Don’t get too enamoured with your fancy, long term plans…
Process technology is hard. It only seems like Intel can get it right, and it can only do it right because it has more money and more Fab space than anyone else in the world. For guys like AMD and NVIDIA, they are at the mercy of lower funded 3rd party Fabs who struggle to keep up with the giant. Sometimes their delays hit the bottom lines of their customers… like what is happening today.Process technology is the best friend, and worst enemy of the graphics world. Throughout the years we have seen improvements in process tech create new opportunities within the 3D graphics industry, and at the same time we have seen it be the downfall of certain products and cycles. Today we see that in full effect, and for a variety of causes.
Looking back through the years we have seen some great advances, and some huge stumbles. Consider NVIDIA and their flawed GeForce FX design on a “not quite ready for primetime” 130 nm process. The original GeForce FX 5800 was the pariah of the industry, all the while ATI released their Radeon 9700 Pro on the older, more mature 150 nm process to great fanfare. But we have also seen massive improvements from one generation to another. Take the Radeon HD 2800 XT, which was a hot and power hungry card using a specialty 80 nm process to achieve high speeds. It was not a big seller, and it had plenty of issues. Revisit ATI about eight months later and we see the 55 nm HD 3870, which matches (and improves upon the HD 2800 XT) in terms of performance, but comes in around ½ of the power and significantly lower heat production.
Who can forget the GeForce FX 5800 Ultra? A bad combination of unwieldy design and a still teething process technology.
The latest generation of products has also benefited from, and has been cursed by, the latest process improvements. AMD first dipped into the 40 nm process from TSMC with a smaller, less complex part which eventually made up the HD 4770 product. It was a small order, yields were not great, but AMD learned a great deal from it. AMD was able to take these lessons and apply them to their HD 5000 series of parts, which allowed them to get more than adequate yields and launch a series of products in mass quantities that left NVIDIA in the dust. NVIDIA came to 40 nm much later than AMD, and had their Fermi GF100 design quite far along before they learned about the issues that TSMC was having with their 40 nm process, and the seemingly uncontrollable variations that led to non-working chips. Eventually NVIDIA did produce some smaller 40 nm parts, and we assume they learned many of the same lessons, but these could not be applied to the GF100 generation of parts in time.
Now seemingly things have settled down. Somewhat. TSMC has resolved most of the issues with their 40 nm process, and it can be considered mature now. There are still some tricky things with it, but nothing that cannot be resolved with design changes. But we are running into a big problem now. While the 40 nm process is all well and good, the next step in process technology is late. Really late.
Several years back TSMC promised 32 nm to its customers in an early 2010 timeframe. Unfortunately, TSMC was unable to deliver adequate volume of 40 nm production until Q1 2010. TSMC knew this was going to be a problem several years ago, and in 2008 they changed their plans. There was not going to be a 32 nm process, and it was going to go directly to 28 nm. This is not unheard of, as TSMC went directly from a half node (55 nm) to another half node (40 nm), and skipping the more traditional “full node” 45 nm completely.
The ATI Radeon 9700 Pro, on the other hand, was a success due to a very forward looking design and a reliance on a mature and stable process.
Unfortunately for TSMC’s customers, they were already assuming that 32 nm was going to be the big jump (and it is), and centered their design plans on that node. GPU designs start at least 2 to 3 years in advance of production, so both NVIDIA and AMD had a lot of work already in place by the time TSMC announced that they would be skipping 32 nm altogether. Not only that, but they were pushing 28 nm out at least a year from their original estimates.
This had several repercussions. First of all, and most obviously, a lot of the work done by AMD and NVIDIA was in some ways wasted. While design rules can be changed, and EDA software can somewhat quickly recompile aspects of the design to work on the new process, it is not an insignificant task to rework a design for a new process. Materials, timings, transistor speeds, and a plethora of other criteria regarding the designs change even when going a half node down.
The inside of GLOBALFOUNDRIES Fab 1. Scratch picture to sniff.
Perhaps the biggest inconvenience for these companies is what to do with their release schedule? AMD had to make some radical changes to their plans going forward. They were expecting to be in 32 nm right now with their next generation of products (Northern Islands), but obviously these had to be changed to 28 nm. So instead, AMD rearranged some design groups and gave the go ahead for the “Southern Islands” series of chips, which would take aspects of Northern Islands, combined with the current 5000 series basics, and have a new “half” generation product which will flesh out their plans for the next year or so. The first product from this redesign is “Barts”, which now inhabit the HD 6800 series of cards. “Cayman” is the big-daddy update, and it is going to be one of the largest chips that AMD has released to date.
NVIDIA changed around their plans as well. They have redesigned their Fermi chips, and now we are seeing the GTX 500 series starting to roll out. A lot of housekeeping was done to these chips, as well as some new functionality when it comes to texturing. Otherwise, it is most apt to call it, “Fermi done right.” NVIDIA will continue to redesign chips on the 40 nm process for the foreseeable future, and we will see the 500 series further fleshed out going through next year.
So what is next? Well, both TSMC and GLOBALFOUNDRIES will have 28 nm on line by early 2012. TSMC is actually hoping for production to happen in Q2 2011, but their track record is not exactly spotless. So this potentially means that we will not see anything come out at 28 nm for some time yet. AMD expects GF to have some risk production in Q3/Q4 2011, but there will be no actual 28 nm products until 2012 per Dirk Meyer in their latest analyst meeting.
Um, yeah GF, we would like to talk about your schedule for process releases…
So in the meantime, things are business as usual. Both companies were expecting a performance and feature boost when 28 nm comes online, but it is certainly later than most everyone had planned for. Instead, we are seeing new designs on the current 40 nm process, and will continue to see derivative parts being introduced over the next six to seven months. Then it will be the buildup to the 28 nm generation. So, we really do not expect to see a big jump in features and performance throughout this next year. Cayman and GTX 580 will essentially rule the roost as the top performing single chip solutions. When we consider that GLOBALFOUNDRIES has yet to ship a 32 nm part, and does not expect to do so until Q2 2011, the likelihood of some fantastic breakthrough on 28 nm will allow production in 2011 is very dim.