LONDON – Microprocessor vendor Advanced Micro Devices Inc. has decided to cancel APUs that Globalfoundries Inc. was set to make for it on 28-nm process technology, according to online reports.
Instead AMD (Sunnyvale, Calif.) will start afresh using the 28-nm gate-last high-k metal-gate manufacturing process technology from alternative foundry Taiwan Semiconductor Manufacturing Co. Ltd. (Hsinchu, Taiwan), the reports said.
It is speculated that the reason for the move is that the 28-nm process at Globalfoundries (Milpitas, Calif.) will only be ready for volume production in mid-2012, which would give the planned processors only six months in the market. An alternative view is that low-yields on the 28-nm process have been at the root of failed negotiations on the manufacturing contract.
If true, the move would be a major blow to Globalfoundries, which was originally created by the spinning off of AMD manufacturing facilities. The plan was to use AMD has an "anchor" customer while ramping up a number of others in foundry mode.
However, AMD has been moving its manufacturing towards TSMC for some time. A change of management at Globalfoundries, implemented in the summer, was said to be related to problems with 32-nm chip yield which had also led to a rewritten supply contract so that AMD only paid for good die.
The risk to Globalfoundries is that it could lose its main customer before it has attracted enough other customers to fill its fabs. Meanwhile TSMC, the world's largest foundry, is a strong competitor, and high volume chipmaker Samsung is also seeking to get into the foundry business and all this at a time when most industry watchers are predicting 2012 to be low-growth year for the chip industry.
I think then we agree more bins would be better, right? So foundry customers should hopefully start using speed bins explicitly in their product line definitions, just like Intel CPUs. The end user would be less unclear about what he/she is getting.
I see Xilinx has 4 speed grades.
Otherwise the customer needs to accept a wide GHz range.
And if the die size is too large, the same number of lost dies would be a larger yield loss.
It is the customer's responsibility in the end.
But lack of customer consistency would mean different yields per customer.
@Chipmonk: some good points you raise in your analysis. Foundry utilization details was something GloFo never discussed in their technology day this fall. But they did make a point about diversification including AMS & MEMS business than digital / memory business dominancy.
Here is the link. And if you read the articles from Semiaccurate and Extremetech than you will see that it was no more than a (wrong) assumption that Glofo is intended to produce the chip. But this was not the case and now all this allegations have no foundation any more. The truth is totally different.
AMD simply decided to let produce TSMC more Radeon GPUs with high margins instead of low margin APUs because
the 40nm Bobacats are still selling very well. It has nothing to do with Glofo. Please, get it!
The story is partly wrong and exaggerated. Already months ago taiwanese media such as digitimes and commercial times reported that the chips are to be produced by TSMC and not Glofo. While Glofo is already busy with Llano and Trinity chips, TSMC also has not so much capacaties in 28nm for all its customers. Therefore, AMD is obiously shifting to more profitable graphic gpus. All in all, the story has little to do with Glofo or problems at AMD but with high demand for hardly available 28 nm products from TSMC.
David Patterson, known for his pioneering research that led to RAID, clusters and more, is part of a team at UC Berkeley that recently made its RISC-V processor architecture an open source hardware offering. We talk with Patterson and one of his colleagues behind the effort about the opportunities they see, what new kinds of designs they hope to enable and what it means for today’s commercial processor giants such as Intel, ARM and Imagination Technologies.