But, although Altera, Qualcomm and Nvidia have publicly complained about lack if 28nm supply from TSMC, Xilinx has said it has no issues.
“We have all the capacity we want and yields are ahead of plan,” senior Xilinx vp Victor Peng tells me.
He points out that Xilinx uses TSMC’s HPL process which is not used by any of the companies which have complained about lack of 28nm supply.
Curiously Peng says that Xilinx has shipped “1,000s of units.” Thousands seems rather small bearing in mind that Xilinx started sampling 28nm over a year ago.
Asked if the company was comfortable with the supply situation for the future, Peng replied: “We’re comfortable with what we’ve got. We have what we need right now.”
The most virulent complainant has been Qualcomm which said it was actively looking for multiple alternative suppliers. Qualcomm CEO Paul Jacobs did not blame yields but cited lack of capacity for constrained supplies which impacted Q1 revenues.
Altera has made no criticism of yields but said it could not get enough supply to meet an up-tick in demand late in Q1.
However Altera, which has always exclusively used TSMC, is not going to add to its fab sources nor go back to the jointly-owned fab model in Washington State which it used to have with TSMC.
Asked if going back to the shared fab model, or adopting a multiple foundry strategy, might be in Altera’s planning, Altera’s senior vp for military, industrial and computing operations, Jeff Waters, tells me: “One of the benefits of sticking to TSMC is the early access we get to new processes, another is the long-term support we get for our process technologies. If you’re hopping around from fab to fab, from foundry to foundry, it’s difficult to get a foundry to support you. And TSMC gives us support.”
Jen-Hsun Huang of Nvidia is the only CEO to publicly accuse TSMC of having poor yields on 28nm and also announce that he’s off to Samsung to get 28nm fab. However some people have pointed out that Nvidia’s yield problem could be down to design flaws.
No one expects process transitions to go smoothly, but there had been a hope after the horrors of 40nm, that lessons learnt during that node would translate into a better experience at 28nm.
The jury is still out.