The ATI Radeon 5XXX Thread

[saaya]

http://www.semiaccurate.com/2009/09/...-shader-count/

and theres a rumor going around that the 5870 will cost 100$ less than a 295, which is 399$ then? thats pretty high... either yields are bad or ati wants to cash in as long as nvidia cant compete... i have to say, i really dont like what im hearing about prices now... it really spoils the anticipation of the new 800 series :/

i really hope those are only the launch prices and they will drop notably after a few weeks...

EDIT: and about gt300.. it should easily beat a 5870, i dont think anybody doubts that, the question is more about how much itll cost... gt300 will definately bring the perf crown back to nvidia once it comes out, but that seems to be a while and it wont be cheap...

i cant help but feel things actually got worse this time around... ati wont offer great price perf anymore cause nvidia has even less to compete with their mainstream parts, and the highend will be nvidia dominated as before but prices will probably be even higher than before... sigh...

[Utnorris]

Very interesting, but until it debuts it's all speculation. Even if the HD5870 comes in at $400, that will still be a good deal if it beats the GTX295 in perfomance, which means Nvidia will have no choice but to drop the price of the GTX295 down to around $350 leading the rest of the GTX series to follow suit. That's far from bad news from a consumer's standpoint.

[onethreehill]

Repost
http://www.xtremesystems.org/forums/...&postcount=418

[labs23]

-Very interesting, but until it debuts it's all speculation.

-Even if the HD5870 comes in at $400, that will still be a good deal if it beats the GTX295 in perfomance, which means Nvidia will have no choice but to drop the price of the GTX295 down to around $350 leading the rest of the GTX series to follow suit. That's far from bad news from a consumer's standpoint.

-Exactly, something not to be taken as is until Sept. 10.

-I'm gonna love that. Should still be great for non-early DX11 adopters, ie to those who wait for a DX11 competition.

[saaya]

Repost
http://www.xtremesystems.org/forums/...&postcount=418

i think this deserves its own thread?
its pretty unlikely that charlie got those numbers wrong so close to the launch...

this can only mean 2 things though... either going from 10.1 to 11 takes a lot more transistors than we thought, or those gpus actually contain some extra shader blocks and other blocks... those could be for redundancy since its a rather big chip and 40nm wafers are still somewhat the swiss cheese of semiconductors im curious though if those redundant parts can be enabled in the near future when yields improve and we will se a 5890 with actually more logic and not just higher clocks...

about pricing:
GTX295 500$
GTX285 350$
GTX275 225$
GTX260 175$
GTS250 125$

4870x2 400$
4890 200$
4870 175$
4850 125$
4770 125$

if the 5870 is 399$ and it beats the 295, veeery likely , then what about the 5850?
it definately beats the 285, lets assume for 100$ less as well? then thats 299$... quite a jump from 5850 to 5870... but its possible...
now with the 5850 at 299$ what happens to the 4800 series? they get replaced by juniper aka 5830 at about the same price?

while ati could def position their cards that way, i doubt they will... it wouldnt generate a lot of drive for people to buy new cards...
make it 349, 279, 200 or even 299, 249 and 175 instead and people will suddenly feel a big urge to upgrade

[jaredpace]

I don't think an HD5870 is going to be faster than a GTX295 on average - perhaps only in a few instances. Computerbase.de has some average performance ratings of the cards relative to the GTX295 based on averages from all the listed benchmarks. I expect 5870 to be 95-98% and 5850 to be 80-85% of a GTX295's overall performance (at this resolution). Less than one week away!


Source:
http://www.computerbase.de/artikel/h...ting_qualitaet

A B3D member made this post with a list of his predictions for the future card's performance in each game. He calls it his "5870 fake review" It can be seen here:

Source:
http://forum.beyond3d.com/showpost.p...postcount=2223

[LordEC911]

i think this deserves its own thread?
its pretty unlikely that charlie got those numbers wrong so close to the launch...

this can only mean 2 things though... either going from 10.1 to 11 takes a lot more transistors than we thought, or those gpus actually contain some extra shader blocks and other blocks... those could be for redundancy since its a rather big chip and 40nm wafers are still somewhat the swiss cheese of semiconductors im curious though if those redundant parts can be enabled in the near future when yields improve and we will se a 5890 with actually more logic and not just higher clocks...

Well, it only took him 2-3 weeks to get the codenames right...
Also, what are you talking about with the diesize/transistor count?
Nothing seems too out of line.

Also, redundant parts of the chip are never going to be "enabled" it doesn't work that way, unfortunately.

[flopper]

A B3D member made this post with a list of his predictions for the future card's performance in each game. He calls it his "5870 fake review" It can be seen here:

Source:
http://forum.beyond3d.com/showpost.p...postcount=2223

His predictions is on the low side also.
7days to go

[570091D]

does anyone else remember rumblings from investors after the 48xx launch? we're pricing this card too low? i believe this pricing scheme i aimed at returning investor confidence, and truly, if amd has a superior product they should be paid for it. that's business.

[saaya]

charlie had the codenames wrong? just hemlock afaik and thats probably cause he listened to sylvie

jared, why would the 5870 be that much faster than the 5850? those numbers look weird, i cant figure out what he based his perf boost over 4890 on..
what shader count did he expect when he made that graph? it doesnt make any sense that with about double the hw he only predicts a 10% boost in some games?

and about 5870 beating gtx295...
4870 to 4870x2 scaling is 55% of a 295 to 90% of a 295... considering that a 5870 appears to have the same hw specs as a 4870x2 but isnt castrated by having it on 2 chips and there is no overhead, id expect a 5870 to be around 110% of a 295 like a 4870x2 would be, if it wouldnt have any overhead caused by being based on 2 chips... and thats assuming the 5870 comes clocked at 750mhz core and 3600mhz mem... according to the rumored specs it looks like like 800-900 clocks for the gpu and 5000mhz for the mem... so it will DEFINATELY beat the 295...

[Smartidiot89]

Charlie says that Juniper 180mm2 will be 800 shaders.

Will there be a "CypressLE" or something like HD4830(RV770LE) had 160 shaders disabled to 640 shaders instead of 800?

[Lightman]

I don't think an HD5870 is going to be faster than a GTX295 on average - perhaps only in a few instances. Computerbase.de has some average performance ratings of the cards relative to the GTX295 based on averages from all the listed benchmarks. I expect 5870 to be 95-98% and 5850 to be 80-85% of a GTX295's overall performance (at this resolution). Less than one week away!

...

I think equaling GTX295 shouldn't be a problem for 5870 specced as rumor says. Remember that for the first time in recent ATi history we are getting not only 2xShader count, 2xRBE but also 2xROP. All this on one die instead of 2 dies. On top of that better scheduler plus other tweaks will mean much better scaling compared to older 4870X2 (no CF overhead). The only limiting factor can be memory bandwidth in some cases.

I'm looking forward to these new cards!


EDIT:
sayya you're typing too fast! We made same point in two different ways and two different posts :p

[saaya]

does anyone else remember rumblings from investors after the 48xx launch? we're pricing this card too low? i believe this pricing scheme i aimed at returning investor confidence, and truly, if amd has a superior product they should be paid for it. that's business.

and those investors know more about running amd than amd does?
im not saying amd has been managed perfectly lately, FAR from it, but its ridiculous to think investors and analysts could do better...

you have to balance out demand, price per unit and units sold... investors arent happy that amd is losing money, but how retarded do you have to be to think the solution is to simply increase prices?

seriously its like they have no concept of commerce at all...

sure, lets them try how high they can price their cards and still sell some...
but im pretty sure if they price them as high as rumored lately, they will sell a lot less and overall make less profits than they would if theyd sell them cheaper...

if you increase the price by 10% and lose 10% customers as a result cause the pricing is too high for them, you actually make less money than selling it at the original price... if you reduce the price by 10% in most cases you end up with a more than 10% increase in volume, sometimes up to 30%... thats why the real money is being made selling mainstream and entry level parts aka juniper and below at tight margins, so why would ati price a highend part, which either way wont make them rich, at such high levels?

it has nothing to do with making money, if they price it that high its probably cause it costs them that much to make and theyd lose money if theyd sell it for less...

[blindbox]

Since Chiphell's picture of Cypress is pretty spot-on, I bet their rumours of 1600sp would be correct too. That article just confirms it further, IMO.

[LordEC911]

and about 5870 beating gtx295...
4870 to 4870x2 scaling is 55% of a 295 to 90% of a 295... considering that a 5870 appears to have the same hw specs as a 4870x2 but isnt castrated by having it on 2 chips and there is no overhead, id expect a 5870 to be around 110% of a 295 like a 4870x2 would be, if it wouldnt have any overhead caused by being based on 2 chips... and thats assuming the 5870 comes clocked at 750mhz core and 3600mhz mem... according to the rumored specs it looks like like 800-900 clocks for the gpu and 5000mhz for the mem... so it will DEFINATELY beat the 295...

I think equaling GTX295 shouldn't be a problem for 5870 specced as rumor says. Remember that for the first time in recent ATi history we are getting not only 2xShader count, 2xRBE but also 2xROP. All this on one die instead of 2 dies. On top of that better scheduler plus other tweaks will mean much better scaling compared to older 4870X2 (no CF overhead). The only limiting factor can be memory bandwidth in some cases.

That last bit was exactly what I was going to say to saaya.
We are doubling, if not more than doubling, almost every aspect of the chip but yet they are only, as far as we know, increasing the bandwidth by 39% (160 vs 115GBps).

Will there be a "CypressLE" or something like HD4830(RV770LE) had 160 shaders disabled to 640 shaders instead of 800?

Eventually. Remember, it isn't AMD/ATi's strategy anymore to rely on salvaged parts.
The last few generations they have been releasing those salvaged parts, seemingly, when the GPU is EOLed, when they know they have a decent stock and need to clear inventory w/o worrying about too many more salvaged parts flowing in.

[purecain]

cant wait... this is truly a very exciting release.....

[Smartidiot89]

Eventually. Remember, it isn't AMD/ATi's strategy anymore to rely on salvaged parts.
The last few generations they have been releasing those salvaged parts, seemingly, when the GPU is EOLed, when they know they have a decent stock and need to clear inventory w/o worrying about too many more salvaged parts flowing in.

True but there is a large gap between Juniper and Cypress so there should be something in between... atleast I hope so. Yields are never perfect so using salvaged part is better then to throw away chips which are semi-broken.

[saaya]

That last bit was exactly what I was going to say to saaya.
We are doubling, if not more than doubling, almost every aspect of the chip but yet they are only, as far as we know, increasing the bandwidth by 39% (160 vs 115GBps).

that is assuming the 4870x2 has more bandwidth than the 4870... which it effectively doesnt... both cards mirror their counterparts memory afaik and hence the extra memory bandwidth is completely useless... if at all, its a minor boost...

dual gpu cards would be way less memory bandwidth limited than single gpu cards if the added bandwidth would really be useful... thats not the case, afaik its even the contrary, dual gpu cards are more bandiwdth limited than single gpu cards... thats why there has been all this talk about getting gpus to share their memory capacity and bandwidth. cause right now, there is double the mem and double the mem bw on a dual gpu card, but it doesnt result in higher overall bw and higher overall memory capcity...

in sfr there might be a boost from higher overall bandwidth of a dual gpu card... but even that is unlikely... and even if true, most of the time xfire and sli run afr cause it brings a notably higher boost...

[LordEC911]

that is assuming the 4870x2 has more bandwidth than the 4870... which it effectively doesnt... both cards mirror their counterparts memory afaik and hence the extra memory bandwidth is completely useless... if at all, its a minor boost...

dual gpu cards would be way less memory bandwidth limited than single gpu cards if the added bandwidth would really be useful... thats not the case, afaik its even the contrary, dual gpu cards are more bandiwdth limited than single gpu cards... thats why there has been all this talk about getting gpus to share their memory capacity and bandwidth. cause right now, there is double the mem and double the mem bw on a dual gpu card, but it doesnt result in higher overall bw and higher overall memory capcity...

in sfr there might be a boost from higher overall bandwidth of a dual gpu card... but even that is unlikely... and even if true, most of the time xfire and sli run afr cause it brings a notably higher boost...

As to your first parts, you are correct, dual GPU cards like the 4870x2 don't increase the bandwidth vs a single GPU card but it also doesn't decrease it.
Each 800SP/40TMU/16ROP RV770 has 115GBps of bandwidth vs a 1600SP/80TMU/32ROP RV870 that has 160GBps.
Basically you are, possibly, more than doubling the potential throughput of the 5870 while only increasing bandwidth 39%, which was my original point.
Sure you don't have the inefficiencies of AFR and a dualGPU setup but you also run into other potential problems/bottlenecks.

I just have a feeling that we are in for a few surprises when we actually get to take a look at the architecture and performance, there is obviously quite a bit of tweaks/info that we haven't heard about/thought about.

PS- Before those Vantage scores were leaked, there was talk about Cypress running near 4890CF, which would put it in the same league as a GTX295.

[jaredpace]

charlie had the codenames wrong? just hemlock afaik and thats probably cause he listened to sylvie

jared, why would the 5870 be that much faster than the 5850? those numbers look weird, i cant figure out what he based his perf boost over 4890 on..
what shader count did he expect when he made that graph? it doesnt make any sense that with about double the hw he only predicts a 10% boost in some games?

and about 5870 beating gtx295...
4870 to 4870x2 scaling is 55% of a 295 to 90% of a 295... considering that a 5870 appears to have the same hw specs as a 4870x2 but isnt castrated by having it on 2 chips and there is no overhead, id expect a 5870 to be around 110% of a 295 like a 4870x2 would be, if it wouldnt have any overhead caused by being based on 2 chips... and thats assuming the 5870 comes clocked at 750mhz core and 3600mhz mem... according to the rumored specs it looks like like 800-900 clocks for the gpu and 5000mhz for the mem... so it will DEFINATELY beat the 295...

well it may beat the 295. like everyone else i dont know - just guessing. I did read on CH that the internal workings of the dual core provide 200% efficient scaling, so it's like a 4870x2 with perfect scaling. if that were the case it would be 110% of the 295 like you said. but i don't think they're actually going to get 200% of a 48701gb based on the rumor that cypress = 16000-18000 vantage and a 48701gb is like ~10,000.

dual gpu cards are more bandiwdth limited than single gpu cards... thats why there has been all this talk about getting gpus to share their memory capacity and bandwidth. cause right now, there is double the mem and double the mem bw on a dual gpu card, but it doesnt result in higher overall bw and higher overall memory capcity

I wonder if they have this fixed for the 5870x2? I think there will be an operational sideport to provide more bandwidth directly between the gpus, but I wonder if they will be sharing the memory?

[Calmatory]

True but there is a large gap between Juniper and Cypress so there should be something in between... atleast I hope so. Yields are never perfect so using salvaged part is better then to throw away chips which are semi-broken.

Right now the yields are lower than in the future. When the yields mature, the demand for salvaged chips is still higher than for those fully functional, so they have to salvage the chips and sell them with lower margins.

In short run in the launch having salvaged chips can be benefitical, but once the yields improve, salvaging the chips becomes more of a problem.

I guess AMD learnt this when they sold fully functional Barton cores as Durons, they had to sell the expensive core for cheaper and cripple it because there was huge demand for the Durons(Mainly because it was possible to mod the Duron to an Athlon, 64k L2 -> 256k L2. Those were the times. )

[Nedjo]

seriously its like they have no concept of commerce at all...

sure, lets them try how high they can price their cards and still sell some...
but im pretty sure if they price them as high as rumored lately, they will sell a lot less and overall make less profits than they would if theyd sell them cheaper...

Seriously saaya, to make your price analysis worthy you ought present some type of research to back up assumptions on equilibrium price-point of market demand and production output...

Yeah I know it sound difficult, and perhaps even AMD didn't do any research, perhaps they just look upon performance of their new parts, and performance of NV offering and decided to put some price-tag (not saying that what Theo And Charley are claiming is true prices - hell I'm pretty sure there's no way they could have official pricing leaked)

[saaya]

As to your first parts, you are correct, dual GPU cards like the 4870x2 don't increase the bandwidth vs a single GPU card but it also doesn't decrease it.
Each 800SP/40TMU/16ROP RV770 has 115GBps of bandwidth vs a 1600SP/80TMU/32ROP RV870 that has 160GBps.

i dont think thats how you can calculate it... by your math a 4870x2 has 1600sps 80TMUs 32RBEs and 230GBps bandwidth... but thats not the case... look at how xfire and sli actually scale and youll see that the extra tmus dont scale, and extra memory bandwidth doesnt scale either you only render half the scene or only every 2nd frame, but in either case you still have to fetch the same data into local memory as you dont know what texture will be on what part of the frame or if the texture will show on the even or odd or both frames in afr... the only thing that really scales with sli and xfire from what ive seen in reviews is: shader power. everything else is working exactly the same in single gpu and multi gpu mode... the gemoetry has to be calculated the same way, even if only half of it or only second frame gets rendered, and all the textures need to be fetched cause by that time you dont know yet if this gpu will render this texture or not... the diference in the rendering pipeline between single and multi gpu mode are so late that most of the work is being done twice with 2 gpus, and 4 times with 4 gpus... games that scale well with xfire and sli are shader power limited... or the gpus run in sfr, in which case geometry and texturing perf gets a boost too iirc... but sfr usually only gives you a 50% boost over single gpu... and you need to work through object dependencies which is a huge pita, thats why nvidia and ati push afr as much as they can cause it gives them better perf...

read what i said, the bw is actually used to fetch the very same textures and the very same infos twice... once for each gpu... the overall effective bandwidth for the 4870x2 and any dual gpu card is still the same bandwidth each gpu has, and actually even less for some reason in real world tests... like i said, dual gpu cards seem to actually be slightly more bw starved than single gpu cards of the same model and clocks and specs, most likely it has something to do with afr/sfr overhead or how sli/xfire work...

according to your math the 4870x2 has double the bandwidth of the 4870, true, but effective bw is the same as the 4870, just like effective memory capacity is the same as 4870 even though you got twice the mem on the card.

so a 5870 has 40% MORE effective bandwidth than a 4870x2 AND higher gpu clocks AND a tweaked gpu AND no xfire overhead... it will clearly be notably faster than a 4870x2...

correct me if im wrong, but thats the impression i got from testing myself and reading reviews and articles about sli and xfire...

[Foamy]

399$ = 275€ = cheap to me .

[eric66]

bs rumours everywhere ati does great job about keeping specs hidden