RSX 1.8 T FLOPS's possibility

[Solid Snake]

"I also scratched my head which kind of wild marketing optimizm lead to these numbers. Things I could think of:

If texture interpolation now can be done in FP32:
Bilinear interpolation per one component takes 4 multiplies, 3 adds and 2 subs: 9 ops, x4 channels = 36 ops.
If result blend operation (when writing the results into FP32 buffer) can be done in FP32, it would add another 12 ops.
Add the original 8 ops of the other shareder unit (4xMADD)

We have 56 ops/cycle/pipe.
At (suppose) PS 32 pipes, it would be 1792 FP ops/cycle.
At 550Mhz, we have 985 GFlops.

Well, we do have a (theoretic) teraflop, we are almost there I am not too sure about number of PS pipes (24 or 32?). Anayway, I did not count the VS pipes (8, likely?). Also, if anamorphic filtering can be done in FP32, that would hike the texture unit FP to twice, and we would be at 1.8 Tflops

Now, before we claim a record, let see how it compares to a Cray supercomputer, model X1E, one liquid cooled cabinet configuration:

2.3 TFlops, memory bandwidth 3200Gb/s

Teraflops could be comparable, but it seems we have a memory bandwidth problem: Cray has 100x more to do any useful job with all these FP units


Roman"

[harl]

I think those flops should be more signifcatn if we
know how many Flops hava a 6800 or a G70

Anyway are there any oficial info about the nuber of pipes and
shaders of the rsx?

I think thay said over 300 million of transitors and 1.8 Tflops
beacause this are impresive numbers and dont really tell anything
about the architecture of the chip

And RSX 1.8Tflops + Cell 0.2 Tflops = 2 Tflops twice the
number of 360

Will Sony/Nvidia tell anything of the rsx sometime or
becaues it's a console chip will keep "secret" the true specs of the chip

Anyway despite the power of the RSX chip I think 22 GB/s are not enough
even if it's dedicated memory, especialy if they want 1080p as
a standard

[Karma]

Oh great, not another one of these........................

[GigaDrive]

The 2.3 Tflops of that Cray X1E supercomputer are *real* Tflops, even if they are peak & theoretical.

whereas the 1.8 Tflops claimed for RSX are the "graphics version" of flops
.in other words, NvFlops.