NVIDIA: Beyond G80...
- Thread starter kyetech
- Start date
A random thought from a refrain oft repeated for certain other products: just because the G80 is a year late, what makes you think the G90 is similarly late?
I don't think the situation is quite the same, but, it might not be quite as long a wait. The G71 was *supposed* to be the G80, after all....
I also would agree that the MUL->MAD enhancement has precedent, but, the MUL appears to be hiding in the interpolater, so, I'm not sure that makes sense. Also, raising issue-breadth risks decreasing maximum utilization, so it's a tradeoff between spending the trannies there or in another cluster.
Yeah, well, I think basic physics tells us that was an impossible goal.Oh, we mind
That hardly matters, however, if you don't even know what the codename stands for. Heck, for all we know, no such codename might even exist!
Uttar
I'm not sure why they are associating weak demand for G80 with the upcoming entry and mid-level variants. They are targeting completely different markets. Weak demand would most certainly be related to the lack of compelling titles and the impending R600 launch. Many, many people are also quite happy with last generation performance and IQ in current games.
Jawed
Legend
http://www.techreport.com/sympoll/polllist.php?dispid=120
Also, talk of G90 and so on seems to have missed out one factor: if G80 turns out to be a dud in comparison with R600, what kind of turn-around is NVidia going to have to do.
It's looking pretty certain, now, that R600 will have ~50% more GFLOPs (345 versus >500). Am I meant to believe that GFLOPs will not be important in the D3D10 generation?
Jawed
Also, talk of G90 and so on seems to have missed out one factor: if G80 turns out to be a dud in comparison with R600, what kind of turn-around is NVidia going to have to do.
It's looking pretty certain, now, that R600 will have ~50% more GFLOPs (345 versus >500). Am I meant to believe that GFLOPs will not be important in the D3D10 generation?
Jawed
Where does that come from? Haven't seen any references about that.
Also, if it should turn out to have 50% more flops, nV can always fight it by pricing their parts lower. Should be easy by then with a few months into production of G80.
Yeah, but if they need some time to get there, the lower pricing will keep them floating so long.
There's a limit to how much NV can accomplish. Assuming that G80 doesn't have huge redundancy built-in, it looks like a cluster costs 80M transistors. If you wanted your 80nm part size to ~= 90nm part, then you might be able to add two clusters. 25% increase....
If you went for the MUL->MAD upgrade, then you have to somehow feed an extra arg (maybe more, considering its MIA status) to the units. That bandwidth is going to cost, and it has to be carried on all 8 clusters. Probably a lot less than 160M trannies, though, and it might boost your numbers by a similar amount.
ROP upgrade plus clockspeed boost+GDDR4 plus ALU tweaking??
One other question -- given how well the 8800s overclocks, why the ban? Are they trying to leave room for a small incremental upgrade?
Jawed
Legend
Yeah.Jawed is probably discounting the MUL until it's found
It's not a trivial thing like Purevideo on early 6800 series, either.
Jawed
heh, I bet it'll be extremely amusing to see how wrong we all are once the actual roadmaps and/or products are out there. Or maybe not and we'll be mostly right, but I seriously doubt that, TBH. As for overclocking, you'd expect them to want to be able to release a small incremental bump with the 80nm shrink indeed, and they wouldn't want to repeat the G71 overclocking fiasco.
What's funny is that if NVIDIA decides to be very aggressive, they could manage what I described as G95 performance not in H2 2008, but rather in H1 2007! SLI is an interesting technology to have at your disposal indeed, and G80 scaling seems to be excellent, so we'll see what happens!
Uttar
What's funny is that if NVIDIA decides to be very aggressive, they could manage what I described as G95 performance not in H2 2008, but rather in H1 2007! SLI is an interesting technology to have at your disposal indeed, and G80 scaling seems to be excellent, so we'll see what happens!
Uttar
heh, I bet it'll be extremely amusing to see how wrong we all are once the actual roadmaps and/or products are out there. Or maybe not and we'll be mostly right, but I seriously doubt that, TBH. As for overclocking, you'd expect them to want to be able to release a small incremental bump with the 80nm shrink indeed, and they wouldn't want to repeat the G71 overclocking fiasco.
What's funny is that if NVIDIA decides to be very aggressive, they could manage what I described as G95 performance not in H2 2008, but rather in H1 2007! SLI is an interesting technology to have at your disposal indeed, and G80 scaling seems to be excellent, so we'll see what happens!
Uttar
I'd dare to speculate that G8x as a base architecture (irrelevant of future codenames) will remain for quite a few years to follow. The rather dramatic changes in G8x compared to predecessors do not sound to me like a short term plan at all.
I haven't a single clue if Microsoft and IHVs are already negotiating for "DX-Uber-Next" (or "D3D11"-whatever). It sounds a bit too early to me right now, but when it starts it would be interesting to see what advancements are being proposed and what how it will shape out after countless of drafts and IHVs screaming in between about HW budgets. That future "deflection point" might be a field worth investigating for major architectural overhauls or maybe even not. Anything in between sounds rather like efficiency tweaks within the same base architecture.
How would I know?
Uttar
Why increase the Multisampling sample densities and not further expand on coverage mask sampling instead? I had a probably dumb idea just a while ago while posting at 3DCenter; IF it would be possible to find a way to use N Transparency Multisampling (or call it alpha to coverage-whatever) for N coverage mask samples (irrelevant if like N/4 z/stencil are being stored), it would be more than just interesting.
Transparency SS or adaptive AA might give excellent results quality-wise, but if your scene is overloaded with alpha test textures, the gain over full scene SSAA isn't all that much after all. The unfortunate thing about Coverage sampling is that you get as much Transparency samples as Z/stencil samples are being stored.
Uhmmm kindly ignore the above if it's a stupid idea