Forthcoming ATI GPUs: Different Processes, Different Suppliers, Hybrid Architecture? [UPDATED]

ATI’s Next Plans: Multi-Fab Manufacturing, Hybrid Architectures

by Anton Shilov
04/01/2010 | 05:14 PM

UPDATE: Adding assumptions regarding performance levels of Southern Islands chips.


The problems with 40nm fabrication technology of Taiwan Semiconductor Manufacturing Company have definitely caught leading fabless chip designers unaware and that caused shortages of certain graphics processors. But there was something that was even worse than current problems: TSMC decided to change its roadmap and it looks like the change of the contract maker’s plans has substantially affected roadmaps of its clients.

This week it transpired, if the market rumours are to be believed, that ATI, graphics business unit of Advanced Micro Devices, had to tangibly reconsider its plans and instead of launching code-named Northern Islands series of graphics processing units (GPUs) this year, the firm will instead release code-named Southern Islands chips in order to refresh the graphics lineup later this year in a bid to offer a breed of processors that offer higher performance than today’s ATI Radeon HD 5000-series code-named Evergreen. The family known as Southern Islands is now claimed to feature architectural elements of both Evergreen and Northern Islands.

Northern Islands and TSMC’s 32nm

The Northern Islands graphics processors were supposed to be made using 32nm bulk process technology, according to AMD’s own slides demonstrated on the 31st of October, 2009, in Japan. Most likely, the Northern Island chips were meant to be made at TSMC. Using 32nm TSMC process technology was a rather logical decision, however, that process was downplayed by TSMC from the very beginning in August 2008 and designing for it essentially was a risk.

There were many talks about half a decade ago regarding usage of high-k metal gate technology (HKMG) with fabrication processes below 45nm; but TSMC decided not to use HKMG in its 32nm fabrication process and keep utilizing strained silicon, silicon oxynitride (SiON) as ultra low-k inter-layer dielectric, copper interconnects and so on. In fact, 32nm fabrication process was generally a shrink for the 40nm/45nm manufacturing technology and was supposed to be ready for mass production in late 2009. Instead of fusing advantages into 32nm fabrication process, TSMC has been concentrating on 28nm with and without HKMG and considered it to be the true next-generation technology.

Considering that TSMC’s 32nm fabrication process was an upgrade to the 40nm tech, it is logical to assume that it would have faced the same issues the 40nm did and, perhaps, in even more dramatic degree. Nevertheless, TSMC did not seem to be worried and in August, 2009, TSMC's board approved capital appropriations of $1116.8 million to “expand 45nm process capacity and install 32nm process capacity”, giving a clear signal to its customers to get ready for 32nm process technology. 

Not surprising that up until at least the end of October last ATI was convinced (perhaps, not completely) about TSMC’s ability to ramp up production of GPUs using 32nm bulk process technology in 2010.

However, already in January, 2010, during the topping ceremony of fab12/phase 5 building in Hsinchu Science Park, which was supposed to produce 40nm, 32nm and eventually 28nm chips, TSMC said that the phase 5 was expected to begin “volume production in the third quarter of this year to satisfy urgent recent increases in customer demand”. Considering that there are no chips built using TSMC’s 32nm process technology announced, “customer demand” term most probably refers to 40nm/45nm production and it may be evidence that TSMC has canned 32nm process technology. Obviously, any products that ATI might have designed for TSMC’s 32nm fabrication tech are now also canned. 

Unnamed Chips and Globalfoundries’ 32nm

It should be noted that when AMD span off its manufacturing assets into a separate company, the chip designer and the Foundry Company (the name of Globalfoundries back then) signed an agreement, under which once the latter establishes a 32nm-qualified process, AMD “will purchase from Foundry Company sales entities, where competitive, specified percentages of its graphics processor unit requirements at all process nodes, which percentage will increase linearly over a five-year period”.

At Computex 2009 in the middle of last year the contract maker of semiconductors based in Sunnyvale, California, demonstrated wafers with test chips produced using 32nm fabrication process. Therefore, AMD’s ATI unit probably could weigh all pros and cons of the new manufacturing technology early enough to make design decisions. There is a big problem though: Globalfoundries has also shelved 32nm bulk fabrication process and ATI probably had to further alter its plans.

Do Not Put All Eggs into Same Basket

TSMC’s problems with 40nm fabrication process have probably demonstrated everyone that reliance on just one contract manufacturer is a huge risk. At present both ATI and its arch-rival Nvidia produce the vast majority of their chips at TSMC as well as certain volume at United Microelectronics Corp. (UMC).

The problem with UMC is that it is lagging seriously behind TSMC in terms of adoption of new technologies. In order to stay on the leading edge of fabrication processes, fabless semiconductor makers have to stay with TSMC and/or go with Globalfoundries, the company that inherited AMD’s tradition to implement new fabrication processes as quickly as possible and ensure high yields and rapid transition.

Considering the fact that ATI suffered quite heavily from TSMC’s 40nm fabrication process and the fact that the graphics chip designer has to make GPUs at Globalfoundries’ fabs, it is very likely that ATI had to split its graphics chips portfolio between the two foundries. Perhaps, Northern Islands were supposed to be made at TSMC, whereas Southern Islands were supposed to be made at Globalfoundries.

However, when TSMC decided to scrap its plans for 32nm fabrication process and Globalfoundries followed, ATI had to majorly reconsider its own roadmap and whatever had been planned before had to be scraped then.

A Bird in the Hand Is Worth Two in the Bush

Both Globalfoundries and TSMC promise that their 28nm high-performance (28HP) process technologies will be available for mass production already in late 2010. But there are chances that TSMC’s process technology will arrive somewhat later than expected, whereas Globalfoundries does not have a track record of quick and flawless ramp up of bulk fabrication processes.

As a result, according to currently available information, the GPU family called Southern Islands will be made using 40nm process technology, which is still not problem free (after the initiation of pre-production in early Q4 2008), but will at least be available in the second half of this year. A bird in the hand is worth two in the bush: TSMC’s 40nm is here, whereas two 28nm fabrication processes are still in the bush, it seems.

Nonetheless, AMD is obliged to acquire graphics processors from Globalfoundries; as a result, when it comes to 28nm process technology, the company will still have to divide its GPU portfolio. It will not be an easy task since Globalfoundries’ and TSMC’s 28nm bulk processes are completely different: the former uses so-called gate-first approach, whereas the latter utilizes gate-last manufacturing method. As a result, it is more than likely that starting from 28nm node there will several different companies producing ATI GPUs.

ATI’s Next-Generation GPUs: Bridging Today and Tomorrow

If the rumours are to be believed, Southern Islands graphics processors will be hybrids that will contain elements of Evergreen as well as elements of the actual next-generation hardware (some call it Northern Islands). In fact, gradual implementation of new features without substantial change in the architecture or rather tangible change in the architecture without implementation of new features are traditional for ATI.

The graphics chip designer implemented a number of important elements of its DirectX 10/11-generation hardware into its ATI R520/R580-generation chips, e.g. ultra-threaded dispatch processor, texture units and render back end were removed from the graphics pipeline itself and so on.

ATI R600 architecture was so forward looking that even the latest ATI Evergreen (R800) chips feature the same general design along with many elements first introduced three years ago.

All in all, it looks like ATI’s next-generation graphics processors are indeed supposed to bridge today and tomorrow. According to reports, the new Southern Islands GPUs will be made using today’s 40nm process technology, will feature current-generation 5-way VLIW stream processors, but will also contain the next-generation ultra-threading engine, memory controllers, caches and other parts.

At present we know only three code-names – Cozumel, Ibiza and Kauai – that presumably belong to the Southern Islands family. This seems somewhat surprising as the Evergreen features four chips and at least six code-names for different products. What we should expect from ATI with the Southern Islands is substantially boosted performance.

Early next year AMD launches its code-named Llano accelerated processing unit (APU, CPU with integrated GPU), which is projected to have about 480 stream processors in its best possible configuration aimed at mainstream systems. Potentially, Llano may easily offer higher or equal computing performance than Redwood (Radeon HD 5600, 400 stream processors). As a result, ATI will have to refresh entry-level and mainstream lineup because otherwise Llano will dramatically affect sales of entry-level solutions like Cedar (Radeon 5400, 80 stream processors), which performance is not much higher than that of ATI Radeon HD 2400-series, a graphics solution from 2007. Once ATI introduces more powerful discrete graphics chips than Cedar, they will automatically have to replace Redwood, Juniper (ATI Radeon HD 5700, 800 stream processors), Cypress (ATI Radeon HD 5800, up to 1600 stream processors) with higher performance offerings. Therefore, at least three new higher-performance desktop Southern Islands chips in the next twelve months are inevitable and four are likely. What is unclear at this point is ATI’s plan for the mobile lineup code-named Vancouver since at present even code-names of GPUs from the family are unknown.

Will these solutions will be truly fast (compared to products that will exist on the market in the launch time-frame) and as power-efficient as the ATI Evergreen? It is rather obvious that the Southern Islands chips sizes will grow bigger and will highly likely consume more than predecessors simply due to usage of current process technology. Still, only time will tell for sure.

ATI did not comment on the news-story.