AMD To Launch Radeon HD4890 "RV790" In April

Printable View

Show 100 post(s) from this thread on one page

03-09-2009, 03:04 PM
ownage

Quote:

Originally Posted by Dainas

Well its plausible, with a return to the 512bit bus the RV790 could -just possibly- compete with a GTX 285... But that wouldnt be like this industry, where bending over and talking out one's.... err, talking big, is all the rage. 256bits of fail onward ho ATI.

Also, after all this time there ought to be a way to squeeze wider buses on a smaller, cheaper, although no less complex pcb.... so don't come at me with the cost effective philosophy, that was thrown out the window by GDDR4 and 5 use instead.

(RV770) AMD had to deal with pin-counts and bonding pads. 512bit would have needed a much bigger die-size just because of the pin-counts and bonding pads. Making the chip bigger would mean less chips per wafer, lower yields etc, higher prices etc. 512bit also doesn't add much performance.
03-09-2009, 03:55 PM
Nedjo

Quote:

Originally Posted by LordEC911

They cannot create a shader core with 1000SPs. It is 960 or 1040SPs.

why not?

Quote:

Originally Posted by Glow9

280 to 285 is 20% more or less so why not?

it's more "of" and much less "or" :yepp:

Quote:

Originally Posted by zerazax

RV790 sounds like the R520->R580 refresh than a new node at this point

and that was a quite of refresh:
http://www.xbitlabs.com/images/video.../cod_candy.gif

Quote:

Originally Posted by OBR

thisHD4890 is in my hands ...NOW and perf is higher only 10 percents above HD4870 1GB! Great improvement Great AMD! GTX 260 216 still better!

:rofl:

Quote:

Originally Posted by OBR

My own Sample of RV 790 is "like 4850" with clocks 850/900 MHz ... better model (like today 4870) with 950/975 MHz could be later, but maybe will never be launched !

:clap:

Quote:

Originally Posted by LOE

shaders are not that important actually, RV770 has a good amount of raw power, it needs more texturing power IMO, thats the bottleneck of the chip

WRONG LOE ;)

http://www.abload.de/img/understandinggraphicsad29f.jpg
03-09-2009, 04:28 PM
Hornet331

Quote:

Originally Posted by onethreehill

RV790 gets more efficient Powerplay
Source: ATI-Forum

Translated

uhng... not again... i only belive it when its really there. :down:
03-09-2009, 04:41 PM
LordEC911

Quote:

Originally Posted by Nedjo

why not?

Because of how the shader core is arranged.
RV770 has 10SIMDs, i.e. clusters, with 16 fat shaders in each cluster.
The next step up is 12SIMDs, 12*16=192*5=960.

If they add two more clusters, hit the clocks they are supposedly hitting and also reworked the ROPs this will be a nice filler before RV870 hits later this year.
03-09-2009, 05:44 PM
Dainas

Quote:

Originally Posted by SNiiPE_DoGG

I dont get it.... the 256bit gddr5 4870 has an excess of bandwidth... even at 900 core I don't see improvements in even synthetics more that 1-2% by raising my ram clocks 50 mhz.

50 mhz is hardly anything with GDDR these days, also GDDR5 on 256bit is one of those things that by principle should have, but dosent really measure up to 512bit GDDR3. Latency or whatnot, dunno. But the HD4870 isnt flooded in wasted bandwidth like the 2900XT was for its day.
03-09-2009, 06:58 PM
trinibwoy

Quote:

Originally Posted by Nedjo

why not?

The total has to be a multiple of the number of lanes per ALU and the number of ALUs per SIMD. Using 5 and 16 for those values as it stands today you can't arrive at 1000.

Quote:

WRONG LOE ;)

Vantage's texture tests are borked and disagree with every other texturing benchmark out there. FM has been "promising" to fix it for ages now. How did you manage to miss all the other texturing tests out there? I'm not convinced that RV770 is THAT texture bound though.
03-09-2009, 07:37 PM
SNiiPE_DoGG

Quote:

Originally Posted by Dainas

50 mhz is hardly anything with GDDR these days, also GDDR5 on 256bit is one of those things that by principle should have, but dosent really measure up to 512bit GDDR3. Latency or whatnot, dunno. But the HD4870 isnt flooded in wasted bandwidth like the 2900XT was for its day.

GDDR5 is QDR... 50mhz = 200mhz

and I repeat that even in synthetics you only see an increase within the margin of error if at all. 4870's gddr5 is well capable of 1200mhz + so care to explain to me why world record 4870x2 benches are only run at 1000mhz ? its because there is an excess of bandwidth.
03-09-2009, 07:41 PM
Vapor

Quote:

Originally Posted by SNiiPE_DoGG

and I repeat that even in synthetics you only see an increase within the margin of error if at all. 4870's gddr5 is well capable of 1200mhz + so care to explain to me why world record 4870x2 benches are only run at 1000mhz ? its because there is an excess of bandwidth.

The biggest reason is that it's pretty difficult to run the high vmem speeds with high system speeds :p:

Basically, the harder you push the CPU/system end of the PC, the lower the 4870X2 memory has to go. It's a weird phenomenon, but it definitely happens.
03-09-2009, 08:04 PM
B.E.E.F.

Quote:

Originally Posted by Vapor

TBasically, the harder you push the CPU/system end of the PC, the lower the 4870X2 memory has to go. It's a weird phenomenon, but it definitely happens.

So the higher the GPU overclock means I have to underclock the memory??
03-09-2009, 08:06 PM
SNiiPE_DoGG

Quote:

Originally Posted by B.E.E.F.

So the higher the GPU overclock means I have to underclock the memory??

without voltmodding yes.... although when I was playing with my voltmodded 4870s I never noticed the need to go below 1000mhz. I wish I had more experience playing with Vmodded ram but alas I am always to impatient to get that mod done once I've sorted out the core mod :D
03-09-2009, 09:53 PM
OnLine

Quote:

Originally Posted by Vapor

The biggest reason is that it's pretty difficult to run the high vmem speeds with high system speeds :p:

Basically, the harder you push the CPU/system end of the PC, the lower the 4870X2 memory has to go. It's a weird phenomenon, but it definitely happens.

Yes, I can only guess this problem is related to the GPU's memory controller.
Faster system, more stress on GPU's memory controller. :p:

//

Texture Units
Texture Fetch Units
Texture Adress Units
Texture Mapping Units
Pixel Shader Texture Units
Vertex Shader Texture Units
?

When card X has worse/better texturing performance than card Y, it could be under some specific circumstances.

I don't know the difference between these, I know nothing, I say nothing. Hopefully RV790 will be more efficient/faster than RV770 at both shaders & textures. :yepp:
03-09-2009, 10:03 PM
RVWinkle

If you look at the DDR 5 roadmaps you can come to the conclusion that memory bandwidth is going to be a non issue for the next 2 years or more. 256 bit differential ddr 5 will have something like 20x the bandwith as 512 bit ddr 3. In the same time period GPU performance with increase by 2x at best.
03-09-2009, 10:08 PM
largon

Quote:

Originally Posted by RVWinkle

If you look at the DDR 5 roadmaps you can come to the conclusion that memory bandwidth is going to be a non issue for the next 2 years or more. 256 bit differential ddr 5 will have something like 20x the bandwith as 512 bit ddr 3. In the same time period GPU performance with increase by 2x at best.

Meaning <256bit buses will become viable for performance parts. Within, say one year, a 128bit bus with conservative speed bin of differential GDDR5 will easily outperform current 256bit GDDR5 solutions, let alone 512bit GDDR3...
03-09-2009, 10:36 PM
RVWinkle

Quote:

Originally Posted by largon

Meaning <256bit buses will become viable for performance parts. Within, say one year, a 128bit bus with conservative speed bin of differential GDDR5 will easily outperform current 256bit GDDR5 solutions, let alone 512bit GDDR3...

I would say that the GTX285 will be the last video card you'll ever see with a 512bit bus but you know they're going to rename it in a few months lol.
03-09-2009, 10:48 PM
grimREEFER

remember ppl, 512bit and gddr5 are not mutually exclusive for next gen cards.
they will have silly amount of bandwidth. why? because they are from the effin future, and have to make current gen cards look slow.
03-10-2009, 12:56 AM
zerazax

Quote:

Originally Posted by grimREEFER

remember ppl, 512bit and gddr5 are not mutually exclusive for next gen cards.
they will have silly amount of bandwidth. why? because they are from the effin future, and have to make current gen cards look slow.

Um, having bandwidth but leaving it unused means that the bandwidth is just wasted.

And, since having larger memory buses means more room being wasted on the chip, they're going to opt for smaller buses so you can get more die per wafer, and ultimately make GPU's cheaper.
03-10-2009, 01:39 AM
gamervivek

Quote:

Originally Posted by SNiiPE_DoGG

GDDR5 is QDR... 50mhz = 200mhz

and I repeat that even in synthetics you only see an increase within the margin of error if at all. 4870's gddr5 is well capable of 1200mhz + so care to explain to me why world record 4870x2 benches are only run at 1000mhz ? its because there is an excess of bandwidth.

i guess calling it Graphics Double Data Rate was a stupid idea then?:rofl:
not sure about all benchmarks but memory bandwidth wouldn't do as much for benchmarks without AA,just like 4850 does well for its core speed when compared to 4870without AA but falls back more than the core speed difference when AA is set to 4x and especially 8x.
03-10-2009, 04:09 AM
SNiiPE_DoGG

Quote:

Originally Posted by gamervivek

i guess calling it Graphics Double Data Rate was a stupid idea then?:rofl:
not sure about all benchmarks but memory bandwidth wouldn't do as much for benchmarks without AA,just like 4850 does well for its core speed when compared to 4870without AA but falls back more than the core speed difference when AA is set to 4x and especially 8x.

I think youve got something with the AA ;) I wish I still had my single 4870 to play with under water..... then I could actually give us some real info to work with :D
03-10-2009, 04:28 AM
ToTTenTranz

Quote:

Originally Posted by Fudzilla

ATI's upcoming 55nm RV790 is the first new chip to come from ATI in the next quarter. We were told to expect it in April time and, at the same time, we learned that RV740, ATI's first desktop 40nm chip should start shipping a bit later.

ATI focuses on the mobile side and has Asus' first notebook design win with 40nm RV740, while desktop parts might start shipping either in late April or in early May time. ATI definitely have caught Nvidia by surprise as Nvidia simply didn’t believe that ATI will beat them to a new manufacturing process once again.

ATI was faster to transition to 65nm, 55nm and now even 40nm so Nvidia will have to wait for 32nm to try to beat ATI which is faster with new shrunk chips. We still have to see Nvidia’s first 40nm GPUs or even hear solid news about the state and availability of such products.

I guess that 40nm RV790 was just a typo after all :shrug:
03-10-2009, 04:55 AM
seanx

Quote:

Originally Posted by ownage

(RV770) AMD had to deal with pin-counts and bonding pads. 512bit would have needed a much bigger die-size just because of the pin-counts and bonding pads. Making the chip bigger would mean less chips per wafer, lower yields etc, higher prices etc. 512bit also doesn't add much performance.

maybe , but do u have proof of this ?
I remember stories of nvidia 512bit cards vs 256bit cards and the theory then
was they didn't "upgrade" the 256bit ones to 512bit purely as a marketing
point , since the alleged preformance of 256bit card (upgraded to 512bit) would be equal to the standard 512bit one....

And I aggree with the idea that the 4870x2 for example is suffering from the
256bit bottle neck ....seems to make sense.
just my useless 5c lol..
03-10-2009, 07:26 AM
Shadowmage

Quote:

Originally Posted by seanx

maybe , but do u have proof of this ?

Yes. Dave Baumann, who's a senior manager at AMD/ATI, confirmed this on Beyond3D forums.
03-10-2009, 07:49 AM
ownage

Quote:

Originally Posted by seanx

maybe , but do u have proof of this ?

Quote:

Originally Posted by Shadowmage

Yes. Dave Baumann, who's a senior manager at AMD/ATI, confirmed this on Beyond3D forums.

And CJ, another ATI guy.
03-10-2009, 07:58 AM
Hell Hound

Um,there cores are not fast enough so they sing some other sad song.Geforce used 512bit and performance was great.Dont give us excuses give us POWER.:D Plus they are a lot of games that love bandwidth,So how is it wasted.
03-10-2009, 08:07 AM
SNiiPE_DoGG

Quote:

Originally Posted by Hell Hound

Um,there cores are not fast enough so they sing some other sad song.Geforce used 512bit and performance was great.Dont give us excuses give us POWER.:D

Different architectures, different die sizes, different cards. that doesnt really make any sense.
03-10-2009, 08:09 AM
ownage

Quote:

Originally Posted by Hell Hound

Um,there cores are not fast enough so they sing some other sad song.Geforce used 512bit and performance was great.Dont give us excuses give us POWER.:D Plus they are a lot of games that love bandwidth,So how is it wasted.

The pros simply have to outweigh the cons.
They could have made the RV770 'virtually' (As CJ described) bigger/larger so pin-counts and bonding pads wouldn't be a problem, it also would have result in a much cooler chip. The same chip, but just made bigger. The pros of a small economical die was probably more rewarding than a big 512bit chip.

Show 100 post(s) from this thread on one page

All times are GMT -8. The time now is 06:06 AM.

XtremeSystems