20nm GPUs unlikely in 2014 due to TSMC issues

Scorpus

Posts: 2,156   +238
Staff member

It's looking increasingly unlikely that either AMD or Nvidia will be able to release a new graphics card this year with a die manufactured using a 20nm process. Issues are aplenty at TSMC, the foundry that produces both AMD and Nvidia GPUs, with the company struggling to meet production demands and repeatedly delaying their final 20nm technology.

Nvidia were originally hoping to release high-end 'Maxwell'-based cards manufactured on a 20nm process this year, but TSMC hasn't been able to meet their end of the bargain. It appears Nvidia has had to re-engineer their GM204 and GM206 Maxwell GPUs for the existing 28nm process, so we may not see these products until December 2014.

Previous rumors surrounding the first GM104-based product, the GeForce GTX 880, suggested there would be 7.9 billion transistors on the GPU die and 3,200 CUDA cores, but having to revert to 28nm may change all of this.

Meanwhile over at AMD, senior vice president Lisa Su has stated that the company will continue to use the 28nm process for the remainder of 2014, before moving to 20nm and FinFET sometime down the track. This will likely mean that Pirate Islands, the successor to AMD's current products, will feature 28nm GPUs.

Intel, on the other hand, aren't struggling with their manufacturing technology as it's all done in-house. Current 'Haswell' CPUs are produced on a 22nm process, and there have been few issues progressing to 14nm for the upcoming launch of 'Broadwell'.

Permalink to story.

 
Disappointing, but oh well things do happen and what with chips on an ever shrinking die its just making it more and more difficult for them to keep up.

Not like this is to big a deal of course, either way the next generation will still be a good improvement all around anyway.
 
The same manufacturer for two biggest competitors - bad news for the market and for the end user. It stinks with stagnant technology and likely price fixation.
 
It's a nuisance really, we're sure to get stripped down models but unfortunately prices won't be stripped to match.
 
How about Nvidia asking intel's foundry (make partnership) to produce their new-20nm-Maxwell GPUs? :D
 
How about Nvidia asking intel's foundry (make partnership) to produce their new-20nm-Maxwell GPUs? :D

Better yet, 14nm :D

It's a nice thought, and may one day be in Intel's model, but for now they'd prefer to sell you their advanced technology one chip at a time. I suppose if the CPU market dries up at some point Intel could always pimp out their FABs and still make a killing.
 
How about Nvidia asking intel's foundry (make partnership) to produce their new-20nm-Maxwell GPUs? :D
Conflict of interest. Intel's Xeon Phi math co-processors compete in the same high profile GPGPU markets as Nvidia's Tesla, and to a lesser degree AMD's FirePro and Sky
The same manufacturer for two biggest competitors - bad news for the market and for the end user. It stinks with stagnant technology and likely price fixation.
Oddly enough, TSMC is a paragon of virtue amongst the sea of corruption that is Chinese/Taiwanese business. I think they are one of the few companies that actively disassociate themselves from institutional bribery.
On the other hand, they are the largest semiconductor foundry business around so do get to set their prices in somewhat of a vacuum, although from what I gather, wafer prices are in line with other fabs (~$3000-3500 for a 28nm wafer, ~$6000 for a 20/16nm wafer initially which will drop by ~25% as the process matures if other process nodes are any indication). Globalfoundries (and possibly UMC) signing up for Samsung's FinFET process licenses should ensure a little more competition and availability relatively soon.
 
Last edited:
@DBZ
Morals and 'business interests' doesn't always go hand in hand (especially when significant financial gains are involved), and to that extent I'll give western weapons manufacturers an edge since they are pretty 'good' at making the 'right deals', recent examples: remember bofors & augusta submarines etc.
 
so, that means nvidia will skip maxwell :D
No, just that Maxwell might be split over 28nm and 16nm. The first Maxwell parts are already available (the GM 107 based desktop GTX 750/750 Ti, and 850M/860M mobile, and GM 108 based mobile 830M/840M).

Nvidia's performance segment boards (GTX 770/760) will be a little long in the tooth by the time the Christmas holiday season rolls around - especially when you consider that they are minor tweaks of a series that debuted in March 2012, so it makes sense to push out a replacement on a known/proven process. The Chinese New Year production shutdown means that if you don't get product out in Nov-Jan, then you're pretty much stuck until March.

It also makes sense (if you can afford it and have wafer allocation) to run the architecture at 28nm since it becomes a relatively low risk option to optically shrink the design for a smaller node while incorporating any revisions into the metal layers. The big GPU successor to GK 110 (Titan/GTX 780/K6000/K40) will certainly be made on the smaller 16nm FinFET process.
 
Back