ATI RV740 review/preview

[pjbliverpool]

On the other hand, the 8800GTS (G92) will definitely be faster in cases not bandwidth limited as it has clocks much closer to the 240 (650 vs 675, 1625 vs 1688, 1940 vs 2200) but with a 14% cluster advantage.

RV740 @ 650 should trail the 8800GTS by a little, perhaps on par with it (and the 9800GTX original) on OCed editions, but ithere might be a 5-10% gap wrt the 240.

My bad, I thought the GTS 512MB had 112 SP's. But I was getting it mixed up with the improved GTS 640MB. Damn, NV's naming schemes were confusing back then!!

The 240 is basically a standard 9800GTX with 1 cluster disabled. Given that the 4850 was on par or margianlly faster than a 9800GTX (at least with 4xMSAA) and that RV740 is a good 5-10% slower than the 4850, I still think it will be pretty even with the 240.

If the performance edge has to go to one though, its pretty likely to be RV740.

 

[Jawed]

I read the link. The PDC-presentation tells us, that D3D10.1 API is the foundation - now i wonder, what techlevel might be used for desktop-effects.

Hot-tracking, which if you've seen video of it in action is a hugely important part of the UI is the only 10.1-specific feature as far as I can tell.

According to this:

http://download.microsoft.com/downl...02-97DB3CF79AD5/GRA-T786_Kancherla_Taiwan.pdf

DWM 1.1 is supported by all 10 and above hardware, which means that NVidia D3D10 GPUs will in fact benefit from the memory savings, so I was wrong about that earlier. The memory savings seem pretty substantial, too. Who ever has less than 30 windows open, eh?

So, NVidia's marketing department shouldn't have to work too hard.

Jawed

 

[CarstenS]

That's why I compared 4Q 08 to 4Q 07.

That's not the only thing I meant. Nvidia would have known that they're in for a difficult time ever since RV770s pricepoint and performance were known. Then the economical crisis started, which they must have seen, would hit their Q4/08 squarely in their face. Now, if I'd have to report bad news to my "owners" (basically), then this is a thing I'd rather do only once and preferrably when everyone else is being grounded too. Then I would not mind making this single report a bit worse than it could have been so to make some leeway wrt to forthcoming quarters, where my inventory is lowered to a bare minimum, my expenses have lowered since all order cancellations are in effect and so forth. With this quarters results being as they are, I would further make some price reductions, basically selling some k-units of maybe 65nm-inventory for barely no margins just to increase market share - and i would strive to stuff all of this into one single quarter, however disastrous that might end.

Then maybe light at the end of the tunnel will shine in Q1/09 alright.


--
Jawed,

Why would D3D10.1 (hw) be necessary for hot-tracking?

 

[entity279]

RV740 is a nice chip but what it essentially doing is filling the hole in AMD's line between RV770 and RV730. NV never had that hole so it's quite normal for them to slightly adjust their current offerings and be done with it. No technological advancements are needed from them here.

Actually, there's already 4830 for that so no gap there. RV 740 is a cheaper replacement (maybe yield-dependent though.).

And if nV never has a hole it will make sense that they won't adjust prices too much while rearranging/rebranding their line-up, right?

LE: Wooow soory for that I was waay too late (never talk on the phone & post, I know..)

 

[Jawed]

I don't know.

Jawed

 

[neliz]

I don't know.

Jawed

I think CarstenS asked the question before you added your last sentence "all 10 and above cards" to your post

 

[Silent_Buddha]

That's not the only thing I meant. Nvidia would have known that they're in for a difficult time ever since RV770s pricepoint and performance were known. Then the economical crisis started, which they must have seen, would hit their Q4/08 squarely in their face. Now, if I'd have to report bad news to my "owners" (basically), then this is a thing I'd rather do only once and preferrably when everyone else is being grounded too. Then I would not mind making this single report a bit worse than it could have been so to make some leeway wrt to forthcoming quarters, where my inventory is lowered to a bare minimum, my expenses have lowered since all order cancellations are in effect and so forth. With this quarters results being as they are, I would further make some price reductions, basically selling some k-units of maybe 65nm-inventory for barely no margins just to increase market share - and i would strive to stuff all of this into one single quarter, however disastrous that might end.

Then maybe light at the end of the tunnel will shine in Q1/09 alright.

Well that problem with that is noone expected Q4 08 to completely drop off the map as much as it did. Most were expecting a worse Q4 than 07 but not MUCH worse. And everyone was expecting a bit of a rebound from Q3 due to the holiday season. Nvidia's Q3 ended end of Oct. The really bad stuff didn't hit until Nov.

If anything I would have expected many companies to preload Q3 with slightly worse financials in order to make Q4 look even better with the holiday buying spree. Either way however, none of that really affects your yearly report, so I doubt many companies go to the trouble to bake their quarterly reports unless they are trying to seek out more investors for an upcoming project.

Although it should be noted that AMD's Q4 ended a month earlier than Nvidia, so there wasn't as much of the really bad weeks in there. But even taking that consideration. Nvidia's graphics division was still hit considerably harder than AMDs graphics division, despite stronger sales (for Q4). It should also be noted that the strength of those Q4 sales mainly came from G8x and G9x which had to have margins slashed even more to compete with significantly smaller Rv7xx dies.

Regards,
SB

 

[homerdog]

ATI took steps to fix this when going from R600 -> Rv670, a relatively quick turn around. We're still waiting for Nvidia to do something. 65nm -> 55nm was a start but it's still not good enough.

G80 -> G92 ?

Regardless I don't see G92 as the chip NVIDIA wants to put up against RV740, I see it as the chip NVIDIA has to put against RV740. Like GT200 vs RV770 - a losing proposition for the Green Team. We can speculate about how much die size really matters in production costs all day, but why would AMD do this if it wasn't significantly cheaper to produce (than HD4830, which shouldn't be much more expensive to produce than GTS240)?

 

[CarstenS]

SB, that's exactly what I'm saying: Nv's 4th Quarter started November, when they saw it coming, they still had 2,5 months to make preparations for an accelerated one-quarter nose-dive giving them the chance to trade profit for market share - all this conveniently buffered by the catastrophic global economy.

 

[DegustatoR]

90% of the performance? I remember the clocks going down on the "Green Edition" so one has to wonder why they put all the extra megahertz on the 9800GTX then. Clocks should be 550/900 instead of 600/900 and it's actually more expensive than the GTS 240. Save the planet, give us more money!

To reiterate: GTS250 > 9800GTX+ > 4850 = 9800GTX > GTS240 =(?) 4830 > 9800GT = 4750 > 9800GTG (no, I don't count MSAA 8x cases as typical).
9800GTG may well have 90% of RV740 performance, why not?

To both AMD and Nvidia, die size means quite a bit more than nothing.

Where did I say that die size means nothing?

And I think we can compare overall losses by AMD's graphics division (ATI) and Nvidia year to year (4th quarter 08 to 4th quarter 07) to have anecdotal evidence that AMDs die size advantage in every market segment where they compete with Nvidia certainly points to all current Nvidia chips being significantly more expensive (and lower margin) than their competing AMD counterpart.

If you do that and compare financial results for the several quarters back you'll find the difference quite small -- if at all existant. I did just that recently. And 4th quarter 2008 isn't a good quarter to base any assumptions upon. (Well, actually this way of comparing margins isn't good at all but since we don't have any other...)

ATI took steps to fix this when going from R600 -> Rv670, a relatively quick turn around. We're still waiting for Nvidia to do something. 65nm -> 55nm was a start but it's still not good enough.

So we should just forget the G80->G92 transition which was exactly the same as R600->RV670?

I do believe that both AMD and NV can count their money better than anyone here. And i do beleive that both are doing what's best for them from the chip cost point of view. If NV's doing GTS240/GTS250 and Green Edition combos against 47x0 line then i do believe that they plan to earn money on these cards (otherwise -- why do them at all?). More or less than AMD on 47x0? Time will tell. But in 3rd Q 08 their income was more or less the same. And GT200 is certainly less appropriate GPU to compete against RV770 than G94b and even G92b against RV740.

It's just that most people here seems to think that RV740 vs G94/92 is going to be a repeat of RV770 vs GT200 scenario while it won't. New G94/92 cards will simply be a temporary solutions based on (very) old GPUs -- it certainly wasn't the case with GT200 cards half a year back. And RV740 is certainly not the RV770 which was AMD's forth GPU on 55nm.

Right now there's nothing for NV to be afraid of in RV740. And when there may appear something they'll probably have their own 40G line for that segment. So the real question is: will RV740 be able to compete against GT21x GPUs?

 

[Silent_Buddha]

SB, that's exactly what I'm saying: Nv's 4th Quarter started November, when they saw it coming, they still had 2,5 months to make preparations for an accelerated one-quarter nose-dive giving them the chance to trade profit for market share - all this conveniently buffered by the catastrophic global economy.

I don't think they would have had a choice. Their Q4 could have been significantly worse if they hadn't traded margins for marketshare. None of their products could compete with the ATI lineup on price and they would have ended up losing both revenue and marketshare. Rather than just revenue only as it turned out.

There really wasn't anything they could do to prevent the revenue loss so they instead (for the first time) took a huge hit on margins in order to maintain marketshare.

All that due to AMD having such a competitive part with a much smaller die size.

Smaller part and non-competitive (IE - Rv670) would have been business as usual.

Which goes back to the point I was making. Die Size is very important when talking about things...

Regards,
SB

 

[Silent_Buddha]

G80 -> G92 ?

Regardless I don't see G92 as the chip NVIDIA wants to put up against RV740, I see it as the chip NVIDIA has to put against RV740. Like GT200 vs RV770 - a losing proposition for the Green Team. We can speculate about how much die size really matters in production costs all day, but why would AMD do this if it wasn't significantly cheaper to produce (than HD4830, which shouldn't be much more expensive to produce than GTS240)?

No arguments about Nvidia HAVING to use G92 since they have no other alternative, this side discussion was originally a response to someone claiming Die Size means absolutely nothing.

And my comment about waiting to see an Nvidia response similar to R600 -> Rv670 is that Nvidia is now in a similar (not same) situation.

ATI was forced to sell R600 at far lower than was profitable with very low margins. Had they not made such a quick transition things could have gotten really ugly really fast.

Nvidia's situation isn't quite as dire, however with no end to the recession yet in sight, it could become dire. And Nvidia is now in the position ATI was in with R600 of having to sell their chips with much lower margins in order to get them to sell versus the competition.

Of course, to the consumer none of this matters unless one business goes out of business or can no longer afford the R&D to keep up. See companies such as Rendition, 3dfx, 3dlabs, S3, etc. for examples. Although Rendition was never really a market leader.

Regards,
SB

 

[Arty]

To reiterate: GTS250 > 9800GTX+ > 4850 = 9800GTX > GTS240 =(?) 4830 > 9800GT = 4750 > 9800GTG (no, I don't count MSAA 8x cases as typical).
9800GTG may well have 90% of RV740 performance, why not?

I'm not if you ran through the numbers from Guru3D, 4750 is about 11% faster than the 4830. That would change your equation to this:

GTS250 > 9800GTX+ > 4850 = 9800GTX > 4750 = (?) GTS240 > 4830 > 9800GT > 9800GTG

Roughly that will put the 9800GTG as 70-75% of RV740 performance. Which is not bad, but I guess it will be a niche product.

Right now there's nothing for NV to be afraid of in RV740. And when there may appear something they'll probably have their own 40G line for that segment. So the real question is: will RV740 be able to compete against GT21x GPUs?

That will be an interesting fight. But I guess ATI could pull a G92 and fight those 40nm (if they are really good & if ATI's 40nm line sucks) with a full RV770. I just see it as ATI having a much more comfortable choice of options than Nvidia.

 

[Tchock]

From my crude guesstimation (mainly just the crappy power regulation on what's supposed to be an ES utilizing components of the utmost quality) RV740 seems to exceed the 75W cap not by far.

Even if TSMC's 40G leaks more extrodinarily than usual, would a 110-120W part make sense in the big picture?

Granted 740 "XT" would have to reach 800-850//1250 Mhz to slot into the 4850 and 4870 in-between, but it would serve as a viable upgrade path to the 4850. And 149USD for an 1GB XT (GTS 250 1GB) part is plain lucrative, mainly on ATI's part.

Did they respin for that, actually?

 

[DegustatoR]

All that due to AMD having such a competitive part with a much smaller die size.

Wrong. All (most of) that is due to sudden stop in global economy. Nobody was buying GPUs from NV in 4th Q, everybody was selling their stock which was bought previously. That's the main reason for bad NV's results during 4th quarter.

I'm not if you ran through the numbers from Guru3D, 4750 is about 11% faster than the 4830. That would change your equation to this:

GTS250 > 9800GTX+ > 4850 = 9800GTX > 4750 = (?) GTS240 > 4830 > 9800GT > 9800GTG

Roughly that will put the 9800GTG as 70-75% of RV740 performance. Which is not bad, but I guess it will be a niche product.

Ah, OK, my bad. But i may be wrong about 9800GT being slower than 4830. My fast check on our own current data gave me this:

So it looks like without AA 9800GT is competitive with 4850 (!) right now. Yeah, i'm a little surprised myself.

As for 9800GTG being a niche product -- i think that's pretty much a given since they even specify that niche in the name of the card -)

That will be an interesting fight. But I guess ATI could pull a G92 and fight those 40nm (if they are really good & if ATI's 40nm line sucks) with a full RV770. I just see it as ATI having a much more comfortable choice of options than Nvidia.

Why is that? Options are exactly the same: use the old 55nm GPU or the new 40nm GPU.

Edit: Another source: http://www.computerbase.de/artikel/...ikkarten_2008/19/#abschnitt_performancerating
9800GT is generally faster than 4830 -- besides AA 8x cases -- although certainly not competitive with 4850. That's closer to what i thought.
Another interesting side-effect of that test is that GTS250 may end up being comparable to 4870...

 

[v_rr]

Another interesting side-effect of that test is that GTS250 may end up being comparable to 4870...

LOL
Dude, GTS250===============9800GTX+

From March 3rd onwards, GeForce 9800 GTX+ will officially be rebranded as GeForce GTS 250. The specs remains the same as the 9800GTX+ but some AIC makers are planning to market those cards as a whole new product with a new cooler.

http://vr-zone.com/articles/cheaper-geforce-gts-250-cards-available-on-march-3rd/6576.html?doc=6576

You put a new Name and a "bling bling" Box and suddenly it is comparable to HD 4870?

I will make you a picture in other way. The RV740 with a slight bump in clocks (think OC, XXX, PCS versions) will catch the GTS250=9800GTX+. This with the price <100$. Nvidia need to drop even more its margins (that are already very very bad) from 130$@GTS250 to <100$ to catch price/performance of this litle 40nm speeder.
The Guru3D review card had 650Mhz clock speed. Think in 750Mhz and 800Mhz versions when they reach public. It will be more then enough to hold GTS250.
G92 is nowhere competitive in price or performance. Remember Nvidia finished 2008 with a loss compared to 2007 with huge money (G80).
Still they lost huge discrete market share. In global share they also lost some points but in that AMD will never catch Nvidia because they have Intel IGP market that is huge compared to AMD market. And IGP acount for 80/90% of the global market? Or more...

In discrete share AMD and Nvidia should be almost 50-50%. And wait until HD 4000 mobile reach the market (1 week missing). That is going to hurt Nvidia as bad as it hurt on desktop side. Nvidia the only thing that is doing is renaming G92 mobile GPU´s while ATI is coming with the whole new lineup. If you research a litle you will see that Nvidia is in a huge trouble in Mobile GPU´s when HD 4000 mobile show up:
http://www.hexus.net/content/item.php?item=17380

Wait for Cebit next week....


PS: That review is a litle old (Cat 8.11).
http://www.computerbase.de/artikel/...i_catalyst_92/10/#abschnitt_performancerating
From Cat 8.11 -> Cat. 9.2 you get 9% boost 1680@4xAA16AF.

 

[mczak]

NV will offer with the GTS 240 (675/1688/1100MHz 112SPs G92b) an equal performance to the RV740XT at the same price point of ~$99 (considering GTS 250 512MB is allready priced at $129).

hmm, the guru3d ES doesn't look like the XT version to me.

Of course full D3D10.1 would be a nice to have, but on the other side the competition is not able to deliver a free select-able high-quality AF on 32-80 TMU cards and it looks that the cards before HD4000 are not able to support OpenCL.

Small nitpick, what 80TMU card would that be .

While it looks that AMD is not able to make it "PEG-only", Nvidia is preping an interesting solution:
http://vr-zone.com/articles/geforce-9800-gt-green--no-external-power-needed/6643.html?doc=6643
- PEG-only (<75W TDP)
- shorter PCB than the RV740XT ES
- ~ 90% performance of RV740XT ES

I already mentioned this in the other thread, but to me this ES board doesn't look like a final version. I suspect it's going to be smaller, and personally I wouldn't be suprised if the non-xt version wouldn't require a PEG connector.
But even if it does, it is competitive with the GTS 240 (with potentially higher margins). The 9800 gt green edition though will not be competitive - more like 70% performance (won't that actually get it close to 9600 gt non-green)

 

[compres]

LOL
Dude, GTS250===============9800GTX+


http://vr-zone.com/articles/cheaper-geforce-gts-250-cards-available-on-march-3rd/6576.html?doc=6576

You put a new Name and a "bling bling" Box and suddenly it is comparable to HD 4870?

I will make you a picture in other way. The RV740 with a slight bump in clocks (think OC, XXX, PCS versions) will catch the GTS250=9800GTX+. This with the price <100$. Nvidia need to drop even more its margins (that are already very very bad) from 130$@GTS250 to <100$ to catch price/performance of this litle 40nm speeder.
The Guru3D review card had 650Mhz clock speed. Think in 750Mhz and 800Mhz versions when they reach public. It will be more then enough to hold GTS250.
G92 is nowhere competitive in price or performance. Remember Nvidia finished 2008 with a loss compared to 2007 with huge money (G80).
Still they lost huge discrete market share. In global share they also lost some points but in that AMD will never catch Nvidia because they have Intel IGP market that is huge compared to AMD market. And IGP acount for 80/90% of the global market? Or more...

In discrete share AMD and Nvidia should be almost 50-50%. And wait until HD 4000 mobile reach the market (1 week missing). That is going to hurt Nvidia as bad as it hurt on desktop side. Nvidia the only thing that is doing is renaming G92 mobile GPU´s while ATI is coming with the whole new lineup. If you research a litle you will see that Nvidia is in a huge trouble in Mobile GPU´s when HD 4000 mobile show up:
http://www.hexus.net/content/item.php?item=17380

Wait for Cebit next week....


PS: That review is a litle old (Cat 8.11).
http://www.computerbase.de/artikel/...i_catalyst_92/10/#abschnitt_performancerating
From Cat 8.11 -> Cat. 9.2 you get 9% boost 1680@4xAA16AF.

You pretty much sum up what i had in mind, that is why I am very confused now with DegustatoR claims.

 

[Tchock]

Slightly off topic, but I've seen the Mobility Radeon 4570 and 4650 get adopted exclusively in quite some models. (HP/Dell)

I think some builders were burnt badly by the G84/86 issue that they probably won't acquire any (low/midrange) GPUs from nVidia for a while- at least until a commanding lead in perf/watt is held. For highend mobile GPUs nVidia still holds quite some influence. Remains to be seen if M98 reverses that.

AMD's IGP shares will remain bad (and shrinking) until they get that dualcore Yukon CPU in, preferably on 45nm, for cheap all-in-ones and thin notebooks.

/ot

Hopefully the 9800GT Green doesn't face the issue the 2600XT did with mobos being unable to supply near-75W loads consistently over PCIe, where it was really bad (especially on Chiphell then).

 

[Freak'n Big Panda]

To reiterate: GTS250 > 9800GTX+ > 4850 = 9800GTX > GTS240 =(?) 4830 > 9800GT = 4750 > 9800GTG

GTS250=9800GTX+ no? There might be a marginal clock boost but seeing as it's the same chip I don't see why anybody would be expecting anymore more than 5%. And 4850 = 9800GTX+. Just depends on the app, sometimes (GRID, bioshock, aoc come to mind) the 4850 absolutely creams anything G92 has to offer.