MS to combine CPU and GPU at 65nm?

Status
Not open for further replies.

Shifty Geezer

uber-Troll!
Moderator
Legend
Scooby quoted this info in the NPD thread :
At the heart of a Valhalla Xbox 360 is a combined 65nm CPU and 65nm GPU on a unified super chip. This also would only need one cooling system and be much quieter and hopefully more reliable than previous Xbox systems (if they include hardware JTAG testing reporting into the design).

Combined 65 nm parts sounds rather optimistic to me. It'll be a big chip with a complex construction method as the eDRAM will need to be on package too. If the eDRAM is present on die in the 65nm Xenos, that's 337 million transistors in one die. Xenon is 165 million. That's nigh on 500 million transistors, the size of a quad-core Pentium. Yields will be lower and there's no market for seconds, so I expect the price to be relatively high. Is it really economical to combine the processors at this level? Previous combo chips like Sony's have always been a good lot of process advancements from the originals so clearly other manufacturers haven't thought an early combination chip is a good move.

Or is this far more likely bunk reporting via the 21st Century's monster child, Rush-Journalism, and in reality the combined chip design isn't targeted for 65 nm but 45nm where we'd expect it?
 
Valhalla sounds like 45nm to me. It makes sense that if they want to combine them, they do it when they shrink the process. It makes more sense than engineering combined chips, then shrinking them. Look at the gpu side. ATI/Nvidia often throw a couple of extra features in when the shrink their dies. They don't usually just add them on their own.
 
Valhalla sounds like 45nm to me. It makes sense that if they want to combine them, they do it when they shrink the process. It makes more sense than engineering combined chips, then shrinking them. Look at the gpu side. ATI/Nvidia often throw a couple of extra features in when the shrink their dies. They don't usually just add them on their own.
Exactly what I was about to say ;)

I guess somebody got something wrong while rumoring this, lack of understanding?
 
IMO that source is unreliable. An insider information that guy previously got was backed up by a MS ship-it award belonging to the alleged source. The problem is - those awards ended around 2002 - way before 360 shipped. How can that guy be considered insider in 2007/2008 if he seems to have no involvement with Xbox team for several years now?
 
The intel core 2 extreme's have almost 600 million at 65nm so its not unprecedented size. The quad core extreme's push over 800 million on 45nm. Admittedly those have a big chunk of cache taking up a lot of the space.

I don't find it that likely to occur myself, but that's not to say MS isn't investigating the idea as it would certainly be a huge coup in terms of cost reduction if they could make it happen.
 
Xbox 360 was designed from the outset to eventually have the CPU and GPU combined into one chip.

It's mentioned in the book 'The Xbox 360 Uncloaked'. I'll dig up the quote later.
 
The intel core 2 extreme's have almost 600 million at 65nm so its not unprecedented size.
Right, but they're also not cheap, which I guess is in fair part to yields. And that's with a market for seconds too. I dunno what cost to manufacture versus retail markup for performance is like, but I can't imagine production is cheap.
I don't find it that likely to occur myself, but that's not to say MS isn't investigating the idea as it would certainly be a huge coup in terms of cost reduction if they could make it happen.
It's probably safe to say it's intended for 45nm, but as an exercise, what are the problems? Is 500 MT at 65nm impossible in the price/heat envelope (bearing in mind > half of that is GPU to boot)? What's the formula for determining when a die becomes economical to be a combo-chip?
 
Well, probably the first step is a integration of edram and Xenos on the same die, and why not Ana also, seem more logical now with the Hdmi to integrate it in the GPU,no?
They can win some silicon and place on the motherboard?
 
Maybe, but what gain would that bring?

Combined chips on a 45 nm process may be possible, but wouldn´t we see a 65 nm GPU before that?

A single cooling solution? I guess it's just poorly worded in the original article (which isn't surprising since it's a blog). That it's [implied to be] still very early in the development stage, it probably is a combined 45nm project. At 65nm it doesn't sound like a cost reduction act.
 
Doesn't really confirm anything, but this seems to indicate that MS is up to something.


MS is leaving the console arena! :runaway:

Anyhow...

Shifty Geezer said:
It's probably safe to say it's intended for 45nm, but as an exercise, what are the problems? Is 500 MT at 65nm impossible in the price/heat envelope (bearing in mind > half of that is GPU to boot)? What's the formula for determining when a die becomes economical to be a combo-chip?

For one thing, there will be fairly significant differences in thermal density during operation, and no doubt the multiple clock-speeds around the chip will cause even more headaches (XeCPU already has multiple frequency operation).

From a manufacturing standpoint, they'd be going from three production lines to one for use in a single Xbox 360, but then they'd better hope they get somewhat decent yields at 45nm for that to be of any advantage. Right now they have production at Chartered (and IBM?) for the XeCPU, and TSMC for the GPU & eDRAM. If they can get a decent size reduction at 65nm, they could combine those latter two production lines into one for the GPU and get double production capacity.

IIRC, people believe the GPU is on 80nm, but after seeing that RealWorldTech chart, it seems more likely that the size difference is down to the difference in equipment capability or engineering, particularly between NEC and TSMC's eDRAM process. Any miniscule difference per transistor would be magnified by the hundreds of millions per chip and lead to minor die sizes.

With a combined chip, they would want to either keep or add more redundancy as well, and perhaps that will play a role in what the chip layout ends up being (just try cutting and pasting the three dice together right now :p).
 
Yes.

Many circuits designed for non-SOI are either suboptimal or simply do not work if ported to SOI, and the same applies in the other direction.

The GPU would have to be reimplemented on the new process.
 
If GPU+CPU+eDRAM is about 500million transistors then why is it not possible? G94 is 500M transistors and 225 mm2 which is only slightly larger than a 90nm xenos.
 
Status
Not open for further replies.
Back
Top