The Official NVIDIA G80 Architecture Thread
- Thread starter Arun
- Start date
Jawed, I think this may be the root problem/difference. As the ALUs clock rises wrt tex units, the relative cost of doing a tex lookup rises, but no amount of latency hiding can mask throughput issues.
There is a limit to the throughput of the tex units, and a maximum speed for the ALU unit. If you want to keep both operating at peak performance, there's a ratio between tex and alu ops that is implied. As the tex units in each cluster can apparently produces 4 Vec4 per its clock, while the alu ops can perform 16 scalar ops per its clock, the ratio of ops looks to be roughly the same as the clock ratio ~2-1/3. Of course, the ALUs are likely busy about a quarter of a clock cycle doing perspective correction for addressing, so a roughly 2:1 ratio. So, assuming full utilization, with each batch running 2 ALU ops per tex lookup, hiding ~200+ cycles looks to be 128 scalar ops, or 32 Vec4-sized ops.
Yes, hello and welcome to B3D! The results on R580 are broken, you (if you're the OP at R3D) completely misunderstand the results presented and G80 effectively branches for free. You* were told this by Demi, DC and others over at Rage3D and they were correct then. They'll be correct here again if you* really want to ask again......Lord CorDox said:
btw: I think the penalty for not believing Demi about something 3D or DC about math is running an S3 Virge for life
Moving your 3D request to the programming forum so as not to pollute this thread.
Where Mike Houston of Stanford --and one of the developers of GPUBench-- has already pitched in. Enjoy! 
http://www.beyond3d.com/forum/showthread.php?p=874511#post874511
http://www.beyond3d.com/forum/showthread.php?p=874511#post874511
Jawed
Legend
Interruption in Programming
http://www.beyond3d.com/forum/showthread.php?t=21971
Damn, that's funny. July 2005.
Jawed
http://www.beyond3d.com/forum/showthread.php?t=21971
Damn, that's funny. July 2005.
Jawed
Well, that's your problem right there. Who says they lack it? More importantly, who says they require support from anywhere external (presumably you're thinking of Microsoft) in the first place? David Kirk pointed out repeatedly that the two concepts of unified programming model and unified hardware assets were independant. . . G80 proves it is true --just not quite the way we all thought he meant it at the time. . .
Well, it would be a lot more granular and variable than that. It would be happening per frame. I don't know if we've decided yet if there would be any min/max stops, or if it might go 128/0 or 0/128 in spots.
Actually, "per frame" isn't even granular enough if we take the usage graphs that both ATI and NV have shown literally, based on B3D member RoOoBo's work.
The way I expect the unified pipelines work is that you just have N threads in flight at any given time, some of which will be pixel threads, while others will be vertex threads. nVidia will have a scheduler that decides which threads to execute during any given clock. Thus, in essence, you just have a cluster of processing units that can be assigned a thread to work on each and every clock, independent of whether it's a pixel or vertex thread. The load balancing would thus be perfect, and would be an improvement even for DX9 games, as you don't have to worry about small triangles eating up your vertex units, or large triangles eating up your pixel units while hardly touching the vertex units.
Actually, you're doing something massively wrong if the drivers have anything to say about it when running realworld applications. And it balances the workload constantly, not just per-frame. I'm not going to write an introduction to the subject however, as plenty of people have summarized it excellently in the past already.
Uttar
Would there be anything stopping nvidia or ati producing cutdown unified chips solely to make DX9 games faster and more efficient whilst saving transistor cost by excluding DX10 requirements not needed? And if not, then obviously it would have to be sooner rather than later but since DX9 games will still be around for a considerable length of time, would it be worth it to either IHV?
I always though unified shaders were needed to be supported at a software level as well as in a hardware level. Therefore DX10/unified shader architecture. Now it could be assumed that the scheduler in G80 is dynamically allocating the shaders per frame.
So, theorectically a unified shader architecture will be more efficent than a fixed function architecture in DX9 games? but if so, why didnt ATi use the R500 and make a derivative for the PC market to face nVIDIA (7800GTX) instead of making (spending much R&D as well) another GPU that became known as R520?
Im presuming that they were going for a safer route instead. Or the decision was made long time ago, even before the R500 project took place.
When is the seond part of the B3D review of G80 coming? if there is any
