ATI Radeon HD 4000 Series discussion

[Papu]

1) Poor FSAA performance. The R6x0 series cards were dogged with FSAA performance issues.

from whats been said in support this has been drastically improved

2) Lack of Physx support, yes it is not really a big deal now.....but will it become one?

ati will support havok on there gpu's so maybe its nvidias lack of havok thats the bigger issue?

3) ATi compatibility issues, this might sound like a mindless fanboy rant here but back in the day of my Radeon 9700 I had a few issues with games due to a lack of support by the Devs for ATi cards (SimCity 4 and Metal Gear Solid). Later the X1800XT had issues to, due to not quite fully supporting Shader Model 3 (a lack of Vertex Texture Fetching). I just dont want to end up getting a card which lacks dev support or something...

back in the day...nuff said..

4)Heat and Noise, lets face it the X2900XT cards were blimin hot and noisy.

good thing the 4850 isnt a 2900 , its improved and 55nm

[zerazax]

Well if that architecture shot is a fake, then its a pretty good fake

[v_rr]

Simple mathematics

several fillrate tests from 4850 in the wild show ~19500MT/s and clock of the Card (TMUs) is 625Mhz

with 32TMUs and 625mhz you get a theoretical fillrate of 20.000MT/s
with 40TMUs at 625mhz you get 25.000MT/s.

i dont think RV770 is that inefficent.

So your first comment was not true
I know that to and has logic, but keep in mind that are separate things.

Anyone dont´know if that is right or not.

TMU´s can be in 32/40 number. Anyway is a very nice jump from 16.

[STaRGaZeR]

TMU´s can be in 32/40 number. Anyway is a very nice jump from 16.

And if nothing else has changed they will be still insufficient. 800SPs/32TMUs ratio is worse than 320SPs/16TMUs. If 16 were not enough before, 32 will be even worse now because of the huge number of SPs. If the chip has 40TMUs instead, it will be the same as before. ATI better has done something to that TMUs or we'll see the same bottleneck in texture perfomance

[zerazax]

Well the raw # of TMU's is a huge boost regardless. And do keep in mind that at one point or another, having more TMU's is pointless if the games dont use them all. Look at G92...64 of them and they blew away G80 in fillrate tests but that didnt mean they were faster.

It'll be interesting to see if bottlenecks still exist or if they happen to have just enough to get over the hump

[AliG]

No. ALU and TMU are independant of each other in the respect that you only need as much texture power as games can use, but as I've said before, I can't see ATI further reducing the ratio, if anything next generation expect to see it increase again as they wouldn't want to risk making a alu monster without enough texture power again.


40 TMUs will be enough, the 8800gtx only had 32 and was fine, and besides the 4xaa doesn't kill performance anymore, jimmyz's review proved that, the drop in fps from 2x to 4x with the r600 was more than 50% in some cases, with COJ (which is pretty demanding) the 4850 lost only 2 fps to the drop. That, and jimmyz was able to get 20 FPS out of his system at stock for crysis at 1920*1200 4xaa 16xaf, that was never possible before for ati, I don't think even the g92 gts could do that (I believe the ultra was the only one due to the high clocks and wide bus)

[madmossy]

TMU's greatly depend on the game, if its a shader intensive game they mean very little, if on the other hand its not then they are vital. Need only look at benchmarks of the gtx280 vs GX2 to see how TMU/Shaders are used in different games.

[v_rr]

And if nothing else has changed they will be still insufficient. 800SPs/32TMUs ratio is worse than 320SPs/16TMUs. If 16 were not enough before, 32 will be even worse now because of the huge number of SPs. If the chip has 40TMUs instead, it will be the same as before. ATI better has done something to that TMUs or we'll see the same bottleneck in texture perfomance

That´s not true. You don´t need such texturing power as that.
G80 with 32 TMU did just well and 8800 ultra is still much faster the 9800GTX in high res with AA and AF.

That´s because it have much more bandwith. So I think that HD 4850 is realy heavy bandwith limited.
GDDR5 in HD 4870 should bring very nice performance jump to feed 800 SP´s.

[XSAlliN]

That´s not true. You don´t need such texturing power as that.
G80 with 32 TMU did just well and 8800 ultra is still much faster the 9800GTX in high res with AA and AF.

Yeah, but Ultra has only 128 Stream Processors - 32 TMU and 24 ROP bring a good balance in that configuration.

Compared to that 800SP /32 TMU + 16 ROP might look like disaster... hope the results are different. The high OC-ers might care only about high results in benchmarks, but the gamers care more about stability and balance between frames. What do I care if I can reach 60 FPS in Crysis from time to time, wille my FPS looks and feel like this:

next second 12 frames -> next second 45 frames -> next second 23 frames -> next second 60 frames -> next second 21 frames -> next second 7 frames -> next second 37 frames -> next second 42 frames -> next second 21 frames -> next second 58 frames -> and so on.

[v_rr]

Yeah, but Ultra has only 128 Stream Processors - 32 TMU and 24 ROP bring a good balance in that configuration.

Look to G92:
64TMU´s and 16 ROPs

They just need to balance things good. Let´s wait and see....

[AliG]

Yeah, but Ultra has only 128 Stream Processors - 32 TMU and 24 ROP bring a good balance in that configuration.

Compared to that 800SP /32 TMU + 16 ROP might look like disaster... hope the results are different. The high OC-ers might care only about high results in benchmarks, but the gamers care more about stability and balance between frames. What do I care if I can reach 60 FPS in Crysis from time to time, wille my FPS looks and feel like this:

next second 12 frames -> next second 45 frames -> next second 23 frames -> next second 60 frames -> next second 21 frames -> next second 7 frames -> next second 37 frames -> next second 42 frames -> next second 21 frames -> next second 58 frames -> and so on.

I don't think it will be that big of an issue, honestly g80 showed that as long as you have enough power, whether balanced or not, you'll be fine. G80 just happened to be very balanced, but g92 was fine with only 16 ROPs, it just needed more bandwidth at the higher resolutions but otherwise outperformed g80, and g80 stayed competitive with only 32 ROPs. Even the 3870 was a great competitor without any visual enhancements as it had enough shader power, just not enough texture.

When people say the 16 tmus and rops bottlenecked the r600, they don't mean that they screwed up the performance, just that they didn't allow for the r600 to reach its full potential

[XSAlliN]

Look to G92:
64TMU´s and 16 ROPs

They just need to balance things good. Let´s wait and see....

Yes but they also have 128 SP so 64 TMU showed even more balance , even at 16 ROPS....

Even the 3870 was a great competitor without any visual enhancements as it had enough shader power, just not enough texture.

Don't remind me that's my current card. I never fall for benchmarks, but new technology like GDDR4 made a sucker out of me... that's my weakness when it comes to upgrades.

[AliG]

but see, that's my point, it doesn't necessarily have to be perfectly balanced. I'd like to see a 1:1 alu:tmu, but that's just not going to happen, and for many reasons, the main being it just isn't necessary and would take up too much space and transistors. Look at gt200, that has 240 shaders, but only 80 TMU, that's a 1:3 ratio, not 1:2 or 1:4 (which has proven to be enough as long as there's enough texture power) and 32 ROPs, which is a 1:~7 ratio. You don't need everything to be perfectly balanced, with the games demanding more and more shader power, it just can't happen anymore. But what you can do is provide enough that it does what you need it to. GT200 is a great example of this because it would be fine with half the texture power as the 9800gx2 beats it even in the higher resolution depending on the game

[XSAlliN]

I know the ROP's don't have a major influence but when it comes to SP's - like 800 of them (not 64 with 5 operations per cycle - the 320 SP's of 3870) I do feel you need a balance TMU, let's say 48 TMU with some bandwidth help might be enough to keep a stable FPS ( +/- 2,3 FPS) - I'm talking about optimized games. A game with poor optimizations will get crappy FPS even on the most perfect configuration.

[003]

ati will support havok on there gpu's so maybe its nvidias lack of havok thats the bigger issue?

No. Unless ATI supports the PhysX API (which they can do, but would also need to support CUDA, which they will never do), it is their problem. PhysX is already the standardized physics processing API designed from the ground up to be done in hardware.

It is already used in multiple versions of 3DMark, and all games that use the Unreal Engine 3, plus many more game engines that I don't know off the top of my head.

Combine that with the fact that all GeForce 8 and up video cards will gain support for PhysX, it really is ATI's problem, and a pretty big one at that. Nvidia made a very, very smart move in buying Ageia. Even if the 4870 X2 is faster than the GTX 280, wait for games that heavily rely on PhysX to start coming out, and watch the tables turn.

It is for this reason that I got a used PhysX card cheaply on e-bay.

[AliG]

Most people just turn on vsync if they have weird framerates that are well above 60 and that usually fixes things


Honestly, think whatever you want, no one can take away your opinion, I just think its not necessary

[Hornet331]

No. Unless ATI supports the PhysX API (which they can do, but would also need to support CUDA, which they will never do), it is their problem. PhysX is already the standardized physics processing API designed from the ground up to be done in hardware.

It is already used in multiple versions of 3DMark, and all games that use the Unreal Engine 3, plus many more game engines that I don't know off the top of my head.

Combine that with the fact that all GeForce 8 and up video cards will gain support for PhysX, it really is ATI's problem, and a pretty big one at that. Nvidia made a very, very smart move in buying Ageia. Even if the 4870 X2 is faster than the GTX 280, wait for games that heavily rely on PhysX to start coming out, and watch the tables turn.

It is for this reason that I got a used PhysX card cheaply on e-bay.

acutally havoc was/is the standard, many game engines have physics engiens based on havoc (thaty why intel bought them )

aegia (aka PhysX) main win was the ut3 engien.

[STaRGaZeR]

That´s not true. You don´t need such texturing power as that.
G80 with 32 TMU did just well and 8800 ultra is still much faster the 9800GTX in high res with AA and AF.

That´s because it have much more bandwith. So I think that HD 4850 is realy heavy bandwith limited.
GDDR5 in HD 4870 should bring very nice performance jump to feed 800 SP´s.

Don't mix NV and ATI here, very different architectures, clocks, etc. RV770 is basicly the same of RV670 with more of all its components. If ATI has just increased the number of units without doing anything else this card will be faster of course, but you'll have the same bottleneck in 45FPS instead of 35FPS for example. I smell a limited card again, I've been hearing "future games will need lots of shading power, so let's crack up the number of SPs without looking at texture perfomance". Heck we are in the PRESENT, and PRESENT games are FAR from needing such enormous shading power. They need a balanced design, just like NV's. By the time all this cards' shading power is needed we will have R1200 or something. Damn I just don't get why they don't learn from their mistakes.

[v_rr]

No. Unless ATI supports the PhysX API (which they can do, but would also need to support CUDA, which they will never do), it is their problem. PhysX is already the standardized physics processing API designed from the ground up to be done in hardware.

It is already used in multiple versions of 3DMark, and all games that use the Unreal Engine 3, plus many more game engines that I don't know off the top of my head.

Combine that with the fact that all GeForce 8 and up video cards will gain support for PhysX, it really is ATI's problem, and a pretty big one at that. Nvidia made a very, very smart move in buying Ageia. Even if the 4870 X2 is faster than the GTX 280, wait for games that heavily rely on PhysX to start coming out, and watch the tables turn.

It is for this reason that I got a used PhysX card cheaply on e-bay.

Havock as far more Games then PhysX.
PhysX was a failure since the begining. They have UT3 and some others, but Havock have far way much more games.

And physX in UT3 adds almost nothing or nothing to normal gameplay.
And remember that if the card is rendering PhysX, it is not rendering pixel and vertex = slower FPS.

[Papu]

No. Unless ATI supports the PhysX API (which they can do, but would also need to support CUDA, which they will never do), it is their problem. PhysX is already the standardized physics processing API designed from the ground up to be done in hardware.

It is already used in multiple versions of 3DMark, and all games that use the Unreal Engine 3, plus many more game engines that I don't know off the top of my head.

Combine that with the fact that all GeForce 8 and up video cards will gain support for PhysX, it really is ATI's problem, and a pretty big one at that. Nvidia made a very, very smart move in buying Ageia. Even if the 4870 X2 is faster than the GTX 280, wait for games that heavily rely on PhysX to start coming out, and watch the tables turn.

It is for this reason that I got a used PhysX card cheaply on e-bay.

what great games support/make proper use of Physx? that crappy demo physix put out to show how great it was , i was able to run it fine on my 2900xt.... havok is used in many great titles including the half life 2 series ,oblivion ect ect. Its much more likely that havok an already established api is more important than crappy Physx..

[Xello]

Havock as far more Games then PhysX.
PhysX was a failure since the begining. They have UT3 and some others, but Havock have far way much more games.

what great games support/make proper use of Physx? that crappy demo physix put out to show how great it was , i was able to run it fine on my 2900xt.... havok is used in many great titles including the half life 2 series ,oblivion ect ect. Its much more likely that havok an already established api is more important than crappy Physx..

While I won't take the flame bate you so blatantly cast (you could read what you wrote, swapping 'physx' for 'havok' and it would be equally irrelevant) I would say that NV's swaying power with developers is not to be sniffed at, as they've proved on multiple occasions.

Whether they're fair in all their dealings is another discussion

Arguing which API is currently 'the best' is pretty pointless:

http://www.havok.com/content/blogcategory/29/73/

http://www.nzone.com/object/nzone_physxgames_home.html

Both have an impressive library so far, though i think it's the games to come (not the games already out) which will be more important here.

Also worth noting may be that Physx has somewhat of a head start in this area, as so far Havok is only supported on the CPU whereas Physx has had its own PPU, and soon to be GPU acceleration.

[Hornet331]

what great games support/make proper use of Physx? that crappy demo physix put out to show how great it was , i was able to run it fine on my 2900xt.... havok is used in many great titles including the half life 2 series ,oblivion ect ect. Its much more likely that havok an already established api is more important than crappy Physx..

but you have to differentiate, many games use havoc, but most of this havoc physic engiens are heavly modified (e.g. HL2) and only have smal parts that are common (like ragdolls).

[Papu]

While I won't take the flame bate you so blatantly cast (you could read what you wrote, swapping 'physx' for 'havok' and it would be equally irrelevant) I would say that NV's swaying power with developers is not to be sniffed at, as they've proved on multiple occasions.

Whether they're fair in all their dealings is another discussion

Arguing which API is currently 'the best' is pretty pointless:

http://www.havok.com/content/blogcategory/29/73/

http://physx.cwx.ru/

Both have an impressive library so far.

flame bate? lol...
it also seems that once you switch on Physx on the nv gpu that it cuts framrates , so they will be advising 2 or 3 gpu solutions so it wont really be put to much use by normal pc users/gamers

[zerazax]

Havok is used in far more games than PhysX

And for all of you saying its a problem... its still up to the developers on how much they implement in. And you know what? Seeing as to how much PC game developers are worried about sales figures, they'll never accept a configuration that alienates 25% of their possible market

[Xello]

seems that once you switch on Physx on the nv gpu that it cuts framrates , so they will be advising 2 or 3 gpu solutions so it wont really be put to much use by normal pc users/gamers

AMD on the other hand have found a magic box that allows acceleration without the performance hit? (any sources?)

Havok is used in far more games than PhysX

For PC games i count 44 for Havok and 74 for Physx (from the 'partial list' on that site), I was quick and may be 1 or 2 off, but i'm way too lazy to do it again