bit-tech.net

AMD could move all ATI GPUs to SOI at 32nm - here's why

Posted on 9th Jun 2009 at 10:58 by Tim Smalley with 7 comments

Tim Smalley
One of the most interesting discussions we had last week was with silicon manufacturing firm Global Foundries, and it talked about its future following its separation from AMD.

Naturally, we talked about process technology and where the industry is heading, and bit-tech came away with a good idea of how Global Foundries hopes to succeed where AMD couldn’t, becoming a leader in silicon manufacturing technology.

Ever since AMD acquired ATI, the new combined company has talked at length about its plans to integrate the GPU onto the CPU die. This decision has been greeted with polarised reactions; Nvidia argues that the integrated GPUs will be ‘outdated’ while AMD and Intel, which later announced its intention to do the same, say it’s the natural progression of technology as CPUs becomes more parallel and GPUs become more general purpose.

Nvidia’s argument does hold some weight if it steps up its push for new technology in the integrated GPU arena. At the moment, we haven’t seen a great deal of that though, as the GeForce 9400M has virtually the same capabilities as the GeForce 8200, which was first introduced back in January 2007 at CES. Products didn’t follow until later that year, but IGP performance hasn’t moved forward massively since then.

With that in mind, process technology is moving fast enough for these new CPUs with integrated GPUs to keep up with Nvidia’s current efforts because CPUs are being refreshed regularly enough for technology and performance improvements to be introduced on a yearly basis.

When AMD moves the GPU onto the CPU die, the company will build the GPU with the same 'Silicon on Insulator' (SOI) process technology that the CPU will be manufactured on. There simply isn’t another outcome if AMD wants to truly integrate the GPU onto the CPU die – you can’t design part of a die using one process technology and the other part using a different one unless they are different pieces of silicon.

SOI has been used for AMD’s CPUs since they were built on a 130nm process many moons ago, and it’s unlikely that the company will switch away from this any time soon – at least, not without a completely new CPU architecture.

Thomas Sonderman, Vice President of Manufacturing Systems Technology at Global Foundries, explained to us how moving to Silicon on Insulator requires a completely different mindset and the decision to move to SOI needs to be made very early on in the design phase. Essentially, you can’t design a piece of silicon for a bulk silicon process and then move it to SOI at a later date – the design changes required would be fundamental and it’d be less costly to start again with a clean slate.

Based on Sonderman's comments, it’s quite clear to us that AMD will move its integrated GPUs to an SOI process when its first Fusion product is released (providing it's truly integrated of course). Moreover, if one GPU in the product family requires a complete re-think, AMD is unlikely to make exceptions for just that chip.

This is especially true when you consider that the architecture work needed for the high-end GPU in a particular generation is carried down into the more mainstream designs – it makes little sense to completely re-architect a chip from the ground up when the hard work has already been done. It’s also incredibly costly too, and for that reason we wouldn’t be surprised if AMD moved its entire family of GPUs to an SOI process in the same timeframe as Fusion’s release.

The theory lines up with some of the organisational changes that have happened inside AMD recently. From what we understand, there are no longer separate GPU and CPU design teams – they all under the Products division within AMD, which is headed by Rick Bergman. There's also the Advanced Technology group, led by Chekib Akrout, which will have taken AMD's most prolific forward thinkers into a separate group who dictate the future technological direction of the company.

As for what this will mean for consumers, it’s not completely clear at the moment, but one thing we can expect is that if the company adopts SOI for its GPUs, it's likely we'll see higher clock speeds than what AMD’s current GPUs ship at today. This should lead to more performance at similar price points, but ultimately the upper end of performance is going to be limited by power budget. With that said though, SOI can help to improve power consumption at the transistor level and that could give AMD a distinct advantage over Nvidia if it chooses to continue using a bulk technology.

Intel always combated AMD with more advanced silicon technology and although it did use a similar technology to SOI, known as TeraHertz, until recently, the chip giant has found that other silicon manufacturing advances have helped it more. Intel is ultimately the benchmark in the semiconductor manufacturing world - the comparison between the process technologies used for AMD and Intel CPUs and ATI and Nvidia GPUs is not as simple as it might seem because up until now (and for at least the next generation), both companies have used similar process technology.

If AMD was to move ATI GPUs to a different process technology with its own set of benefits, it could put Nvidia at a disadvantage because the graphics industry has always been a power struggle. Being on the right process technology is paramount if, as a manufacturer, you’re going to strike the right balance between performance, yield, power consumption and profits because while you can compete with an older process technology, you’re always at a slight price disadvantage due to the smaller number of chips you're going to get from a wafer.

7 Comments

Discuss in the forums Reply
GFC 9th June 2009, 12:16 Quote
Mm.. 3ghz gpus that overclock to 4ghz.. here we come! ^^
Aracos 9th June 2009, 20:15 Quote
So basically I would be getting a GPU that's a CPU but only used for graphics? Not the best example but a Phenom II that's built for graphics?
pdemon 9th June 2009, 21:55 Quote
Eh, not quite. It would be more like taking a 4870 and building it with the same tools & materials as a Phenom II. Architecture and process are at different levels in design hierarchy. Admittedly the process does effect the overall design and certainly causes some changes in layout but it does not determine the function of a device.
salesman 10th June 2009, 00:56 Quote
This was an excellent read Tim. I throughly enjoy reading stuff like this.
Slyr7.62 10th June 2009, 05:36 Quote
Quote:
Originally Posted by salesman
This was an excellent read Tim. I throughly enjoy reading stuff like this.
Same here. Thnx Tim, this was quite an interesting article. ;)
alpaca 10th June 2009, 13:27 Quote
Quote:
Originally Posted by Slyr7.62
Quote:
Originally Posted by salesman
...
... ;)
i do agree!
confusis 11th June 2009, 01:06 Quote
ATI/AMD are hopefully back on the up :)
Log in

You are not logged in, please login with your forum account below. If you don't already have an account please register to start contributing.



Discuss in the forums
Fractal Design Node 804 Review

Fractal Design Node 804 Review

17th September 2014

SK Hynix SH910A SSD 256GB Review

SK Hynix SH910A SSD 256GB Review

16th September 2014

1200W PSU Roundup 2014

1200W PSU Roundup 2014

15th September 2014