5900 series behind the corner

[SKYMTL]

I am also getting weird 2560x1600 results ( I suppose that is what you are talking about) and I see that nobody has a working GPU-Z for these babys. However, I have the feeling that the problem is not with the memory bus, and from my math we should be loking at a 2x256 bit card. Bu anyway, we'll see more about that once the reviews are out.

You're actually the 4th editor who has confirmed the issues with me. Honestly, the card is behaving like it can't access its full memory alotment in situations that need a large framebuffer. It IS supposed to be a 2x 256-bit bus but ironically the ATI documentation does not mention anything about the bus...just clock speeds. That's a contrast to the HD 5800 and HD 5700 series documents that showed quite clearly their 256 and 128 bit specs.

[emertX]

You don't really think the card has a 128 bit interface, do you???

[Monstru]

You're actually the 4th editor who has confirmed the issues with me. Honestly, the card is behaving like it can't access its full memory alotment in situations that need a large framebuffer. It IS supposed to be a 2x 256-bit bus but ironically the ATI documentation does not mention anything about the bus...just clock speeds. That's a contrast to the HD 5800 and HD 5700 series documents that showed quite clearly their 256 and 128 bit specs.

Please believed me that I had the EXACT same feeling when I was reading the documentation and I noticed that the bus width is not there. Also, if you think that there is no GPU-Z out there capable of reading the specs correctly, you might get a "fishy" image about all this. So I also gave this a lot of thinking today, but I strongly think that the bus is 2 x 256 bit, due to the bandwidth that they show in the specs. However, the fact that a card which has a key selling point in working with 3 2560x1600 monitors, but however performs slower exactly in that resolution is at least...ironic
Luckily for ATI, the performance in 1920 and 1680 compensates for the behaviour in 2560 in some games, and most of the gamers do not own a 30" monitor yet.

LE - unbelievable - GPU-Z 0.3.6 does not work with ATi reference samples but it does work with retail cards...sigh...

[SKYMTL]

You don't really think the card has a 128 bit interface, do you???

Probably not, but it DOES behave like it is using a 2x128-bit interface or that it is in some way crippled at high resolutions.

[SKYMTL]

Please believed me that I had the EXACT same feeling when I was reading the documentation and I noticed that the bus width is not there. Also, if you think that there is no GPU-Z out there capable of reading the specs correctly, you might get a "fishy" image about all this. So I also gave this a lot of thinking today, but I strongly think that the bus is 2 x 256 bit, due to the bandwidth that they show in the specs. However, the fact that a card which has a key selling point in working with 3 2560x1600 monitors, but however performs slower exactly in that resolution is at least...ironic
Luckily for ATI, the performance in 1920 and 1680 compensates for the behaviour in 2560 in some games, and most of the gamers do not own a 30" monitor yet.

What I am more interested in knowing is what is holding it back. Especially versus a HD 5850 Crossfire setup. Another possibility is memory timings but I am just brainstorming here.

1920 and 1680 resolutions are besides the point here since performance will be CPU limited even at 4 / 8x AA in literally every game. Trust me, with a 4Ghz i7, the graphs look pretty darn flat...

[Manicdan]

the 128 has to be a typo. read the context

this graphics card again may use 128-bit memory bus interface.

hes comparing to the 5870 specs, and must have wrote down the wrong thing.

[SKYMTL]

the 128 has to be a typo. read the context



hes comparing to the 5870 specs, and must have wrote down the wrong thing.

Yeah, but the thought did cross my mind prior to reading that. Monstru does make a good point about the bandwidth numbers though; they indicate a 2x 256-bit layout.

My search continues for what's crippling performance...

[gurneyhck]

What about the fact that it has only 2GB (1GB per GPU) ? Doesn't that have anything to do with it?

[Monstru]

My graphs really do not look flat in FC2 or Hawx at 1920

Anyway, the bandwidth might be enough for 5850, due to the crippled Cypress, but it is not enough for 2 "full-option" Cypress like HD 5970 has. What is changed from HD5870? The bus is the same, the number of shader procesors is the same, the only different things are the clocks. Which give less bandwidth/ GPU compared to HD 5870.

[iTravis]

Hearing how you guys talk about cripple performance at 2560x1600 makes me even more depressed.

[eric66]

What I am more interested in knowing is what is holding it back. Especially versus a HD 5850 Crossfire setup. Another possibility is memory timings but I am just brainstorming here.

1920 and 1680 resolutions are besides the point here since performance will be CPU limited even at 4 / 8x AA in literally every game. Trust me, with a 4Ghz i7, the graphs look pretty darn flat...

can you tell us which game are those

[SKYMTL]

My graphs really do not look flat in FC2 or Hawx at 1920

Anyway, the bandwidth might be enough for 5850, due to the crippled Cypress, but it is not enough for 2 "full-option" Cypress like HD 5970 has. What is changed from HD5870? The bus is the same, the number of shader procesors is the same, the only different things are the clocks. Which give less bandwidth/ GPU compared to HD 5870.

Yeah, Warhead either I guess.

DX10 is another matter altogether.

The performance issues could be due to loosened timings on the memory. Has anyone been able to read the voltage used for the GDDR5. I haven't got there in my testing but it is possible that ATI lowered the GDDR5 voltage in order to keep power consumption under 300W. Therefore, they also needed to loosen timings in order to stay at the desired speed.

Man, I'm grasping for straws here...

[Hornet331]

ble i really hope ati gets the shortage of cards sorted out with the release of the 5970, just looked at retailers and nothing there... well yeah maybe for 450€+ (a single 5870)

[gurneyhck]

Yeah, Warhead either I guess.

DX10 is another matter altogether.

The performance issues could be due to loosened timings on the memory. Has anyone been able to read the voltage used for the GDDR5. I haven't got there in my testing but it is possible that ATI lowered the GDDR5 voltage in order to keep power consumption under 300W. Therefore, they also needed to loosen timings in order to stay at the desired speed.

Man, I'm grasping for straws here...

That is a good possibility. But if it is a memory "issue" then u can test this by OC'ing the memory and see if it still performs "strange", right?

[eric66]

Yeah, Warhead either I guess.

DX10 is another matter altogether.

The performance issues could be due to loosened timings on the memory. Has anyone been able to read the voltage used for the GDDR5. I haven't got there in my testing but it is possible that ATI lowered the GDDR5 voltage in order to keep power consumption under 300W. Therefore, they also needed to loosen timings in order to stay at the desired speed.

Man, I'm grasping for straws here...

how about giving some voltage and then testing it

[Monstru]

Yeah, Warhead either I guess.

DX10 is another matter altogether.

The performance issues could be due to loosened timings on the memory. Has anyone been able to read the voltage used for the GDDR5. I haven't got there in my testing but it is possible that ATI lowered the GDDR5 voltage in order to keep power consumption under 300W. Therefore, they also needed to loosen timings in order to stay at the desired speed.

Man, I'm grasping for straws here...

1.10v for the GDDR.

The problem has only one source, one big fat engineering...let's call it compromise. The memory clocks. 1000MHz with 2 x 256 bit bus simply does not cut it for these beasts. Of course, they clocked the memory down so they can lower the volts, so they can keep the power consumption and heat disipation low.

Solution is pretty darn simple. 1.15v in GDDR, 1200MHz on the GDDR, and 2560x1600 is back on track, that's what I did

[SKYMTL]

1.10v for the GDDR.

The problem has only one source, one big fat engineering...let's call it compromise. The memory clocks. 1000MHz with 2 x 256 bit bus simply does not cut it for these beasts. Of course, they clocked the memory down so they can lower the volts, so they can keep the power consumption and heat disipation low.

Solution is pretty darn simple. 1.15v in GDDR, 1200MHz on the GDDR, and 2560x1600 is back on track, that's what I did

Memory is @ the same speed as the HD 5850 but two HD 5850 cards beat it in high resolutions most of the time.

[Piotrsama]

Memory is @ the same speed as the HD 5850 but two HD 5850 cards beat it in high resolutions most of the time.

Drivers issues maybe?

[zalbard]

From the techtree article. If this 2x128-bit memory interface information is correct, expect a MASSIVE hit in performance in high texture DX10 and DX11 games.

EDIT: that would also explain some results some editors (myself included) have been getting.

No way, I don't believe in 128 bit. Why would they do it? Makes absolutely no sense at all. Also I think the performance loss would be brutal... in all resolutions.

[Monstru]

Memory is @ the same speed as the HD 5850 but two HD 5850 cards beat it in high resolutions most of the time.

Because the 5850's GPU does less operations/second compared to a Cypress. Imagine that HD 5970's GPU's can process more data then 5850's GPU, but the bandwidth is not enough to insure an optimum flow of data. Now, I am not a programmer, but what does that lead us to?

[Calmatory]

Cant imagine how hard it is going to be to get our hands on these, considering the availability of 5850/70's

Dell hasn't ordered any of these, so all the manufactured products go directly to the usual channels.

[Astennu]

Strange 1000 MHz GDDR5 should be enough for a 725 MHz Cypress core with 1600 shaders. I think its a driver isssue.

Look at those overclockers that do 1300-1400 core with only 1250-1300 MHz memory. And still have a huge boost. at the start i thought the RV870 and the RV840 where memory bandwith bottelneked but if you look at the OC results you gain way more by overclocking the core.

I think and hope this is a driver issue because i have a 30" screen. And a HD5850 @ 900-1125 is to slow in some games

[SKYMTL]

Because the 5850's GPU does less operations/second compared to a Cypress. Imagine that HD 5970's GPU's can process more data then 5850's GPU, but the bandwidth is not enough to insure an optimum flow of data. Now, I am not a programmer, but what does that lead us to?

I don't think we are talking about the same thing.

This card is supposedly = 2x HD 5870 cores underclocked to HD 5850 specs.

Bandwidth (on paper) should therefore be ~ the same as two HD 5850 1GB cards.

Therefore this card should = 2x HD 5850 (+/-). However, in higher resolutions in DX10 games, it looks to be getting spanked by two HD 5850 cards.

.
.
.
.

As for "driver issues", I don't play along with that. In newer games, sure. However, we are talking about HawX DX10 and Far Cry 2 DX10...two games that were released months and months and months ago. If ATI is having "driver issues", that is simply an embarrassment of epic proportions.

[eric66]

have u tried to overclock memory ? or gpu or bring it to 5870 specs ? all you do is whinne here lol

[zalbard]

As for "driver issues", I don't play along with that. In newer games, sure. However, we are talking about HawX DX10 and Far Cry 2 DX10...two games that were released months and months and months ago. If ATI is having "driver issues", that is simply an embarrassment of epic proportions.

That actually is. Crossfire almost didn't work at all before 9.10 final for 5870. In all games. So I'd count on that.