SAN FRANCISCO— Taiwan Semiconductor Manufacturing Co. (TSMC) announced Monday (Oct. 24) it had achieved volume production of its 28-nm process with wafers having already shipped to customers.
The foundry said its 28-nm line-up includes 28-nm high performance (28HP), 28-nm high performance low power (28HPL), 28-nm low power (28LP), and 28-nm high performance mobile computing (28HPM), though 28HPM would only go into production at the end of 2011. 28HPM Wafers have, however, already sampled to “most mobile computing customers for their product-design use,” the firm said.
Customers who have already signed on to use TSMC’s 28-nm process include Altera Corp., Advanced Micro Devices Inc. (AMD), Nvidia Corp., Qualcomm Inc. and Xilinx Inc., with speculation abounding that Apple Inc. too may have silently joined the Taiwanese fab’s clientele. In August EE Times reported Apple was bolstering its ties with TSMC in order to distance itself from longtime foundry partner, Samsung Electronics Co. Ltd., after a spate of legal differences between the two.
Sources revealed at the time that TSMC would make Apple’s 40-nm A5 dual-core processor on a foundry basis for the iPad 2 tablet, saying the Cupertino, Calif.-based firm would also work with TSMC on the 28-nm process for its A6 processor. Whether this is the case or not, however, is yet to be confirmed.
Samsung, however, has also been reported to be ramping Apple's A6 on its own 28-nm CMOS process, out of its wafer fab in Austin, Texas, but with the levels of legal animosity rising between the two firms over competing tablet and smartphone products, it’s thought to make sense for Apple to diversify its foundry reliance.
TSMC has also said the 28-nm production is taping out at more than double the rate of its 40-nm predecessor, with over 80 customer product tape-outs already. The foundry says the 28-nm process has also surpassed the previous generation’s production ramps and product yield in the same time-frame, which it attributes to “closer and earlier collaboration with customers.”
The news of healthy yields will be a relief to customers like AMD and Nvidia for whom the previous cancellation of the 32-nm process and the delay achieving stable 28-nm production has had an adverse impact. It has meant that the graphics processors (GPUs) of both firms have had to rely on the 40-nm process for more than two and a half years, with yield issues severely crippling supply in the early stages.
German publication Heise Online recently reported AMD would be first to ship new 28-nm GPUs in December of this year, with Nvidia set to release its new GPU generation on 28-nm in the first quarter of 2012. Both reports are currently unconfirmed. It’s believed, however, that the first batch of GPUs, from both firms, are likely to be lower-end, mainstream cards, which would test out the new 28-nm process before stretching it for higher-end GPUs due out later in 2012.
While TSMC’s senior vice president Jason Chen noted in a press release that his firm was “first to 28-nm volume production,” Globalfoundries spokesman Jason Gorss told EE Times its high-k metal gate (HKMG) offerings had already been in production “for months.”
“Our 28-nm process is ready,” said Gorss, pointing out that Globalfoundries, unlike TSMC, had also produced wafers on the 32 nm process, with HKMG and that 28-nm was simply a shrink of that previous offering. Products on Globalfoundries’ 28-nm process would emerge sometime in 2012, Gorss noted.
AMD’s Llano (A series APU) was produced on Globalfoundries’ 32-nm process and has been on store shelves since June, said Gorss, adding it was the only HKMG product that could currently be found on store shelves. AMD’s long awaited Bulldozer is also made on Globalfoundries’ 32-nm process.
Another TSMC customer commenting on the 28-nm announcement today was FPGA maker Xilinx, which claims to have been shipping its programmable chips on TSMC’s 28-nm process with HKMG to customers since April.
Xilinx’s senior vice president of worldwide quality and new product introductions Vincent Tong said the 28-nm HPL process being used to build the firm’s 7 Series FPGA and processing families had lowered static power by 50 percent, while also increasing both raw performance and usable performance.