ATI RV740 review/preview

[AnarchX]

Oh, and http://www.xtremesystems.org/forums/showthread.php?p=3798488&posted=1#post3798488

Oooo-Wait. TWO power connectors? WHAT

According to VR-Zone a RV770CE based card. That would match to the HD4870-like PCB, which is shown by Powercolor.

But Fuad reports this card will go on sale for only $79:
http://www.fudzilla.com/content/view/13790/34/

A bit strange for a 8-layer PCB RV770 SKU with 900MHz GDDR5, which should have a performance equal to the HD4770, since it has a higher BW. (Would have a 128-Bit RV770 still 16 ROPs?)

After this Fuad reports, that RV740 has a bad yield @ 750MHz and a better on 700MHz, so that AMD will launch a RV740 based HD4730 with 700MHz GPU-clock:
http://www.fudzilla.com/content/view/13792/34/

With 900MHz GDDR5 this card would be still ~4770 perfomance, while have an improper name.

Does anybody see through this?

 

[Jawed]

So has "HD4750" ever appeared in AMD material? Was it always going to be HD4730?

Worse, though, it appears some HD4730s have a butchered RV770 and others have RV740. Can't help thinking this is either varying by region or crossed wires in the rumours.

Jawed

 

[CarstenS]

To my knowledge, it hasn't.

 

[Jawed]

In the imaginary "which will be first, GT300, Larrabee or RV870?" thread, right now I'm thinking about voting Larrabee.

Jawed

 

[Tchock]

Fuad removed the RV740 4730 news.

 

[Jawed]

Wow, retractions are pretty rare on the tabloid sites. Can't remember having seen one before.

Jawed

 

[CarstenS]

In the imaginary "which will be first, GT300, Larrabee or RV870?" thread, right now I'm thinking about voting Larrabee.

Jawed

Larrabee definitely has the advantage of being shown around on wafers, thus not only having passed Tape-Out but at least silicon back from fab. But from what I hear, G300 having had it's tape out doesn't necessarily mean that it happened only a few days ago.

 

[Jawed]

Larrabee definitely has the advantage of being shown around on wafers, thus not only having passed Tape-Out but at least silicon back from fab. But from what I hear, G300 having had it's tape out doesn't necessarily mean that it happened only a few days ago.

Agreed. The June tape-out rumour for GT300 could have been NVidia smoke-n-mirrors. Though if TSMC is really as leaky as rumoured, I don't see how that would stand up. On the other hand NVidia could simply have beaten the June tape-out by 3 weeks e.g. if cutting GT212 allowed GT300 to move forward.

Anyway with all this noise about 40nm being a disaster zone and both AMD and NVidia being in the same sort of boat (NVidia's being trickier with it's bigger chip and being new to GDDR5?) it seems Intel's in-house 45nm, which will have 2 years production under its belt is sitting pretty, even if it isn't as dense as TSMC's 40nm.

Actually, wouldn't it be interesting to know how big GT200, say, would be on Intel's 45nm. Would it be smaller?

And, I dunno, Intel could still spring a 32nm Larrabee on us, couldn't it?

Jawed

 

[Humus]

In the imaginary "which will be first, GT300, Larrabee or RV870?" thread, right now I'm thinking about voting Larrabee.

Jawed

I'd expect Larrabee to be last out of the pack considering it's "early 2010" according to Intel.

http://www.pcmag.com/article2/0,2817,2347159,00.asp

 

[Jawed]

I'm envisioning a repeat of R600, afflicting TSMC GPUs on 40nm

I sure hope not.

Jawed

 

[Tchock]

Hotboy updated that thread with a bit of info.

1. nV seems to get some special pricing from TSMC, so die size is not the only factor when counting into the cost (rather captain obvious, considering nV's volume is definitely bigger)

2. nV might not be able to get high-speed GDDR5 that fast in the near future, indicating that either the RAM tech isn't maturing that good, or that ATI booked the majority of supply (again). Probably indicates that GT300 will not be 256-bit too.

 

[LordEC911]

Hotboy updated that thread with a bit of info.

1. nV seems to get some special pricing from TSMC, so die size is not the only factor when counting into the cost (rather captain obvious, considering nV's volume is definitely bigger)

Interesting, so I am assuming this is based on overall volume?
So will it factor in to other chips or just this one?
Hmmm...

 

[Jawed]

Certainly makes sense NVidia would get cheaper prices, with such a disparity in volume. But if we're talking about prices per wafer and NVidia actually pays per good chip?

Could NVidia be sucking up so much 40nm volume AMD can't get enough?

Would NVidia need anything more than about 1.1GHz GDDR5 any time within the next 6 months?

Jawed

 

[Arun]

I'm envisioning a repeat of R600, afflicting TSMC GPUs on 40nm

I sure hope not.

Remember: the problem with R600 wasn't 80nm, it was 80HP. Similarly, AFAIK the *leakage* problem with 40nm isn't 40nm; it's 40HP, or rather, any 40nm process (including some 40G) with SiGe straining. I can say pretty confidently that a reasonably conservative 40G process without SiGe straining actually has lower leakage than 65G. However, that might not be much of a consolation when most of the 40nm performance gains were supposed to come from that and so presumably RV740, RV8xx, GT21x, and GT30x are all implemented with it. And of course, there's also the yield problem which I don't know anything about except that 65nm certainly had its fair share of those in the same timeframe too, see: http://twitter.com/ArunDemeure/status/1706653190

BTW, NVIDIA doesn't pay per-good chip (at least for sufficiently mature nodes). Hasn't in a long, long time. I suspect AMD mostly pays per-wafer too right now, although certainly they have done much more per good chip than NVIDIA in the last 5 years.

 

[rjc]

Joy Division come back to life and touring?

Better than babelfish, which gave "heaven is a place on earth" for the same text, that would be much worse....

There's nothing in this hotboy thread. There's ANOTHER thread where he comments on how horrendous TSMC Shanghai is, perhaps due to TSMC's more pennywise oriented purchasing of equipment.

I saw the tsmc shanghai stuff, which made me pay attention to the other things he was saying...so your saying he is mixing things he knows in with random speculation? That makes it much harder.

On the update google gives as:

Future period of time NV possibly will not be able to obtain high frequency GDDR5, the modern age will be at least this
The NV generation of labor cost has been lower than A, this is also a reason which later will fade out slowly. Therefore everybody later do not use the core area to judge the cost
The core income is not much, QE also more than 50,000.

• So that is possibly no GDDR5 supply.
• Lower labor cost than AMD in the past advantage now "fading out slowly". Therefore if they have a different wafer cost cannot directly compare die areas to assess relative costs.
• "Core income not much"...whose core income?? nvidia(ie price they are selling each chip for) or tsmc(for the wafer)
• "QE more than 50000"....is that a cost ie wafer more than 50000yuan ~ $US7300 or quantity more than 50000 wafers ordered (that is way too much should be < 20k at most per month).

From reading thread, does anyone know what "glue water" means? seems to indicate some kind of progress not exactly sure. Also "arab league" and "river crab" if have the time.

Finally the 4730...they are dumping excess 4870 boards + RV770 chips severely disabled. Worked out roughly awhile ago average cost for RV770 ~$25...if yield for RV740 is below 50% cost will approach $30 per chip.

 

[LordEC911]

Would NVidia need anything more than about 1.1GHz GDDR5 any time within the next 6 months?

Jawed

I wouldn't think so unless they went with a 256bit bus.
4.4ghz GDDR5 on a 384bit bus, smallest I could see them going, would still be ~211Gbps.

 

[Tchock]

river crab - 河蟹， samely pronounced (without the accents) as 和谐， which means harmony (??)

I've seen them interused on CHH, but I'm not a mainland Chinese so I don't get their intarweb slang

 

[rjc]

river crab - 河蟹， samely pronounced (without the accents) as 和谐， which means harmony (??)

I've seen them interused on CHH, but I'm not a mainland Chinese so I don't get their intarweb slang

Ah ok. This sentence: "RV740 雷的不轻阿" gets translated as "RV740 does thunder heavy arab league" was thinking "arab league" was some kindof slang for the mainstream market or low priced cards but not really sure.

Understand they talk about trying the water for new chips, with "glue water" glue sounds like a verb, perhaps a more or less aggressive plan.

Sorry for distracting the thread....

Re wafer discussion, wafers can also be rejected if yield is too low, but this is usually something much below what would normally expect ie say for RV740 if they are averaging 50% yield then rejection is set at like 15-20% say.

 

[Jawed]

Remember: the problem with R600 wasn't 80nm, it was 80HP.

I think it was libraries...

Similarly, AFAIK the *leakage* problem with 40nm isn't 40nm; it's 40HP, or rather, any 40nm process (including some 40G) with SiGe straining. I can say pretty confidently that a reasonably conservative 40G process without SiGe straining actually has lower leakage than 65G. However, that might not be much of a consolation when most of the 40nm performance gains were supposed to come from that and so presumably RV740, RV8xx, GT21x, and GT30x are all implemented with it. And of course, there's also the yield problem which I don't know anything about except that 65nm certainly had its fair share of those in the same timeframe too, see: http://twitter.com/ArunDemeure/status/1706653190

65nm, yep we saw the pain and quite a bit more for NVidia than AMD.

BTW, NVIDIA doesn't pay per-good chip (at least for sufficiently mature nodes). Hasn't in a long, long time. I suspect AMD mostly pays per-wafer too right now, although certainly they have done much more per good chip than NVIDIA in the last 5 years.

How long before a node is "sufficiently mature"? How long was it before AMD stopped paying per good chip on 55nm? Why would AMD be paying per good chip on 40nm right now?

What does AMD's and NVidia's involvement in getting a node up to speed buy them?

Jawed

 

[Tchock]

RV740 雷的不轻阿



地雷 - landmine/trap/ aka RV740 does have quite some problems at TSMC.



Is this the first time TSMC is implementing SiGe Straining btw?