More info about RSX from NVIDIA

It should. We'll see. Atleast we now have confirmation of significant ROP and TMU unit changes in the G70/RSX. The G70 is simply not just an NV4x with 8 more pipes.
 
DemoCoder said:
It should. We'll see. Atleast we now have confirmation of significant ROP and TMU unit changes in the G70/RSX. The G70 is simply not just an NV4x with 8 more pipes.
For the record, can someone explain the changes.
 
There's alot of information on the internet already from people who jumped the gun on the NDA. Official word should come around tomorrow morning.
 
According to this page: http://review.gzeasy.com/graphics/2005-06-21/graphics_2225_3.html
The chart on page 1 of this thread is incorrect. The pixel shader ALU's are 4D and texturing is still coupled to the first/top ALU. They are essentially unchanged from NV40 with the exception of an add stage inserted in the first/top ALU making it MAD/DOT capable. Apparently the vertex shaders are still limited to point sampling. The ROPS have been tweaked to better support alpha to mask, render to texture, blending, etc. Obvious additions are two more vertex units and eight more pixel units. Other tweaks to the architecture have been made to help improve clock speeds as they seem able to hit 500Mhz+. A nice extention of the NV40 architecture but nothing substantially different.
 
The numbers seem to work out...

2 Vec4 + scalar + normalize (assuming Vec3)
2 * 8 + 2 + 9 = 27 flops per clock

648 / 24 pixel pipelines = 27 flops per clock

If the texture units are still coupled then the Xenos and RSX look fairly similar...
- RSX: Vec4 & Scalar + normalize (24 bi-linear textures per clock)
- Xenos: Vec4 & Scalar + unknown 33% claim (16 bi-linear textures per clock)
 
Rockster said:
According to this page: http://review.gzeasy.com/graphics/2005-06-21/graphics_2225_3.html
The chart on page 1 of this thread is incorrect. The pixel shader ALU's are 4D and texturing is still coupled to the first/top ALU. They are essentially unchanged from NV40 with the exception of an add stage inserted in the first/top ALU making it MAD/DOT capable. Apparently the vertex shaders are still limited to point sampling. The ROPS have been tweaked to better support alpha to mask, render to texture, blending, etc. Obvious additions are two more vertex units and eight more pixel units. Other tweaks to the architecture have been made to help improve clock speeds as they seem able to hit 500Mhz+. A nice extention of the NV40 architecture but nothing substantially different.
Link doesn't work.
 
kaigai03l.gif
 
Just going by those game benchmarks it seems like the G70 is just your average % 30-40 perfomance increase from the last gen.

Every gen we hear that the new card will be twice as fast but it never happens and we end up 30-40% faster.

Of course Im hoping they have a mad ass AA for free and HDR with not much frame rate hit, now that would be something, but as it stands those Far cry and Half Life 2 benchmarks are very dissapointing after the wait.

Dont mention to me "its cpu limited". Its a 4000+ Athlon, how much of a funting faster cpu will we need?
 
Agisthos said:
Dont mention to me "its cpu limited". Its a 4000+ Athlon, how much of a funting faster cpu will we need?

You need Cell :devilish:

I don't think you can write off the impact of the CPU here, I think it's fairly obvious the 7800 is CPU-bound in some of these benches. You can see the difference starts increasing as you try and put the load back on the GPU (increasing res and AA/AF), but I think the CPU is still the limiting factor in a lot of those cases.
 
Titanio said:
Agisthos said:
Dont mention to me "its cpu limited". Its a 4000+ Athlon, how much of a funting faster cpu will we need?

You need Cell :devilish:

I don't think you can write off the impact of the CPU here, I think it's fairly obvious the 7800 is CPU-bound in some of these benches. You can see the difference starts increasing as you try and put the load back on the GPU (increasing res and AA/AF), but I think the CPU is still the limiting factor in a lot of those cases.

True but I have seen benchmarks where it was also shown the 6800 Ultra was cpu bound too.

From the perspective of a PC card it is a straight apples to apples comparison for me, cos current cpu's is what we are using. PS3 is a different matter.....
 
Agisthos said:
Titanio said:
Agisthos said:
Dont mention to me "its cpu limited". Its a 4000+ Athlon, how much of a funting faster cpu will we need?

You need Cell :devilish:

I don't think you can write off the impact of the CPU here, I think it's fairly obvious the 7800 is CPU-bound in some of these benches. You can see the difference starts increasing as you try and put the load back on the GPU (increasing res and AA/AF), but I think the CPU is still the limiting factor in a lot of those cases.

True but I have seen benchmarks where it was also shown the 6800 Ultra was cpu bound too.

From the perspective of a PC card it is a straight apples to apples comparison for me, cos current cpu's is what we are using.

Of course, unless you're planning CPU upgrades, you perhaps may as well match your GPU to your CPU in a way that might save you money. Some games coming down the pipe might push it harder regardless of the CPU, however.
 
One of the things about the G70 is that it could even have some specific design features for dual core cpu's. Nvidia claims the next driver release will give %10-%30 increase with dual core, and thats just with current cards.
It would be interesting to see a purpose built graphics card for mutiple core cpu's being able to offload work onto a seperate core, just like they are talking about the Cell being able to do
 
So TMU is still partially coupled to the first ALU, and they're summing fp16 math ops count with fp32 bits math ops count :devilish:
kaigai01l.gif
 
Umh..from that char I'd say G70 pixel processors can do 20 (programmable) 32bit flops per cycle, not 27. This would put programmable floating point power on pixel shading on par with Xenos.
 
The RSX pixel pipelines seem to match R500's ALUs even when the R500 dedicates all units to pixel shading in addition to the claimed 33% bonus. Of course, they were summed up FP16 and FP32 operations but I'm sure you could use those FP16 operations for shading that doesn't require much precision.

Meh, I guess we have to wait for both consoles to be released to see which console is more powerful.
 
Back
Top