is set to continue to use immersion lithography, while Intel seems to still be hoping to adopt EUV (extreme ultraviolet) techniques. Considering TSMC recently pointed out they didn't think EUV would be ready for their 22nm process, however, it would be rather surprising if Intel managed to use such tools more than 2 years before their competitors think they will ready for mass production!