-
Competitor rules
Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.
Radeon HD 5870 QuadFire User Video
- Thread starter Final8y
- Start date
Caporegime
- Joined
- 8 Jul 2003
- Posts
- 30,063
- Location
- In a house
Yeah noticed that, was slower than 3x cards, still, an awesome setup that, but had to laugh at the Crysis and Warhead scores
Yeah noticed that, was slower than 3x cards, still, an awesome setup that, but had to laugh at the Crysis and Warhead scores
TBFH, i gave up on that dog crysis years a go, its a joke.
Caporegime
- Joined
- 8 Jul 2003
- Posts
- 30,063
- Location
- In a house
Yeah, i still don't know why everyones bothering with it, has to be a game problem, it just can't be because cards are not powerful enough, no way, not now
Exactly. If it can't perform well on cards several generations beyond what was available at the time then you have to accept that it was either a) poorly coded, or b) too demanding for the market. It doesn't take a genius to figure out which it was. Personally I thought Far Cry 2 looked better and it ran amazingly well.
Soldato
- Joined
- 9 Jan 2008
- Posts
- 6,521
- Location
- London/Camberley
I have been following his thread on guru3d and he has one insane setup!
Soldato
- Joined
- 26 Apr 2004
- Posts
- 9,851
- Location
- Milton Keynes
Absolutely insane stuff; that'd be the sort of machine I'd build; completely water cooled if I was earning in the region of 100K a year.
As it stands though I just couldnt justify spending £1.2k on the cards alone, plus £1.2 on the screens, without the rest, thats probably a 3.5-4K machine, given some current experience with them, I'd probably get Kobalt to do it, because they look damn good.
Edit: Hmmm...not sure who edited my sig, it was within spec....
As it stands though I just couldnt justify spending £1.2k on the cards alone, plus £1.2 on the screens, without the rest, thats probably a 3.5-4K machine, given some current experience with them, I'd probably get Kobalt to do it, because they look damn good.
Edit: Hmmm...not sure who edited my sig, it was within spec....
There's a question about where does a large portion of his system memory disappear to. That's easy to answer: the system has to place what it wants to upload into a window area so that the graphics card can then DMA it over PCI-E into the GPU memory.
I think the current way graphics engines are too basic. The CPU has to prepare the data into a shadow buffer (CPU/memory bandwidth) then add that GPU-DMA is another chunk of memory bandwidth. So multiply that lot by 4.. the older, simpler game engines aren't written for optimising that sort of processing hence strangle themselves.
I think the current way graphics engines are too basic. The CPU has to prepare the data into a shadow buffer (CPU/memory bandwidth) then add that GPU-DMA is another chunk of memory bandwidth. So multiply that lot by 4.. the older, simpler game engines aren't written for optimising that sort of processing hence strangle themselves.
I thought that was the case only for DX9 games?
DX11 seems to only provide information about Unordered Access Views (UAVs). Now the article here seems to indicate that this is GPU reads.. however GPGPU terminology for using memory pointers for parameter inputs is "gather" these were originally limited to 1, the CTM increased these and DX has increased too.
The old ATI hardware used to 'read-through' into the shadowed memory then cache the data. Hence the slow texture load on the first use if the system didn't have enough GPU memory to pre-cache the textures (which equate to arrays in.
Shared memory - is the shared memory between fragment processors (stream processors) on the GPU.
The old bottle neck of not being able to simultaneously process data on the card and transfer data to/from main memory may still exist if the 5800 command/memory controller can only service one memory access at a time (GPU-GPU memory). This is still unclear - I'll dig further..
Fermi I did note can simultaneously process two memory requests - a major boon for compute type processing.
According to this: http://developer.amd.com/gpu_assets/Shader Model 5-0 and Compute Shader.pps
The UAV is part of gather/scatter so not about GPU-System memory transfers.
Looking through the ATI/AMD 5800 architecture and review slides, nobody has mentioned anything on this. Why am I asking this?
In processing where there's not enough memory to pre-cache al graphics textures and 3D points for all the physics etc, then the system will have to mode data in/out. Without being able todo system-gpu and gpu-gpu transfers simultaneously this results is far far slower performance. In theory, the idea GPU should be able to DMA in & out to system memory why the stream processors are processing other independent data from memory..
CPU -> GPU physics data transfer
GPU -> GPU processing of physics data
GPU -> CPU physics data transfer
CPU -> GPU graphics data (textures/vertexes)
GPU -> GPU graphics rendering
repeat from first...
So because of this the GPU needs loads of memory to keep both the hi-res graphics AND the physics in the GPU memory:
CPU -> GPU physics data transfer/fragment shader programs
CPU -> GPU graphics data (textures/vertexes, shaders)
...
GPU -> GPU processing of physics data
GPU -> GPU process graphics data with physics data
GPU -> GPU graphics rendering
.. <at some point the game code has to process the entire scene to update with game AI move/user moves etc!>
So you end up doing everything in the GPU but you require a massive amount of GPU memory to effectively work.
In future I can see the move to push the entire game into the GPU, thus all the parallel processing (even full raytraced rendering) would be run within the GPU. CPU would result to coordination.
The problem is.. AMD have failed to spot this logical bottleneck.. nV's Fermi does provide 2 DMAs simultaneously (according to the slide set). This makes nV's performance for coupled physics/graphics {potentially} higher.
DX11's API allows the programmer to specify processing groups - however the AMD hardware cannot work efficiently with this at the moment.. perhaps the 5900 series
The UAV is part of gather/scatter so not about GPU-System memory transfers.
Looking through the ATI/AMD 5800 architecture and review slides, nobody has mentioned anything on this. Why am I asking this?
In processing where there's not enough memory to pre-cache al graphics textures and 3D points for all the physics etc, then the system will have to mode data in/out. Without being able todo system-gpu and gpu-gpu transfers simultaneously this results is far far slower performance. In theory, the idea GPU should be able to DMA in & out to system memory why the stream processors are processing other independent data from memory..
CPU -> GPU physics data transfer
GPU -> GPU processing of physics data
GPU -> CPU physics data transfer
CPU -> GPU graphics data (textures/vertexes)
GPU -> GPU graphics rendering
repeat from first...
So because of this the GPU needs loads of memory to keep both the hi-res graphics AND the physics in the GPU memory:
CPU -> GPU physics data transfer/fragment shader programs
CPU -> GPU graphics data (textures/vertexes, shaders)
...
GPU -> GPU processing of physics data
GPU -> GPU process graphics data with physics data
GPU -> GPU graphics rendering
.. <at some point the game code has to process the entire scene to update with game AI move/user moves etc!>
So you end up doing everything in the GPU but you require a massive amount of GPU memory to effectively work.
In future I can see the move to push the entire game into the GPU, thus all the parallel processing (even full raytraced rendering) would be run within the GPU. CPU would result to coordination.
The problem is.. AMD have failed to spot this logical bottleneck.. nV's Fermi does provide 2 DMAs simultaneously (according to the slide set). This makes nV's performance for coupled physics/graphics {potentially} higher.
DX11's API allows the programmer to specify processing groups - however the AMD hardware cannot work efficiently with this at the moment.. perhaps the 5900 series
Last edited:
You are correct.. Actually running an i7 975 at the same speed on "Heaven", "Resdient evil 5" and "3dmark06" You can see I used the xeon 3580 on my 3dmark vantage score.. As well as everything else. Cheers for checking it out guys!
Thanks so much! Love the avatars here! Just awesome!
I did have a question about something. Why do you guys think crysis runs so smooth with 2 5870's while playing, but with 3 it feels laggy.. But when I switch on v-sync, it's smooth again.. Trying to fix it right now..Taking a break from benching so I thought I'd play for a bit. I know the framerates are higher with 3 cards, but it feels really slow without v-sync on..
Monitor maybe? hmmm.
I did have a question about something. Why do you guys think crysis runs so smooth with 2 5870's while playing, but with 3 it feels laggy.. But when I switch on v-sync, it's smooth again.. Trying to fix it right now..Taking a break from benching so I thought I'd play for a bit. I know the framerates are higher with 3 cards, but it feels really slow without v-sync on..
Monitor maybe? hmmm.