Possible reasons for the R600 delay to Q2

[nonamer]

The "LAST" R600 speculation thread is looking not so "LAST" anymore. I hope the mods don't mind if I post outside of it.

Let's look at the facts. If it is to believed that the R600 is based on the Xenos, then ATI had the tech ready for a very long time. It was only a matter of bringing it to the PC with the addition of DX10 features and a wider bus. All of this should be ready some time in 2006, and it was already hard to believe that it could be delayed to 2007 in the first place. Now it's been delayed yet again and it's looking like a mid-2007 product. That's a total of about 18 months to modify an existing design!

I can't believe ATI could screwup something that badly when they already had a working version of the thing in 2005. Could the merger have disrupted development that badly? Was their a major design flaw or obstacle that we don't know about?

My guess is that the G80 caught them completely off-guard. Some time in early to mid 2006, ATI got wind that the G80 was fully unified shader architecture and had better performance than what they were going to get. Rather than getting blown away, they decided to delay the R600 until they could tweak it until it was competitive. Unfortunately they're still tweaking.

 

[northfieldz]

Yes, obviously ATi/AMD can not handle its own NV30 too well, hence the delay, delay and delay.

I think thats the only reasonable explanation, and anyone who is not fooled by his own fanboyism should has figured the obvious out.

 

[3dilettante]

I'll just copy from my post in the rumors thread.


I suppose there are a few non-technical and non-compelling reasons.
1) AMD's strategy is to have some kind of unified launch of Barcelona and R600.
I don't find this particularly compelling, since it would be paper-launching a delayed server CPU with a delayed GPU lineup.
In the worst case (Barcelona is not significantly better than Kentsfield, R600 no better than G80 or G81), AMD gets to announce that it's an also-ran in two markets at the same time.

2) AMD hates to make money.

3) The late-stage development process was disrupted by the buyout.
I think the corporate side would be able to insulate the engineers from most of this.

4) Any other excuses people can think of?

Technical:

1) Yields are low.
Possible, this is a new design on a new process. This is possibly a factor in the following reasons.

2) Bin splits are bad (possibly due to new process). The chip comes out functional, but the clocks for the majority are too low for the high end.
Perhaps the delay for a new high-end respin would make sense for a family launch. The stockpile of low-clocked R600s could make up a good portion of the lower performance parts.

3) Clocks of the design are lower than desired, which leads to:

3a) The chip is too hot because it is being pushed to the upper edge of its clock envelope.

3b) It is clock-limited by something (if not by heat from 3a).

3c) It needs to be clocked higher.

Reason 3 is consistent with some rumors.
The rumored cooling solutions and the need for more power seems to show silicon at the edge of acceptable power consumption. This may be in part due to the adjustment to a new process, or due to clock and voltage bumps made for the sake of hitting higher clocks.

There isn't much hard evidence R600 isn going scalar like G80. This may indicate that utilization for the vector units is low enough to impact any peak numbers it can reach, so the clock must rise to compensate.

There hasn't been much substantive talk of clock domains. This means ATI is likely trying to push the higher levels scheduling logic along with the ALUs. Even should the ALUs be capable of being pushed faster, the scheduling logic might not be. It should be noted that Nvidia doesn't try to push the full core to high clocks, and keeps things in the slow domain several hundred MHz less than ATI's entire-chip clock.

There are other possible problems. Perhaps the way R600 handles branches is less efficient than G80, which devotes units to keep them from tying up the ALUs. Perhaps the data paths are more complex for R600, which would increase heat and suppress clock rates.
Perhaps the wide memory bus leads to clock issues, though one would think ATI could muster a workable ring bus.
Tweaks to the memory controllers and ring bus design would need a respin.

Maybe a utilization problem is worsened by difficulties with driver development, though it wouldn't necessarily need a hardware respin (if that is what is happening). The time frame of the delay seems to be long enough to allow one more go at the fab.

 

[turtle]

Amongst all other things everyone has posted, which I mostly agree with, I wonder if they are going for a massive platform launch.

Think about it, towards the end of Q2 it's possible we could see dual socket Barcelona (8-core K10) on socket 1207 using a RD790 motherboard, capable of 2x16 or 4x8, which could be used for either crossfire or quadfire (which they seem to pushing) as well as one/two cards + lower-end card (rv610/rv630) for physics (marketing term "APP"). Considering AMD cannot compete with Intel at this moment on the cpu front, intel nor nvidia on the motherboard front, and questionably may not be able to compete with R600 vs nvidia on the gpu front (wither G80, the lower-end, or the refresh) for whatever reason, not to mention there isn't a motherboard currently out there that will allow for their 2+1 setup the R600 is pushing besides the lone DFI RD600 board, waiting and blowing their load all at one time might be a plan to swoon people into upgrading their whole system, or at least demonstrating a massive powerhouse platform at review time for people to pick and choose what works for them. By this time it's possible they could have more CTM apps available (gpu-assisted decoding, physics API?) and could showcase some bad-ass stuff.

Granted, this isn't a good excuse, and perhaps the more probable reason is stockpiling the newest chip revision which obviously would be in short supply come even April, allowing for more SKUs, and waiting for a family launch. None-the-less, if any or all of the other product segments have new products available around this time (CPU, new chipset/platform), I can see why they would wait for better allocation and plan a large multi-product launch.

Sheesh though, obviously something did go wrong somewhere. I'm betting at this moment they are wishing they had planned for the initial R600 to be based on 65nm, as by the time this thing rolls around it'll be time for that process to take over as it will be mature, even on the high-end and for complex chips.

 

[Thorburn]

which could be used for either crossfire or quadfire (which they seem to pushing)

Last time I counted ATX only had 7 expansion cards slots so I can't imagine an R600 QuadFire arrangement (I get the feeling that with rumored power consumption and a single slot cooler they'd probably catch fire a bit).
Still, if they are gonna do it I might buy some shares in PSU manufacturers.....

 

[Geo]

Use the other thread.