AMD starts gpu production in their fabs in 06?

[arjan de lumens]

I think that's overstating things by about 2x, or at least 50%.

If they went 1T ram, then that's almost a 5x overstatement.

1T RAM = 1 transistor per bit = 440M transistors for 55 MBytes, so he is pretty much correct on the actual transistor count. However, it must be noted that transistors for DRAMs can usually be packed a great deal more densely (about 3-8x) than random-logic transistors, so the die area for the 55MByte DRAM would be much smaller than that of e.g. an R580-class GPU core. (With a dedicated DRAM process, a 64 Mbyte DRAM on 90nm these days takes roughly 71mm2; an eDRAM or 1T-RAM block would presumably be larger but not dramatically so.)

 

[RussSchultz]

me = teh stupid. Forgot x8 bits.

 

[The_Wolf_Who_Cried_Boy]

Any news on whether AMD is bringing any new fabs up in the next 2 years? I recall that they just had one go online in the US this year.

epic

Supposedly another pair of 300mm FABs are on the way by 2008. One in India and one in New York.



As for Z-RAM, I posted this at AMDZone, aren't the colors pretty...

Z-RAM uses 1 transitor per cell and supposedly has all manner of power and speed advantages over other on die memories.

A little light reading for any interested.
http://www2.dac.com/data2/43rd/43ex...544c4/$FILE/ISI Z-RAM overview April 2006.pdf

 

[asicnewbie]

Does anyone have a feel for how ATI's upcoming 65nm technology compares to TSMC's 65nm? TSMC is apparently in production with 65nm already: http://www.tsmc.com/english/b_technology/b01_platform/b010101_65nm.htm

You mean AMD's (not ATI's) upcoming 65nm technology? I don't think its fair to compare a pure-play merchant foundry, who exists solely to sell wafer-capacity to contract buyers, to AMD's in-house foundry. AMD designs and manufactures a limited number of designs, doesn't have to serve any external customers, and completely controls the end-to-end (from architecture to final-silicon) design, production, and testing process. As a merchant foundry, TSMC must multitask its production-lines for thousands of customers, each with its own mask-set and processing requirements. Suffice it to say, each business can excel, but market success (and history) imply each will excel in different areas.

IBM's SOI technology (which AMD licenses) probably delivers better transistor speed (faster switching times + faster interconnect). I say probably, because I am not a process engineer. As far as I'm concerned foundry == place where mask-set + $$$ go in, wafers come out, with absolutely horrible (2+ month) I/O latency SOI does increase per/wafer processing-costs, so AMD is probably paying more per wafer than they would for a non-SOI wafer.

For 65nm, TSMC is emphasing "improved low-power performance" (vs their 90nm G process.) At 90nm, the leakage was pretty terrible. And by terrible, I mean certain operating conditions could shoot leakage through the roof (4-5X worse than TSMC 130nm leakage.) TSMC's strategy is consistent with a merchant foundry serving a broad range of consumer electronic applications. TSMC's strategic direction is a combo of market factors and customer feedback: a) mobile ASIC design-starts are growing faster than non-handheld design-starts, b) after shockingly horrible transition from 130nm -> 90nm, designers have gotten over the "MHz race", and c) the world needs more 12" wafer-fabs because, you know, every startup could be the next ATI/NVidia/Broadcom.

Thinking further ahead, AMD's newfound foundry connections (from ATI) could potentially mean that AMD can de-emphasize investment in their own factories.

Must real men *still* have fabs?

As long as AMD insists on SOI-production, they will have limited choices. Besides their own fab and Chartered, AMD cannot just sign-on with another merchant fab. Likewise, unless Intel decided to license/share their process-tech with a fab-partner, Intel must continue investing in its manufacturing capacity.

Intel does fab a number of (non-CPU) products on third-party foundries (TSMC, I believe).

 

[{Sniping}Waste]

Its posable for AMD to produce GPUs around the end of this year. Alot will need to be done like redue the layout of the GPU to fit AMDs fabs (Fabs from one company to another are not the same and equipment are not the same) but it shouldent take to long with AMD layout ppl helping out in this (AMD layout ppl will know what to change, the areas of the old layout that could cause yeild problems) and then E-Beam the new quartz masks.

 

[JoshMST]

Yeah, I think most people are being too optimistic about AMD producing parts for ATI this close out of the gate. The hurdles for both companies to do such a thing is pretty tremendous.

ATI uses a standard cell design that works on both TSMC and UMC's lines, but even still verification of each product running on the different lines takes up time and money. That standard cell is not designed for AMD's processes, so will most likely not run. AMD processes SOI wafers, so their Fabs are set up to do just that. I am not sure, but you just can't throw standard wafers down those lines and hope that it will work. ATI would have to redesign their GPU's from scratch using AMD specific standard and custom cells, as well as take SOI into account (and if they want 65 nm, then they have to take SiGe into account as well).

On the AMD side, things look even worse. AMD is trying to ramp 65 nm production as we speak, but it is going to take a good year to get Fab 36 to run the max # of wafers through, and AMD is going to dedicate all of its 65 nm process to its Athlon 64 chips. So no room for ATI at 65 nm in Fab 36 for a long while. Then you must consider that once AMD has transitioned 100% over to 65 nm and Fab 36, they are going to convert Fab 30 into Fab 38 to handle 65/45 nm and 300 mm wafers. So, during that conversion that Fab is going to be down, and not available for 90 nm production once AMD closes those lines down. New York is not going to be open until late 2008 at the earliest. By then AMD will have Fab 38 up and running, then perhaps ATI can start to utilize Fab space at Fab 36 to make their parts. So, theoretically, we have two years for ATI to continue using 3rd party foundries.

As the previous poster said, it is awfully hard to compare TSMC's 65 nm process to AMD's. AMD utilizes SiGe, DSL, and SOI to improve per transistor performance, cut down leakage, and improve transmission speed. TSMC looks to be using Low-K to save on power, but does not look to integrate much more to help transistor performance. It looks like TSMC's process is more designed around decreasing die sizes while offering performance slightly higher than their current 90 nm products.

 

[Sunrise]

ATI would have to redesign their GPU's from scratch using AMD specific standard and custom cells, as well as take SOI into account (and if they want 65 nm, then they have to take SiGe into account as well).

What makes it even worse is that business-wise it doesn´t make a lot of sense to build an already well-designed ASIC from scratch using completely different base characteristics for AMD´s FAB, which will take an estimated time of about 3-6 months just for changing the design according to the new rules, but also another 3-6 months for tape-out, prototyping and avoiding any potential hassles with sequential control. So, apart from all material-induced differences, there are a lot of calculations that need to be done here.

You wouldn´t waste that much time on such things if you weren´t desperately in the need for it, which ATi isn´t (they have just recently deepened their relations with UMC), since they´ve planned parts in the pipeline that look at least one and a half year ahead. So, if producing anything in AMD´s own FABs should make any sense at all, this can only mean that they introduce a very sophisticated and (combined) CPU+GPU (vector processor) package, which wouldn´t be similiar to anything we´ve seen to date.

TSMC looks to be using Low-K to save on power, but does not look to integrate much more to help transistor performance. It looks like TSMC's process is more designed around decreasing die sizes while offering performance slightly higher than their current 90 nm products.

TSMC also has 65nm + SOI in the pipeline and some additional enhancements that are power-specific (in comparison to their 90nm Nexsys tech), but that alone doesn´t have to mean anything, because we haven´t seen if SOI will help things or if it makes matters even worse.