By using this site, you agree to our Privacy Policy and our Terms of Use. Close

Forums - Nintendo - Wii U GPU new info and more speculation

timmah said:
dahuman said:
errorpwns said:
 


It loses marginally to the 360?

 

http://www.youtube.com/watch?feature=player_embedded&v=0eH3DmUgokk

 

Really now? 

here let me embed for you:

NFS Most Wanted running MAX settings on HD5500, does this look similar to the above video (minus the annoying blur effect)? Why yes, it does.


No way it's a 160SP 6450, which can't even run the game at low settings:

Don't know what's worse, the framerate or the guy's driving skills.

Edit: wait didn't you had another video before? This one: http://www.youtube.com/watch?v=10Ubc4Y7f40



Nintendo and PC gamer

Around the Network
osed125 said:

Don't know what's worse, the framerate or the guy's driving skills.

Edit: wait didn't you had another video before? This one: http://www.youtube.com/watch?v=10Ubc4Y7f40

Yes, then I realized it was for the 5450, not the 6450 (though both perform equally bad).



My take: We have a GPU roughly on par with the HD5550 (maybe better due to customization) or a downscaled HD6570 with customization (or even something in the same ballpark from their embedded line, such as a modified E6760). The initial ports had trouble due to poor optimization, resulting in performance issues either from CPU code issues or poor memory management (not using texture compression, not properly utilizing EDRAM, etc.), by no fault of the devs because they haven't had time to learn the system. Results porting from the PS4/Nextbox should be similar to scaling the graphics back on PC games from a mid-range gaming rig with a 6670, 7750, or 7770 type card to a more modest 'budget' rig with a 5550-6570 type card. Ports can run well enough and look fine (if optimized correctly), but will be scaled back graphically by some margin and most likely run at 720P. Still nowhere near the Wii-PS360 gap by any stretch.



timmah said:

My take: We have a GPU roughly on par with the HD5550 (maybe better due to customization) or a downscaled HD6570 with customization (or even something in the same ballpark from their embedded line, such as a modified E6760). The initial ports had trouble due to poor optimization, resulting in performance issues either from CPU code issues or poor memory management (not using texture compression, not properly utilizing EDRAM, etc.), by no fault of the devs because they haven't had time to learn the system. Results porting from the PS4/Nextbox should be similar to scaling the graphics back on PC games from a mid-range gaming rig with a 6670, 7750, or 7770 type card to a more modest 'budget' rig with a 5550-6570 type card. Ports can run well enough and look fine (if optimized correctly), but will be scaled back graphically by some margin and most likely run at 720P. Still nowhere near the Wii-PS360 gap by any stretch.


6570 is 480:24:8 part, its DDR3 version with same memory bandwith and same clock as 5550 performs some 40-45% better than 5550. If that's what inside, it shouldn't have had problem running NFS @30pfs/720p from the get go. But, that part is 118mm^2 at 40nm, so I don't think it fits...

My pick is still Redwood LE (5550) or RV730 (4650) shrank to 40nm - 160SP config looks just...not sure what words to use if Nintendo went with that, but who knows. Anyway, even if it's 320:16:8 config inside, we're looking at some 8x between PS4 and WiiU...really shame they haven't gone for at least something like that rumoured E6760 with proper memory bandwith.



HoloDust said:
timmah said:

My take: We have a GPU roughly on par with the HD5550 (maybe better due to customization) or a downscaled HD6570 with customization (or even something in the same ballpark from their embedded line, such as a modified E6760). The initial ports had trouble due to poor optimization, resulting in performance issues either from CPU code issues or poor memory management (not using texture compression, not properly utilizing EDRAM, etc.), by no fault of the devs because they haven't had time to learn the system. Results porting from the PS4/Nextbox should be similar to scaling the graphics back on PC games from a mid-range gaming rig with a 6670, 7750, or 7770 type card to a more modest 'budget' rig with a 5550-6570 type card. Ports can run well enough and look fine (if optimized correctly), but will be scaled back graphically by some margin and most likely run at 720P. Still nowhere near the Wii-PS360 gap by any stretch.


6570 is 480:24:8 part, its DDR3 version with same memory bandwith and same clock as 5550 performs some 40-45% better than 5550. If that's what inside, it shouldn't have had problem running NFS @30pfs/720p from the get go. But, that part is 118mm^2 at 40nm, so I don't think it fits...

My pick is still Redwood LE (5550) or RV730 (4650) shrank to 40nm - 160SP config looks just...not sure what words to use if Nintendo went with that, but who knows. Anyway, even if it's 320:16:8 config inside, we're looking at some 8x between PS4 and WiiU...really shame they haven't gone for at least something like that rumoured E6760 with proper memory bandwith.

It could still theoretically be a custom chip based on the architecture of the E6760, it is certainly very, very customized, so we really don't know what part they used as the basis for the chip. My assumption from seeing performance in the NFS-MW video compared to other cards is, it CANNOT be a 160SP 6450, not a chance, and performance seems to be AT LEAST in the neighborhood of the stock 5550 up to 6570 range based on how it's running that game. As far as memory bandwidth, we've heard developers complain about the CPU (though that should be ok when properly utilized due to a short pipeline and OOOE), but there's been nothing but praise for the memory architecture so far, so maybe that's not a big deal. Not a single developer has said anything about Memory bandwidth issues. Every system Nintendo has built since GC has been very efficient and well balanced, so I assume they would continue that tradition.

Keep in mind, we have a part number for the memory, but don't know to what degree it's been customized, if at all. We don't know for certain what the memory clock rate is, and nobody has actually presented a benchmark to test real-world performance. How does the memory controller use EDRAM? Can the system use some type of predictive prefetch to EDRAM to accelerate real-world memory performance? We have a die shot of the GPU, but no solid answer as to what part it's based on. We also don't know wat the 30% of the GPU that is 'unknown' does - we are fairly certain it has a hardware tessellation unit, but what else did Nintendo bake into the chip? Did Nintendo & AMD create some crazy DX11 type fixed functions to give developers 'free' lighting or other 'free' effects? Maybe some kind of proprietary hardware texture compression is available to negate a potential RAM bottleneck? Some of that is a bit far fetched, but this is Nintendo we're talking about. We have no idea the details of the 'unknowns', and I would assume that 30% does something... I wish Nintendo would just release the damn specs!

EDIT: No way that wold give PS4 an 8x advantage. Last I heard, the rumor was about 1.3TFLOPS GPU (and I'm betting it's going to be downclocked for heat & power reasons), which is about a 3-4x difference based on raw power if you assume the 5550. Wii-PS360 was around a 20x power differential.



Around the Network
timmah said:
HoloDust said:


6570 is 480:24:8 part, its DDR3 version with same memory bandwith and same clock as 5550 performs some 40-45% better than 5550. If that's what inside, it shouldn't have had problem running NFS @30pfs/720p from the get go. But, that part is 118mm^2 at 40nm, so I don't think it fits...

My pick is still Redwood LE (5550) or RV730 (4650) shrank to 40nm - 160SP config looks just...not sure what words to use if Nintendo went with that, but who knows. Anyway, even if it's 320:16:8 config inside, we're looking at some 8x between PS4 and WiiU...really shame they haven't gone for at least something like that rumoured E6760 with proper memory bandwith.

It could still theoretically be a custom chip based on the architecture of the E6760, it is certainly very, very customized, so we really don't know what part they used as the basis for the chip. My assumption from seeing performance in the NFS-MW video compared to other cards is, it CANNOT be a 160SP 6450, not a chance, and performance seems to be AT LEAST in the neighborhood of the stock 5550 up to 6570 range based on how it's running that game. As far as memory bandwidth, we've heard developers complain about the CPU (though that should be ok when properly utilized due to a short pipeline and OOOE), but there's been nothing but praise for the memory architecture so far, so maybe that's not a big deal. Not a single developer has said anything about Memory bandwidth issues. Every system Nintendo has built since GC has been very efficient and well balanced, so I assume they would continue that tradition.

Keep in mind, we have a part number for the memory, but don't know to what degree it's been customized, if at all. We don't know for certain what the memory clock rate is, and nobody has actually presented a benchmark to test real-world performance. How does the memory controller use EDRAM? Can the system use some type of predictive prefetch to EDRAM to accelerate real-world memory performance? We have a die shot of the GPU, but no solid answer as to what part it's based on. We also don't know wat the 30% of the GPU that is 'unknown' does - we are fairly certain it has a hardware tessellation unit, but what else did Nintendo bake into the chip? Did Nintendo & AMD create some crazy DX11 type fixed functions to give developers 'free' lighting or other 'free' effects? Maybe some kind of proprietary hardware texture compression is available to negate a potential RAM bottleneck? Some of that is a bit far fetched, but this is Nintendo we're talking about. We have no idea the details of the 'unknowns', and I would assume that 30% does something... I wish Nintendo would just release the damn specs!


As you  might remember, 5550 comes in 3 flavours (all with same core and mem clock, over 128bit bus) - GDDR5 (27VP rating), DDR3 (21.8VP) and DDR2 (16.5VP). As you may see, memory bandwith is bottleneck, and is sole purpose for difference in perfomance between 3 versions.

5550 DDR3 is more than enough to have NFS sort of visuals at 720p. Problem is, WiiU's suggested bandwith (DDR3/64 bit) matches the lowest version (DDR2/128bit), 12.8GB/s in both cases. Where I supose the problem lies is overcoming that bottleneck and pushing that (hypotetical) 5550 to behave more like it's DDR3 over 128bit bus.

But, what puzzels me most is, as you said, that nobody complained about memory subsystem, on the contrary, I think I remember some devs saying it's well balanced. Now, that 6450 is 64bit card, and those 64bits are actually well balanced for it...

I'm still thinking it's some 320SP config inside, IMHO 6570 would be overkill for such bottlenecked system, but, as I said, I really wish they actually went at least with it and 128bit bus (and of course, better CPU, but that comes from my WiiHD vs WiiU standpoint).



HoloDust said:
timmah said:

It could still theoretically be a custom chip based on the architecture of the E6760, it is certainly very, very customized, so we really don't know what part they used as the basis for the chip. My assumption from seeing performance in the NFS-MW video compared to other cards is, it CANNOT be a 160SP 6450, not a chance, and performance seems to be AT LEAST in the neighborhood of the stock 5550 up to 6570 range based on how it's running that game. As far as memory bandwidth, we've heard developers complain about the CPU (though that should be ok when properly utilized due to a short pipeline and OOOE), but there's been nothing but praise for the memory architecture so far, so maybe that's not a big deal. Not a single developer has said anything about Memory bandwidth issues. Every system Nintendo has built since GC has been very efficient and well balanced, so I assume they would continue that tradition.

Keep in mind, we have a part number for the memory, but don't know to what degree it's been customized, if at all. We don't know for certain what the memory clock rate is, and nobody has actually presented a benchmark to test real-world performance. How does the memory controller use EDRAM? Can the system use some type of predictive prefetch to EDRAM to accelerate real-world memory performance? We have a die shot of the GPU, but no solid answer as to what part it's based on. We also don't know wat the 30% of the GPU that is 'unknown' does - we are fairly certain it has a hardware tessellation unit, but what else did Nintendo bake into the chip? Did Nintendo & AMD create some crazy DX11 type fixed functions to give developers 'free' lighting or other 'free' effects? Maybe some kind of proprietary hardware texture compression is available to negate a potential RAM bottleneck? Some of that is a bit far fetched, but this is Nintendo we're talking about. We have no idea the details of the 'unknowns', and I would assume that 30% does something... I wish Nintendo would just release the damn specs!


As you  might remember, 5550 comes in 3 flavours (all with same core and mem clock, over 128bit bus) - GDDR5 (27VP rating), DDR3 (21.8VP) and DDR2 (16.5VP). As you may see, memory bandwith is bottleneck, and is sole purpose for difference in perfomance between 3 versions.

5550 DDR3 is more than enough to have NFS sort of visuals at 720p. Problem is, WiiU's suggested bandwith (DDR3/64 bit) matches the lowest version (DDR2/128bit), 12.8GB/s in both cases. Where I supose the problem lies is overcoming that bottleneck and pushing that (hypotetical) 5550 to behave more like it's DDR3 over 128bit bus.

But, what puzzels me most is, as you said, that nobody complained about memory subsystem, on the contrary, I think I remember some devs saying it's well balanced. Now, that 6450 is 64bit card, and those 64bits are actually well balanced for it...

I'm still thinking it's some 320SP config inside, IMHO 6570 would be overkill for such bottlenecked system, but, as I said, I really wish they actually went at least with it and 128bit bus (and of course, better CPU, but that comes from my WiiHD vs WiiU standpoint).

...But the 6450 can't do what we saw in that video.

Keep in mind, the developers do not indicate that there is any RAM bottleneck, quite the contrary in fact. This leads to one or more of the following possibilities: (a.) the bandwidth assumptions are wrong, (b.) the memory controller alleviates the 'bottleneck' by utilizing EDRAM and maybe some sort of predictive prefetch, (c.) there is some *really* good hardware texture compression that effectively eliminates any memory bottleneck. One or a combination of these possibilities would remove the 'bottlenecked system' argument... if games are coded properly. Early ports have never been an indication of a system's power, so I reserve judgement for at least a year. Speculation says RAM bottleneck, developers say the opposite, I believe the devs.



timmah said:

EDIT: No way that wold give PS4 an 8x advantage. Last I heard, the rumor was about 1.3TFLOPS GPU (and I'm betting it's going to be downclocked for heat & power reasons), which is about a 3-4x difference based on raw power if you assume the 5550. Wii-PS360 was around a 20x power differential.

PS4 is rumoured at some 1.84TFLOPs (1152shaders x 2ops x 800MHz)

But you can't compare GPUs by flops  alone - suggested PS4 GPU is rated at around 140-145 VP (similar to 7850). Depending on what's inside WiiU, you have as low as 16.8 to max of some 21.8 (for 5550 and DDR3 64&128bit). So that's between 6.5 and 8+ for 5550 level GPU.



HoloDust said:
timmah said:

EDIT: No way that wold give PS4 an 8x advantage. Last I heard, the rumor was about 1.3TFLOPS GPU (and I'm betting it's going to be downclocked for heat & power reasons), which is about a 3-4x difference based on raw power if you assume the 5550. Wii-PS360 was around a 20x power differential.

PS4 is rumoured at some 1.84TFLOPs (1152shaders x 2ops x 800MHz)

But you can't compare GPUs by flops  alone - suggested PS4 GPU is rated at around 140-145 VP (similar to 7850). Depending on what's inside WiiU, you have as low as 16.8 to max of some 21.8 (for 5550 and DDR3 64&128bit). So that's between 6.5 and 8+ for 5550 level GPU.

The U was rumored at ~800GFLOPS originally, it'll be interesting to see if the PS4 rumor is accurate. I still have a feeling both systems' (PS4/Nextbox) GPUs will be downclocked to some degree vs. their desktop counterparts to reduce heat & power usage.



Does anyone think and AMD HSA technology might be used in the GPU?