Panajev2001a
Veteran
Taken form a post of mine on another board...
So 1 TFLOPS Broadband Engine...
[church lady]let's do some math, shall we ?[/church lady]
Let's say that counting a bit of efficiency and AI and physics and stuff that we get for pure T&L something like 250-300 GFLOPS ( 25-30 % ) in good scenarios ( optimized )...
That is something like 40-48x higher FLOPS rating than PlayStation 2's theoretical max ( 6.2 GFLOPS )...
RAW polygon pushing power, for the EE, should be around 102.85 MVertices/s ( 5 cycles fast transform for VU1 and 7 cycles transform for VU0... no lighting.. gross calculation as 6.2 GFLOPS doesn't all transfer to that amount of polygons per second... )...
Let's say that accounting various factors we have 35x that polygon pushing power ( even if we ammitted up to 48x the GFLOPS rating for the Broadband Engine T&L... ), that would give us... ~3.6 GVertices/s...
At 60 fps that would be ( no lighting... ) 60 MPolygons/frame...
Let's admit inefficiency and stuff and let's add lighting and let's drop our figure by a factor of at least 6.67-7.5... we reach 9-8 MVertices/frame
This is approximately 8x the amount of polygons titles like S.T.A.L.K.E.R. plan to use with NV35 and R350 ( this game is not out yet... ) at good frame-rate...
[...]
I would not mind photon mapping ( cheaper than radiosity it seems computationally ) and a 2.5-3 MVertices/frame polygon count which would mean at 60 fps... 150-180 MVertices/s ( you still have to add all the pixel shading and texture tricks the APUs in the Visualizer would do )...
Counting that I had lighting in mind when I said 8-9 MVertices/frame... going to 2.5-3 MVertices/frame means a ~3x performance hit enabling photon-mapping...
180 MVertices/s is only ~1.75x higher than EE's output ( RAW polygons for the EE, no lighting... )...
Considering a more realistical 15 MVertices/s ( I think it can be pushed more even in real world conditions ) with lighting... we get that 180 MVertices/s is like ~12x higher than 15 MVertices/s... better comparison would be having numbers for the EE doing photon mapping... although the EE would hit memory and bandwidth limitations and get much more hurt than the Broadband Engine ( designed for massive data flows... )...
[...]
Generally we can expect PlayStation 3 to be more like 200-300x PlayStation 2's RAW specs ( the 1,000x performance figure takes other things into account... )...
2.5-3 MVertices/frame / 300 = 8,330-10,000 Vertices/frame which means 0.5-0.6 MVertices/s at 60 fps using photon-mapping...
[...]
I think that photon mapping is neat ( and cheaper than raytracing... they said the same in that Halo 2 article/presentation ) especially for simulating global illumination without using too many lights...
Considering actual polygon counts, lighting and texturing ( and effects done with the Rasterizer ) on PlayStation 2 ( and even Xbox )... having nice pixel shading and higher resolution textures ( better texture filtering as well ) would make 1.5-2 MVertices/frame look very good ( this is still 1.5x-2x the polygon count per frame than Stalker ) especially if we have global dynamic illumination with photon mapping...
What do you think ?
Comments are welcomed...
So 1 TFLOPS Broadband Engine...
[church lady]let's do some math, shall we ?[/church lady]
Let's say that counting a bit of efficiency and AI and physics and stuff that we get for pure T&L something like 250-300 GFLOPS ( 25-30 % ) in good scenarios ( optimized )...
That is something like 40-48x higher FLOPS rating than PlayStation 2's theoretical max ( 6.2 GFLOPS )...
RAW polygon pushing power, for the EE, should be around 102.85 MVertices/s ( 5 cycles fast transform for VU1 and 7 cycles transform for VU0... no lighting.. gross calculation as 6.2 GFLOPS doesn't all transfer to that amount of polygons per second... )...
Let's say that accounting various factors we have 35x that polygon pushing power ( even if we ammitted up to 48x the GFLOPS rating for the Broadband Engine T&L... ), that would give us... ~3.6 GVertices/s...
At 60 fps that would be ( no lighting... ) 60 MPolygons/frame...
Let's admit inefficiency and stuff and let's add lighting and let's drop our figure by a factor of at least 6.67-7.5... we reach 9-8 MVertices/frame
This is approximately 8x the amount of polygons titles like S.T.A.L.K.E.R. plan to use with NV35 and R350 ( this game is not out yet... ) at good frame-rate...
[...]
I would not mind photon mapping ( cheaper than radiosity it seems computationally ) and a 2.5-3 MVertices/frame polygon count which would mean at 60 fps... 150-180 MVertices/s ( you still have to add all the pixel shading and texture tricks the APUs in the Visualizer would do )...
Counting that I had lighting in mind when I said 8-9 MVertices/frame... going to 2.5-3 MVertices/frame means a ~3x performance hit enabling photon-mapping...
180 MVertices/s is only ~1.75x higher than EE's output ( RAW polygons for the EE, no lighting... )...
Considering a more realistical 15 MVertices/s ( I think it can be pushed more even in real world conditions ) with lighting... we get that 180 MVertices/s is like ~12x higher than 15 MVertices/s... better comparison would be having numbers for the EE doing photon mapping... although the EE would hit memory and bandwidth limitations and get much more hurt than the Broadband Engine ( designed for massive data flows... )...
[...]
Generally we can expect PlayStation 3 to be more like 200-300x PlayStation 2's RAW specs ( the 1,000x performance figure takes other things into account... )...
2.5-3 MVertices/frame / 300 = 8,330-10,000 Vertices/frame which means 0.5-0.6 MVertices/s at 60 fps using photon-mapping...
[...]
I think that photon mapping is neat ( and cheaper than raytracing... they said the same in that Halo 2 article/presentation ) especially for simulating global illumination without using too many lights...
Considering actual polygon counts, lighting and texturing ( and effects done with the Rasterizer ) on PlayStation 2 ( and even Xbox )... having nice pixel shading and higher resolution textures ( better texture filtering as well ) would make 1.5-2 MVertices/frame look very good ( this is still 1.5x-2x the polygon count per frame than Stalker ) especially if we have global dynamic illumination with photon mapping...
What do you think ?
Comments are welcomed...