NVIDIA Fermi: Architecture discussion
- Thread starter Rys
- Start date
F**K MS...I need to find a new pic host..It's throwing a 509 error for me.
Last edited by a moderator:
Try my link.
Thanks for the link, it worked for me (the other one didn't).
Thnx. Seeing die's at such resolutions is an eye-opening experience.
Jawed
Legend
Being a bit silly for a moment
which would take 129 Eyefinity 6 cards to run
Jawed
If a pixel on a monitor corresponded with a "transistor", you'd need a grid of 22x35 2560x1600 monitors to see the transistors
which would take 129 Eyefinity 6 cards to run Jawed
Mintmaster
Veteran
That's because you're double counting the edge vertices. They get reused in adjacent patches. Okay, I'll agree that you can't cache the verts on every edge, so it'll be lightly lower than two, but that's it.In general with lower tessellation factors and anisotropy, you'll get notably less than 2 triangles per extra vertex. Go count some tessellated patches' triangles and vertices if you don't believe me. I counted one tessellated patch with 146 triangles and 87 vertices, earlier
Well it's there. SV_InsideTessFactor. There's nothing weird about it. You're right, that image is using the same tesselation factor for all edges on a mesh, so it's not going to give you any idea of what fancy tesselation structures look like.
It has nothing to do with comfort zone. If you have tiny triangles, all samples fit in one quad most of the time. If you can only feed one triangle every three clocks, then you will only get one quad into the shading engine every three clocks.
You're confusing threads with quads or doing something else. One thread every four clocks is the max you will ever need because a rasterizer can only generate 4 quads per clock at most and a thread has 16 quads in it. If you have one pixel triangles being fed to each rasterizer once every six clocks (on average), then each rasterizer will only produce a quad once every six clocks and sit idle 83% of the time. The thread will not be full until it gets 16 quads, which means one thread every 96 clocks.
The rasterizers can handle six times the triangle throughput of the TS before they become a bottleneck. Before that, though, setup will be a bottleneck, but even that can handle 3x the speed of the TS.
You're missing the point. This rasterizer can run at 6.25% efficiency with single pixel triangles instead of the 1% it does now due to limitations by the tesselator & setup. That's a big difference.
So you're saying 50M triangles before culling? No game does anywhere near that, so what's the point of this strawman example?
So? I'm talking about triangle counts per frame in Heaven. Why do you care about count per call?
Again, missing the point. What I just showed you is that the performance hit is entirely due to the 3-cycle per tri hit of the tessellator. There is no evidence that inefficiency of the "rasterisation/fragment shading/back-end" makes tessellator improvements pointless. Your argument is completely bunk.
ATI can chop the performance hit of tessellation by a factor of six without major architectural changes. All they have to do is make the tessellator generate triangles faster and double the speed of the triangle setup/culling.
Is this finally confirmed? That the rasterizer will always produce full threads, possibly composed of multiple triangles?
Mintmaster
Veteran
I would be shocked if this wasn't the case, as it would be a collossal loss of efficiency to save barely any silicon. I know for sure that Xenos did.
If a pixel on a monitor corresponded with a "transistor", you'd need a grid of 22x35 2560x1600 monitors to see the transistors
Jawed
I think you would need a bit less of screens because every chip is composed from several layers containing transistors.
If Fermi is anything like modern CPU it will have between 9 and 11 layers.
Sorry for OT!
Mintmaster
Veteran
Those are metal layers, not transistor layers. Wiring all those transistors to each other is not an easy thing...
Jawed
Legend
Whereas I think this is precisely the black hole that increased TS throughput would be feeding, thus pointless
I suspect addressing this problem is amongst the things that got deleted from Cypress.
Really?
Care to share the source for that info?
Jawed
Here's a curious question, why do different areas of the die have different colors? Is it false color, or diffraction? I would assume that it's diffraction, and that different areas of the chip might have different density of structures and/or regular vs irregular structure, but I'm curious if other factors could be involved, such as doping composition, metal layers, etc. The ultimate chip would be one that is simultaneously architecturally efficient, and produces beautiful die shots. Try optimizing for both!
Try optimizing for both!Isn't that why G100 is ~3b transistors?The ultimate chip would be one that is simultaneously architecturally efficient, and produces beautiful die shots.
Deadly Towers
Newcomer
oblique lighting with filters?
Perfect use case for Carell's Holographic hemisphere. :smile:If a pixel on a monitor corresponded with a "transistor", you'd need a grid of 22x35 2560x1600 monitors to see the transistors
Jawed
Here's a curious question, why do different areas of the die have different colors? Is it false color, or diffraction? I would assume that it's diffraction, and that different areas of the chip might have different density of structures and/or regular vs irregular structure, but I'm curious if other factors could be involved, such as doping composition, metal layers, etc. The ultimate chip would be one that is simultaneously architecturally efficient, and produces beautiful die shots.
My guess would be diffraction.
aaronspink
Veteran
Here's a curious question, why do different areas of the die have different colors? Is it false color, or diffraction? I would assume that it's diffraction, and that different areas of the chip might have different density of structures and/or regular vs irregular structure, but I'm curious if other factors could be involved, such as doping composition, metal layers, etc. The ultimate chip would be one that is simultaneously architecturally efficient, and produces beautiful die shots.
take a bowl or bucket of water. Put a thin coat of oil on it. Voila!
