ATI 4xxx Series Thread

[Janooo]

Another slide with 800sp's.

 

[Munky]

What's with all the talk of crappy, inefficient shaders? The r6xx cards have enough shader performance to be competitive, the problem is the cards are bottlenecked in other areas, such as texturing and z-output. The r6xx cards do well in 3dmock precisely because it stresses shaders more than typical games.

 

[SlowSpyder]

Originally posted by: SickBeast

Originally posted by: allies

Originally posted by: SickBeast

Originally posted by: taltamir
everyone is using older stuff in their architecture... rebuilding from scratch is an insane endeavor that will lead to nowhere fast.

With the R600-based stuff as borked as it is, they should be more drastically revamping the architecture if not starting over. The shader architecture is horribly inefficient, the GPU lacks TMUs, it cannot perform proper AA, plus they wasted a ton of die space making the thing 512-bit unnecessarily.

NV started from scratch after their 5900-series cards and came out with the 6800-series which was much better. The G80 is also a new architecture, and again it performs quite well.

From what I understand, any semiconductor company pretty much has to 'start from scratch' every time there is a major die shrink involved. Usually they make a part, then to a minor die shrink refresh, then make a new part altogether.

By saying they should start over, I obviously don't mean that they should throw away everything they know about GPU design. I simply mean that the parts need to be completely revamped.

The R770 already addresses most of the points you make:

• (rumor)Revamped AA
  2x TMUs
  More shaders (maybe tweaking involve)
  256mbit bus with GDDR5

What more do you want again

I want proper shaders, not 800 crappy ones! :evil:

I suppose that's my main complaint with the recent AMD architectures. Their shader design is terrible. If they're simply adding 500 more crappy and inefficient shaders, I really don't see the point.

It just seems like it has taken AMD a long time to work the bugs out of R600. IMO a completely revamped GPU with better shaders and proper AA would have been the way to go. Hopefully the 4850/70 are just that. I'm not holding my breath. :beer:

If every fifth shader is a full fledged shader, then if it has 800 shaders that means worst case the 4800 cards will have 160 shaders working for it, right? With the addition of the simpler shaders that games can be coded to take advantage of, or I imagine at the driver level improvements can be made, there could be even more shader power then the 160 full fledged shaders in better case scenarios, I don't think that will be a bottle neck.

I still haven't seen what the big problem is with AA being done through the shaders, maybe I'm missing some benchmarks that are out there? I thought that at 8xAA the 3870 out performed the 8800 cards by a fair margin? Also, wasn't AA through the shaders a DX10.1 thing??? I could be wrong on that, but I thought it was? At any rate, the 48x0 cards are said to implement AA differently anyway.

I think AMD rushed the R600 out the door to get a product out there, that's why it seems to be as unblanaced as it is. I think it's a great architecture for them to build and expand on, as it had it's strong points and AMD is addressing it's weak points with the 48x0 cards. Infact I have an R600 based card and love it... there isn't any game I can't play at 1680x1050 without any problems at good settings. Even Crysis plays fine for me at 1680x1050 all settings at 'High' no AA. But, I still play all my games on XP, so no idea how DX10 does, I'm sure that takes a bigger toll then DX9. My biggest decision will be do I keep my 2900 Pro that works great or upgrade? I think I'm going to take a pass on the Nvidia hardware again this round, unless when the NDA's lift the new Nvidia stuff turns out to good to ignore.

 

[Janooo]

Originally posted by: SlowSpyder
...

If every fifth shader is a full fledged shader, then if it has 800 shaders that means worst case the 4800 cards will have 160 shaders working for it, right? With the addition of the simpler shaders that games can be coded to take advantage of, or I imagine at the driver level improvements can be made, there could be even more shader power then the 160 full fledged shaders in better case scenarios, I don't think that will be a bottle neck.

I still haven't seen what the big problem is with AA being done through the shaders, maybe I'm missing some benchmarks that are out there? I thought that at 8xAA the 3870 out performed the 8800 cards by a fair margin? Also, wasn't AA through the shaders a DX10.1 thing??? I could be wrong on that, but I thought it was? At any rate, the 48x0 cards are said to implement AA differently anyway.

I think AMD rushed the R600 out the door to get a product out there, that's why it seems to be as unblanaced as it is. I think it's a great architecture for them to build and expand on, as it had it's strong points and AMD is addressing it's weak points with the 48x0 cards. Infact I have an R600 based card and love it... there isn't any game I can't play at 1680x1050 without any problems at good settings. Even Crysis plays fine for me at 1680x1050 all settings at 'High' no AA. But, I still play all my games on XP, so no idea how DX10 does, I'm sure that takes a bigger toll then DX9. My biggest decision will be do I keep my 2900 Pro that works great or upgrade? I think I'm going to take a pass on the Nvidia hardware again this round, unless when the NDA's lift the new Nvidia stuff turns out to good to ignore.

HQAF is the biggest issue for the 6xx cards.
When you scroll down you can see where it sinks.
HQ -> AI switched off.

 

[Nemesis 1]

Originally posted by: SickBeast

Originally posted by: allies

Originally posted by: SickBeast

Originally posted by: taltamir
everyone is using older stuff in their architecture... rebuilding from scratch is an insane endeavor that will lead to nowhere fast.

With the R600-based stuff as borked as it is, they should be more drastically revamping the architecture if not starting over. The shader architecture is horribly inefficient, the GPU lacks TMUs, it cannot perform proper AA, plus they wasted a ton of die space making the thing 512-bit unnecessarily.

NV started from scratch after their 5900-series cards and came out with the 6800-series which was much better. The G80 is also a new architecture, and again it performs quite well.

From what I understand, any semiconductor company pretty much has to 'start from scratch' every time there is a major die shrink involved. Usually they make a part, then to a minor die shrink refresh, then make a new part altogether.

By saying they should start over, I obviously don't mean that they should throw away everything they know about GPU design. I simply mean that the parts need to be completely revamped.

The R770 already addresses most of the points you make:

• (rumor)Revamped AA
  2x TMUs
  More shaders (maybe tweaking involve)
  256mbit bus with GDDR5

What more do you want again

I want proper shaders, not 800 crappy ones! :evil:

I suppose that's my main complaint with the recent AMD architectures. Their shader design is terrible. If they're simply adding 500 more crappy and inefficient shaders, I really don't see the point.

It just seems like it has taken AMD a long time to work the bugs out of R600. IMO a completely revamped GPU with better shaders and proper AA would have been the way to go. Hopefully the 4850/70 are just that. I'm not holding my breath. :beer:

This statement has me abit concerned as to whats actually happening here . In this thread.

I suppose that's my main complaint with the recent AMD architectures. Their shader design is terrible. If they're simply adding 500 more crappy and inefficient shaders, I really don't see the point.
So what do you know about the shaders in the new 4000seies of gpu's . If nothing why are you so intent at labling ATI shaders as shiity on the 4000's.
I was almost banned for talking about 10.1 and NV not having the capability to use it.

So in a DX10.1 game would those shitty shaders work better or not. OR are you to against unified shaders. Until we get a game that runs DX10.1 correctly we won't know how bad those ATI shaders are. NV can hold back progress only so long befor it catches up to them . AND it will. But once we get that game we can go back and see how NV held back progress threw the himtbp programm.

I had to look to see if I was in the ATI thread. With all the negitive remarks about a card you know next to nothing about.

I had the 800xtpe 4 months befor almost all others . it was great card. the 4870 isn't going to disappoint anyone. except NV .

 

[chewietobbacca]

http://my.ocworkbench.com/2008...MD_HD4850_HD4870-1.htm

edit: thakns for fixing link PC Surgeon

 

[NoStateofMind]

http://my.ocworkbench.com/2008...MD_HD4850_HD4870-1.htm

 

[Kuzi]

I wasn't checking the thread for the last few days, but I noticed everyone now is saying that the RV770 chips will have 800 SPs instead of 480.

The 3870 had 320 SPs and was bottlenecked for having only 16 TMUs. So having 800 SPs and only 32 TMUs on the new 4850/4870 chips will make the bottleneck even "worse".

Unless I'm missing something. I highly doubt they will have 800 SPs, 480 SPs seems more realistic, especially that ATI is still using the 55nm process.

 

[Nemesis 1]

Do you understand dx10.1 those shaders are way more than u think they are. There unified. So adding 16 more TMUs. will add nothing to the 4000series Let me ask ya this. higher clocked shaders. higher clocked gpu. If the memory improvements have been made and its pooled . I think you haven't a clue. Plus the other improvements. The 2 dies vs . one die is the stupidest argument I ever heard. As long as it 1 card its 1 gpu. Sharing all its resources . A bit differant than what use to be .

What if 4870 more than doubles the performance of 3870 how far ahead in performance do you think NV is . LOL.

 

[golem]

I thought both NV and Ati had unified shaders?

 

[RussianSensation]

So it appears on the GPU side the only difference between 4850 and 4870 are clock speeds?

4850 625mhz vs. 4870 750mhz (+20%) => 1000 vs. 1200 gigabflops (+20%)?

We also know NV hit 933 GigaFlops. Of course you cannot simply compare theoretical shader processing rate across competitors in real world gameplay.

1. 8800GT (336GFlops) card >>> HD3870 (496GFlops)
2. 9800 GX2 (768Gflops) >>>> HD 3870 X2 (1056GFlops)

From this we can also deduce that HD4870 should be slightly faster than HD3870 X2 (let's assume 1.2x of its performance) if we simply compare AMD to AMD. We have already seen slides from NV that GTX 280 is 80-120% faster than 9800 GX2. 9800 GX2 is itself 20-30% faster than HD3870 X2 (or lets say 1.2x the performance). Simplifying this we get:

HD 4870 = 9800 GX2.

Unless ATI can scale 80-100%, we should expect the GTX card to outperform 4870 x2s somewhat. It will then rest on ATI to actually be able to have 80% scaling across a variety of current games....

 

[dennilfloss]

Originally posted by: Kuzi
I wasn't checking the thread for the last few days, but I noticed everyone now is saying that the RV770 chips will have 800 SPs instead of 480.

The 3870 had 320 SPs and was bottlenecked for having only 16 TMUs. So having 800 SPs and only 32 TMUs on the new 4850/4870 chips will make the bottleneck even "worse".

Unless I'm missing something. I highly doubt they will have 800 SPs, 480 SPs seems more realistic, especially that ATI is still using the 55nm process.

It's a trap. Someone has hinted at 45nm process being involved.:laugh:

To repeat a quote from a Rage3D member about the 4870X2

Now witness the power of this fully armed and operational battle station!

 

[Nemesis 1]

Originally posted by: RussianSensation
So it appears on the GPU side the only difference between 4850 and 4870 are clock speeds?

4850 625mhz vs. 4870 750mhz (+20%) => 1000 vs. 1200 gigabflops (+20%)?

We also know NV hit 933 GigaFlops. Of course you cannot simply compare theoretical shader processing rate across competitors in real world gameplay.

1. 8800GT (336GFlops) card >>> HD3870 (496GFlops)
2. 9800 GX2 (768Gflops) >>>> HD 3870 X2 (1056GFlops)

From this we can also deduce that HD4870 should be slightly faster than HD3870 X2 (let's assume 1.2x of its performance) if we simply compare AMD to AMD. We have already seen slides from NV that GTX 280 is 80-120% faster than 9800 GX2. 9800 GX2 is itself 20-30% faster than HD3870 X2 (or lets say 1.2x the performance). So simplyfing things for plausibility perspective we get:

HD 4870 = 9800 GX2.

Unless ATI can scale 80-100%, we should expect the GTX card to outperform 4870 x2s somewhat. It will then rest on ATI to actually be able to have 80% scaling across a variety of current games....

Which games are you talking about . Seems to me ATI hands NV abeating in some games why is that. Xfire scales better than sli . But according to Available info. The 4870x2 has better interconnects /buffering / Full memory utilization of both cores. Crysis the end all benchmark. According to NV . A game that was created using NV hardware for NV . Just doesn't cut it in my book . Ya I have the game . Yes its pretty . But other than that it sucks. Now a new simulator game would be interesting that really pushes all the hardware in both ATi/Nv gpus. But alot of those games were developed using ATI GPUs.

 

[Nemesis 1]

Originally posted by: golem
I thought both NV and Ati had unified shaders?

No NV doesn't have unified shaders. NV did manage to get MS to remove DX10.1 from the orginal DX10 spec. Because they didn't have the hardware.

 

[Extelleron]

Originally posted by: Nemesis 1

Originally posted by: golem
I thought both NV and Ati had unified shaders?

No NV doesn't have unified shaders. NV did manage to get MS to remove DX10.1 from the orginal DX10 spec. Because they didn't have the hardware.

This part here just isn't true.

GeForce 8 cards have unified shading processors just like AMD's R600-based cards do. I'm not sure what you are getting at. The stream processors in G80 are capable of processing either pixel, vertex, or geometry shaders. This is different in comparison to GeForce 7 / X1900 because those cards had seperate units for vertex / pixel shading, and did not do geometry shading at all because they were not DX10 spec.

 

[golem]

Originally posted by: Extelleron

This part here just isn't true.

GeForce 8 cards have unified shading processors just like AMD's R600-based cards do. I'm not sure what you are getting at. The stream processors in G80 are capable of processing either pixel, vertex, or geometry shaders. This is different in comparison to GeForce 7 / X1900 because those cards had seperate units for vertex / pixel shading, and did not do geometry shading at all because they were not DX10 spec.

Thanks Extelleron, that was my understanding of the term unified shaders also.

 

[Extelleron]

We do have one benchmark of HD 4850 now, compared to HD 3870.

Perlin Noise (ALU test in 3D Mark)

HD 3870: 175
HD 4850: 335

So ~91% higher ALU performance than HD 3870. Confirms the 800SP rumor; 800SP @ 625MHz gives roughly 2.01x 320SP @ 775MHz, so the real-world increase is very close to theoretical.

Considering HD 4870 will be clocked 20% higher than the 4850, it should have higher ALU performance than the 9800GX2, which scores ~380 in Perlin Noise.

http://forum.beyond3d.com/show...1173914&postcount=3041

 

[Fallen Kell]

Originally posted by: Extelleron
We do have one benchmark of HD 4850 now, compared to HD 3870.

Perlin Noise (ALU test in 3D Mark)

HD 3870: 175
HD 4850: 335

So ~91% higher ALU performance than HD 3870. Confirms the 800SP rumor; 800SP @ 625MHz gives roughly 2.01x 320SP @ 775MHz, so the real-world increase is very close to theoretical.

Considering HD 4870 will be clocked 20% higher than the 4850, it should have higher ALU performance than the 9800GX2, which scores ~380 in Perlin Noise.

http://forum.beyond3d.com/show...1173914&postcount=3041

Which still doesn't seem like it will be enough if the performance of the GTX280 really is 80-120% faster than a 9800GX2 (as most people are saying/showing). As much as I hope an ATI product will come out and take the crown, I just don't know if it will be this round yet again.

 

[Extelleron]

Originally posted by: Fallen Kell

Originally posted by: Extelleron
We do have one benchmark of HD 4850 now, compared to HD 3870.

Perlin Noise (ALU test in 3D Mark)

HD 3870: 175
HD 4850: 335

So ~91% higher ALU performance than HD 3870. Confirms the 800SP rumor; 800SP @ 625MHz gives roughly 2.01x 320SP @ 775MHz, so the real-world increase is very close to theoretical.

Considering HD 4870 will be clocked 20% higher than the 4850, it should have higher ALU performance than the 9800GX2, which scores ~380 in Perlin Noise.

http://forum.beyond3d.com/show...1173914&postcount=3041

Which still doesn't seem like it will be enough if the performance of the GTX280 really is 80-120% faster than a 9800GX2 (as most people are saying/showing). As much as I hope an ATI product will come out and take the crown, I just don't know if it will be this round yet again.

We're talking about a pure ALU test here. No way will GTX 280 be 80-120% faster in an ALU test. This test has very little dependence on bandwidth or other performance (though it does matter a bit, as seen by 8800 Ultra edging out 8800GTS 512MB). GTX 280 should not be much faster than 9800GX2 here as SLI/CF scaling is perfect and 9800GX2 still has more SP's (256 vs 240) and higher clock (1500 vs 1296). GTX 280 should be faster per clock though so it will probably edge out the 9800GX2 by a good bit.

If you look at performance here: http://www.firingsquad.com/har..._performance/page8.asp

If you take the HD 4850's supposed performance increase over the 3870 (95% of theoretical gain) then HD 4870 should score ~50 FPS in Perlin Noise. That makes it considerably faster than the 9800GX2 and ~2x 8800 Ultra in this test. I would not be surprised if GTX 280 and HD 4870 have comparable performance in this ALU benchmark, and certainly the 4870 should equal or edge out the GTX 260.

Of course this is an ALU test, not overall performance; this will measure shader performance and nothing else. We see that clearly as HD 3870 already edges out the 8800GT slightly in Perlin Noise. But in pure ALU performance, I believe single-GPU HD 4870 will be very close to GTX 280.

 

[SlowSpyder]

Originally posted by: RussianSensation
So it appears on the GPU side the only difference between 4850 and 4870 are clock speeds?

4850 625mhz vs. 4870 750mhz (+20%) => 1000 vs. 1200 gigabflops (+20%)?

We also know NV hit 933 GigaFlops. Of course you cannot simply compare theoretical shader processing rate across competitors in real world gameplay.

1. 8800GT (336GFlops) card >>> HD3870 (496GFlops)
2. 9800 GX2 (768Gflops) >>>> HD 3870 X2 (1056GFlops)

From this we can also deduce that HD4870 should be slightly faster than HD3870 X2 (let's assume 1.2x of its performance) if we simply compare AMD to AMD. We have already seen slides from NV that GTX 280 is 80-120% faster than 9800 GX2. 9800 GX2 is itself 20-30% faster than HD3870 X2 (or lets say 1.2x the performance). Simplifying this we get:

HD 4870 = 9800 GX2.

Unless ATI can scale 80-100%, we should expect the GTX card to outperform 4870 x2s somewhat. It will then rest on ATI to actually be able to have 80% scaling across a variety of current games....

Don't for get the video memory difference. The GPU cores seem very similar, just a clock speed difference. But, the 4870 will have GDDR5 memory (latest rumor says 3.6GHz GDDR speed) while the 4850 will use GDDR3 memory (latesr rumor I saw was 2.0GHz GDDR speed). So, the 4870 will have a faster core out of the box, a better cooler, and much, much more memory bandwidth.

So, let's assume all the numbers are correct, that the only difference between the cores is a mere 125MHz, but the difference in the memory is 1.6GHz (both using the same 256bit bus). With that we can pretty much conclude that either the 4850 is going to be somewhat bandwidth starved, or the 4870 will have more bandwidth then it needs. I don't see a 125MHz difference in core speed warranting a 1.6GHz change in memory speed... my gut says that the 4850 will be somewhat memroy starved, and the 4870 will pull away from the 4850 by a good margin, more then just the 125MHz core difference would account for due to having the memory bandwidth available to it that it needs vs. the 4850 that will probably have to wait on the memory at times.

 

[chewietobbacca]

No one is really saying GTX280 is going to be 80-120% faster than the 9800GX2. Most leaks and sources I've heard from say on average 50-60% faster and that can easily be double if the game doesn't scale well w/ SLI/multi-GPU configs and < 30-40% if the game does scale well

 

[Extelleron]

Originally posted by: chewietobbacca
No one is really saying GTX280 is going to be 80-120% faster than the 9800GX2. Most leaks and sources I've heard from say on average 50-60% faster and that can easily be double if the game doesn't scale well w/ SLI/multi-GPU configs and < 30-40% if the game does scale well

People are probably mistaking the nVidia slide comparing GTX 200 performance to the 3870 X2 as comparing performance to the 9800GX2. GTX 280 showed around that much of a performane lead over the GTX 280 in most apps. Most around 80%.

 

[SolMiester]

Originally posted by: Kuzi
I wasn't checking the thread for the last few days, but I noticed everyone now is saying that the RV770 chips will have 800 SPs instead of 480.

The 3870 had 320 SPs and was bottlenecked for having only 16 TMUs. So having 800 SPs and only 32 TMUs on the new 4850/4870 chips will make the bottleneck even "worse".

Unless I'm missing something. I highly doubt they will have 800 SPs, 480 SPs seems more realistic, especially that ATI is still using the 55nm process.

I agree

 

[Martimus]

Originally posted by: Nemesis 1

Originally posted by: golem
I thought both NV and Ati had unified shaders?

No NV doesn't have unified shaders. NV did manage to get MS to remove DX10.1 from the orginal DX10 spec. Because they didn't have the hardware.

All direct X 10 cards use unified shaders. What nvidia does not do is use the shaders for AA, which is required in DX10.1. That is where you are confused.

 

[Cookie Monster]

Theres no link between the DX API and unified shader architecture. You could still have your fixed function GPU ala with pipelines that supports DX10 if it meets the requirements.