ATI RV740 review/preview

[Arun]

Wouldn't the logical thing to do is to start mass production of 40nm parts while minimizing production of 55nm while that inventory keeps trickling down, and introduce 40nm with a hard launch once 55nm inventory is at the desired level?

Of course, if yields are still quite low and improving very rapidly, then there's good reason to delay 40nm production as much as you can - which clearly complicates matters greatly... I'm not sure they'll be improving so rapidly (on the matter of just one month or two) that it'd matter as much as you could naively assume, though.

As for AMD ramping wafers, I'd *guess* it's mostly RV740/RV790/RV730 at this point, with most of it being the first. I would suspect that they cut production more than consumer sales were really reduced (i.e. at or below the level of inventory replenishments in the supply chain which were clearly insufficient), so now they need to increase production back to meet the real channel sell-through (i.e. consumer spending).

 

[rjc]

Wouldn't the logical thing to do is to start mass production of 40nm parts while minimizing production of 55nm while that inventory keeps trickling down, and introduce 40nm with a hard launch once 55nm inventory is at the desired level?

I think thats roughly what they are doing for their 2 other parts the 32shader/64 bit one and the 64 shader/128bit one. Both those have cost advantages on 40nm. Its a harder proposition for the new G92 as chip is presently roughly equal in cost to 55nm part. Possibly new part has reduced power consumption and can use a cut down board, that might be enough to get it moving forward.

Of course, if yields are still quite low and improving very rapidly, then there's good reason to delay 40nm production as much as you can - which clearly complicates matters greatly... I'm not sure they'll be improving so rapidly (on the matter of just one month or two) that it'd matter as much as you could naively assume, though.

Think first one(GT218) is still supposed to be May although its been tremendously quiet, hinting perhaps at a delay. After is the G96 replacement in June. Might be July or August before see the new G92 which is 4months guess this is enough to give them a bit more yield.

As for AMD ramping wafers, I'd *guess* it's mostly RV740/RV790/RV730 at this point, with most of it being the first. I would suspect that they cut production more than consumer sales were really reduced (i.e. at or below the level of inventory replenishments in the supply chain which were clearly insufficient), so now they need to increase production back to meet the real channel sell-through (i.e. consumer spending).

They have a new chipset RS880 on 55nm not sure how much that will need. Suspect the plan is for the 40nm G96 and RV730 to duke it out and the RV740 vs new G92. According to link i posted in the RV790 thread, AMD was ordering 2k wafers in Jan, round about 4-5k per month now and they(i guess tsmc) hoping for 10k in June...that seems a touch early for dx11 parts doesnt it?

 

[Jawed]

They have a new chipset RS880 on 55nm not sure how much that will need. Suspect the plan is for the 40nm G96 and RV730 to duke it out and the RV740 vs new G92. According to link i posted in the RV790 thread, AMD was ordering 2k wafers in Jan, round about 4-5k per month now and they(i guess tsmc) hoping for 10k in June...that seems a touch early for dx11 parts doesnt it?

Why would AMD continue production of RV730? It's bigger, 146mm2, than RV740 at 137mm2.

Jawed

 

[Arun]

Why would AMD continue production of RV730? It's bigger, 146mm2, than RV740 at 137mm2.

40nm will be substantially more expensive than 55nm for a few quarters at the very least, and both the memory and the PCB are likely more expensive. So I'd argue the question is: why not?

 

[Tchock]

Arun, I'm pretty sure there will be another RV740 SKU using GDDR3 with <75W board power, and less PCB layers for reference (even less for Palit!) in the end.



They better switch now I reckon, RV730 might be more profitable if we compare the G/DDR3 parts but RV740 offers more SKU flexibility in the channel (stopping 730 now and provide 740 alternatives if needed is a good way to prevent a G92-esque overstock). 730 IMHO is more of an OEM's part, funny that it's not used much on there yet.

 

[Jawed]

40nm will be substantially more expensive than 55nm for a few quarters at the very least,

Why? TSMC has spare capacity, so it's a buyer's market.

and both the memory and the PCB are likely more expensive.

A GDDR3-equipped RV740 board shouldn't be any more complex.

Jawed

 

[Jawed]

RV620/635 GPUs seemed pretty unpopular from day one (let alone their predecessors) and they're still on sale

There's 105 2400XTs in stock here:

http://www.ebuyer.com/cat/Graphics-Cards-ATI/subcat/ATI-Radeon-2000-series

priced to shift

So, it appears AMD has an inventory problem with GPUs two generations older than RV740. I admit AMD might have sold them all by now and it's only AIBs/OEMs and retailers holding this inventory.

Delays in the introduction of RV740 (anything from 3 to 6 months depending...) arguably provide a reason for RV730 production to have continued, but November, when orders went pear-shaped is now 5 months back.

Jawed

 

[v_rr]

Think first one(GT218) is still supposed to be May although its been tremendously quiet, hinting perhaps at a delay. After is the G96 replacement in June. Might be July or August before see the new G92 which is 4months guess this is enough to give them a bit more yield.

VR-Zone managed to score a GT218 card drawing which should give us an idea how the actual card might look like. Nvidia is working on the entry level GT218 cards currently and we will probably see some samples during CeBIT in March and a launch in Apri

http://vr-zone.com/articles/nvidia-gt218-card--specs-surfaced/6529.html?doc=6529

Cebit in March showed nothing and April until now nothing showed. Looks like they barely manage to build ultra-low-end GT200@40nm.

 

[trinibwoy]

Why do you keep posting Vr-zone's guesswork all over the place as if Nvidia promised samples during CeBIT?

 

[v_rr]

Why do you keep posting Vr-zone's guesswork all over the place as if Nvidia promised samples during CeBIT?

So you have another source? Please make us know of it

 

[Arun]

A GDDR3-equipped RV740 board shouldn't be any more complex.

Yeah, that's probably true, as well as Tchock's point. However:

Why? TSMC has spare capacity, so it's a buyer's market.

The point is they don't have spare capacity for the immersion litography tools. Those are the largest cost adders for 40nm, and I'd assume they'd normally be buying them gradually. So right now, any capacity switching from 55 to 40 (more than already expected in the very short term) means extra capital expenditure - and probably at higher tool costs than 6 months down the road or whatever.

Of course, if NVIDIA was somehow late to 40nm and TSMC did order the tools for the expected capacity for both AMD and NVIDIA, that'd be quite different - but I frankly have no idea about that, nor would I suspect it'd be such a long-term phenomenon that it'd affect wafer prices which are normally negotiated well in advance AFAIK. Of course, the magnitudes are hard to be sure about...

 

[Jawed]

The point is they don't have spare capacity for the immersion litography tools. Those are the largest cost adders for 40nm, and I'd assume they'd normally be buying them gradually. So right now, any capacity switching from 55 to 40 (more than already expected in the very short term) means extra capital expenditure - and probably at higher tool costs than 6 months down the road or whatever.

AMD hasn't come out of left-field and said, "40nm GPUs or else!" - this has been ramping up for ages and ages now. We have no idea why 40nm is "late" - e.g. TSMC might have been doing nothing more than curbing its own expenditure.

Also, NVidia is quite clearly delaying (inventory, what other reasons?), so that exposes 40nm capacity for other clients.

Of course, if NVIDIA was somehow late to 40nm and TSMC did order the tools for the expected capacity for both AMD and NVIDIA, that'd be quite different - but I frankly have no idea about that, nor would I suspect it'd be such a long-term phenomenon that it'd affect wafer prices which are normally negotiated well in advance AFAIK. Of course, the magnitudes are hard to be sure about...

And the small matter of a recession... Remember, NVidia basically cancelled its wafers after November. NVidia was presumably buying about twice as much capacity as AMD in the months before that.

Anyway, this apparently comes down to: AMD wants to replace RV730 with RV740 versus TSMC wanting to get wafers rolling, preferably not 40nm. TSMC also has the prospect of AMD going shopping for GPU production at GF within 18 months (will AMD use TSMC for 32nm GPUs?). What's TSMC going to do, throw a spanner in 40nm and let AMD spin on re-spins trying to get it working, thus saving their own money?

Jawed

 

[Tchock]

Well nVidia would go GF too if the process is plain better (and unless TSMC gets to 28nm fast I'm not optimistic on them), so maybe TSMC's wrench is on them instead

Has ATI ever used a foundry other than TSMC for their in-house chips? nV has TSMC, UMC (G92), IBM East Fishkill (NV40) so far.

 

[cal_guy]

Well nVidia would go GF too if the process is plain better (and unless TSMC gets to 28nm fast I'm not optimistic on them), so maybe TSMC's wrench is on them instead

Has ATI ever used a foundry other than TSMC for their in-house chips? nV has TSMC, UMC (G92), IBM East Fishkill (NV40) so far.

Yes they produced a few at UMC.

 

[trinibwoy]

So you have another source? Please make us know of it

Another source for what?

 

[DegustatoR]

The GT218 is supposed to debut next month, the GT216 shortly after, i think these are mainly intended as oem parts. The GT215(a 192 bit G92) is delayed till they can clear the G92/G94 stock, it is approx the same cost to produce as current 55m parts so nvidia are not going to hurry with it.

How about GT214? 8)

 

[Jawed]

Another thing that affects the cost of RV740, positively, is the proportion of the die consisting of ALUs.

Also, since it has 16-wide SIMDs instead of 8-wide in RV730, this reduces the relative cost of redundancy - i.e. 16/17ths is a lower proportion of redundant die within the SIMDs than 8/9ths.

So these two factors combined should make each chip more resilient against wafer-level imperfections.

Jawed

 

[Jawed]

Innovation v Rebranding

http://vr-zone.com/forums/418725/hd-4770-rv740-performance-and-price.html

Blimey, that last slide says it all.

But we're expecting NVidia to respond within 1-3 months with a new GPU, aren't we?

Worth noting, that's 16 full RBEs there. Also the memory on HD4770 is crippled to only 51.2GB/s.

Jawed

 

[3dilettante]

Interestingly, some of those performance improvements over the competition would have still been notable on a graph that didn't start at .8.

 

[Fusion]

Jawed, you're very knowledgeable............. Much respect

Could you please PM me with the references you use since i am in the fourth year
of electronics engineering and our graduation project will be a fully working Graphics card!

So i need as much references i can to educate myself and my team regarding this stuff..

Sorry for this Off-Topic post....