Ati X850xt 43 billion shader ops against Xenos 48 billion ?

[Love_In_Rio]

How could this small difference be explained if it is said the XGPU will be more or less twice as powerfull as X850xt ?
It is suppossed that dev alpha kits have a X850xt card, right ? then, if its shader power is so near from a Xenos, why they claim that dev alpha kits are 30 % of the final hardware ?

 

[marconelly!]

Well for one, dev kits have only 256MB of memory from what I've heard. Final hardware - 512. That's a big difference right there without even going into CPU and GPU.

 

[jvd]

Re: Ati X850xt 43 billion shader ops against Xenos 48 billio

How could this small difference be explained if it is said the XGPU will be more or less twice as powerfull as X850xt ?
It is suppossed that dev alpha kits have a X850xt card, right ? then, if its shader power is so near from a Xenos, why they claim that dev alpha kits are 30 % of the final hardware ?

Just because the x850xt is capable of 43 billion shader opps (which is useless for comparing ) doesn't mean it can reach that amount . The xenos has alot of new tech that they are using to reach optimal rendering power , it has a unified shader set up and edram .

 

[nAo]

shader ops kill another one..

 

[Megadrive1988]

How could this small difference be explained if it is said the XGPU will be more or less twice as powerfull as X850xt ?
It is suppossed that dev alpha kits have a X850xt card, right ? then, if its shader power is so near from a Xenos, why they claim that dev alpha kits are 30 % of the final hardware ?

ok to try to answer only the question of why Xenos is not twice as powerful as X850 XT


Xenos does not have straight up twice as much processing resourcesl as X850 XT (R480) or even twice as much as X800 (R420)


the R480/X850 and R420/X800 have 22 pipelines. 16 pixel plus 6 vertex pipes

the Xenos is said to be *equivalent* to 32 pipelines, total.

(remember the 48 unified shader ALU pipes are not comparable head-to-head with 22 pipes of current ATI VPUs)

look at transistor count of the GPU cores: 160 million for R420/X800 and R480X850 compared to 232 million for Xenos

(we're counting rendering/shader core, not the eDRAM of Xenos)

that's the most I can think of right now, focusing on just one thing:
X850 vs Xenos.

 

[nAo]

Short answer: you can't compare different GPUs using shader ops as comparison metric, cause shader ops ARE NOT a well defined/designed concept.
Shader ops a meaningless marketing term!

 

[dukmahsik]

How could this small difference be explained if it is said the XGPU will be more or less twice as powerfull as X850xt ?
It is suppossed that dev alpha kits have a X850xt card, right ? then, if its shader power is so near from a Xenos, why they claim that dev alpha kits are 30 % of the final hardware ?

ok to try to answer only the question of why Xenos is not twice as powerful as X850 XT


Xenos does not have straight up twice as much processing resourcesl as X850 XT (R480) or even twice as much as X800 (R420)


the R480/X850 and R420/X800 have 22 pipelines. 16 pixel plus 6 vertex pipes

the Xenos is said to be *equivalent* to 32 pipelines, total.

(remember the 48 unified shader ALU pipes are not comparable head-to-head with 22 pipes of current ATI VPUs)

look at transistor count of the GPU cores: 160 million for R420/X800 and R480X850 compared to 232 million for Xenos

(we're counting rendering/shader core, not the eDRAM of Xenos)

that's the most I can think of right now, focusing on just one thing:
X850 vs Xenos.

xenos has 332 million trannies i believe

 

[Sonic]

The Xenos is a more efficient graphics architecture. It has been designed for a specific purpose, which is the Xbox 360. Regardless of the amount of shader ops it is capable of compared to R480/420 series it really doesn't matter. The architecture as a whole is extremely flexible and able to do things using more advance shaders that are beyond what the R420/480 are capable of.

 

[Carl B]

xenos has 332 million trannies i believe

That number includes the edram however. You can't include that if all you are looking to do is ballpark the pipeline figures.

 

[Alpha_Spartan]

How do u measure shader ops? Is it consistent across designs? I think not.

 

[The GameMaster]

Let's see...

Each shader pipeline is capable of 2 Shader Operations per cycle (4 ALU operations per shader operation)... *Please correct me if I am wrong here*

The nVidia Geforce 6800/ATI Radeon x850 video cards have 16 pixel pipelines and 6 vertex pipelines for a total of 22 pipelines. 22 pipelines multiplied by 2 would make 44 shader operations per cycle for both the Geforce 6800/Radeon x850. The Geforce 6800 Ultra is clocked at 430MHz and the ATI Radeon x850xt is clocked at 520MHz so that should come out to... 18.9 Billion Shader Operations per second (75.6 Billion ALU operations/second) for the Geforce 6800 and 22.8 Billion Shader Operations per second (91.2 Billion ALU operations per/second) for the Radeon x850xt. Where are they getting the ~44 Billion ShaderOps/second from? Unless they are counting just ALU operations per second in the shaders as opposed to shader operations per second which takes 4 ALU operations per shader operation.

In XENOS the design is 48 unified pipelines with 64 simultaneous threads, and again each pipeline is capable of 2 shader operations per pipeline (4 ALU operations per shader operation). So 48 pipelines multiplied by 2 would make 96 shader operations per cycle... and since it is clocked at 500MHz that would come out to be roughly 48 Billion Shader Operations per second (192 Billion ALU Operations per second - the number that Microsoft SHOULD of used considering the math that Sony used with the RSX) (provided the design can only do pixel or vertex operations instead of both at the same time... if that was the case then it would be 96 Billion Shader Operations per second (384 Billion ALU Operations per second) as technically each pipeline would be capable of doing 2 pixel AND 2 vertex shader operations per cycle... need clarification on that.)

With the RSX... and if it is based on the Geforce 7800 design and provided nothing major got changed (outside of changing the interface)... the Geforce 7800 has 24 pixel pipelines and 8 vertex pipelines for a total of 32 pipelines. 32 pipelines multipled by 2 would make 64 shader operations per cycle... but we have two clock rates we are looking at. According the specifications released by Sony they are listing 550MHz for the RSX (which is an awfully high clock rate for a GPU with nearly 300million transistors) and the clock rate for the ultra version of the Geforce 7800 which is a more down to earth 430MHz. At 430MHz at 64 shader operations per cycle the Geforce 7800 would be capable of 27.5 Billion Shader Operations per second (110 Billion ALU operations per second), now provided Sony *CAN* achieve 550MHz with that same architecture then that would come out to be 35.2 Billion Shader Operations per second (140 Billion ALU operations per second - the number that Sony was touting with RSX).

Ok final comparison...
*nVidia Geforce 6800 - 16pp/6vp - 44shops/cyc - 400MHz - 18.9 Bshops/sec (75.6 BALUOps/sec)
*ATI Radeon x850XT - 16pp/6vp - 44shops/cyc - 520MHz - 22.8 Bshops/sec (91.2 BALUOps/sec)
*nVidia Geforce 7800 (G70) - 24pp/8vp - 64shops/cyc - 430MHz - 27.5 Bshops/sec (110 BALUOps/sec)
*PS3 RSX (unconfirmed)- 24pp/8vp - 64shops/cyc - 550MHz - 35.2 Bshops/sec (140 BALUOps/sec)
*XBox360 XENOS - 48up - 96shops/cyc - 500MHz - 48 Bshops/sec (192 BALUOps/sec)

...and if each pipeline in XENOS can acutally process pixel and vertex data simultaneously...

*XBox360 XENOS - 48up - 192shops/cyc - 500MHz - 96 Bshops/sec (384 BALUOps/sec)

I need to find out if XENOS is capable of processing pixel and vertex data at the same time, but for now I am going to say no... in which case I would believe the 96shops/cyc - 48Bshops/sec (192 BALUOps/sec) number for now.

Regardless I am certainly interested in finding out more details on the RSX as well details on XENON (The CPUs in the XBox360). I will say that the numbers that nVidia/Sony has been giving us has been without context to make them look better than it seems. If I am off on any of the math then please let me know, especially in regards to the 2 Shader Operations per cycle per pipeline data as well as the 4 ALU operations per shader operation.

The GameMaster...

 

[DemoCoder]

*XBox360 XENOS - 48up - 96shops/cyc - 500MHz - 48 Bshops/sec (192 BALUOps/sec)

...and if each pipeline in XENOS can acutally process pixel and vertex data simultaneously...

*XBox360 XENOS - 48up - 192shops/cyc - 500MHz - 96 Bshops/sec (384 BALUOps/sec)

There's some voodoo numerology going on here.It doesn't matter whether the XENOS can do VS/PS/GS simultaneously, since they all share the same ALUs. The fact that it can do VS and PS at the same time doesn't magically double the throughput of the 48 ALUs. The only number that matters is ALU throughput. I think your counts of 2 shops/cyc for both XENOS ALUs and RSX ALUs is apples to oranges.

Stick to the number of 128-bit SIMD vector ops that can be done per clock.

 

[The GameMaster]

There's some voodoo numerology going on here.It doesn't matter whether the XENOS can do VS/PS/GS simultaneously, since they all share the same ALUs. The fact that it can do VS and PS at the same time doesn't magically double the throughput of the 48 ALUs. The only number that matters is ALU throughput. I think your counts of 2 shops/cyc for both XENOS ALUs and RSX ALUs is apples to oranges.

Stick to the number of 128-bit SIMD vector ops that can be done per clock.

Yea I know... that is why I am trying to clarify the number of ALUs present per pipeline in XENOS. Which is why I am choosing to believe 96 Shader Operations per Cycle or 48 Billion Shader Operations per Second (the quoted number in the specification list) which should come out to 192 Billion ALU Operations per Second. I am also trying to clarify the number offered by Sony/nVidia as I have reason to believe it is *NOT* over 136 Billion Shader Operations per second and instead is ALU Operations per second which would translate to roughly 35.2 Billion Shader Operations per Second at 550MHz. This is also assuming the RSX is using 24 pixel pipelines and 8 vertex pipelines.

I am just trying to figure out what is going on here and unfortuntaly we do not have enough details for us to verify Sony/nVidia's claims. Hopefully we will find out more details on the RSX when nVidia launched the Geforce 7800.

The GameMaster...

 

[DemoCoder]

The RSX has 2 ALUs per pipe, however it can execute more than just 2 SIMD ALU ops. As an example, it has a free NRM_16 op that would require 3 ALU ops on the R500 (dp, rsq, mul). So counting scalar ops in the figure is somewhat misleading, since alot of the ops are not comparable.

 

[nAo]

I am also trying to clarify the number offered by Sony/nVidia as I have reason to believe it is *NOT* over 136 Billion Shader Operations per second

136 ops per clock, not 136 gops per second.

and instead is ALU Operations per second which would translate to roughly 35.2 Billion Shader Operations per Second at 550MHz. This is also assuming the RSX is using 24 pixel pipelines and 8 vertex pipelines.

I wonder how you can translate ALU ops per second (what are ALU ops? please give us a clear definition) to shader ops per second (what are shader ops?) if you don't know a thing about RSX architecture.
You're just having some fun with numbers without even trying to make some sense, imho.

 

[jpr27]

1.) Wouldnt the unified structure and the edram play a role in this as well? Im guessing that RSX will have around 32-48 pipes but that just a guess on my part

2.) Do the ALU's on the Xenon or RSX have any new or extended functions in their designs?

3.) I would think that although we are talking about the GPU's on the respective consoles wouldnt the CPU's play into the shader equation as well? (For example: Some people say that the cell has the advantage in this area but I'll wait to see their final release before deciding)

Im in no way a hardware expert but just some questions

Over all I agree with Democoder on this as an Apple - Oranges comparison. I think you have to take into account the system as a whole