G80 vs R600 Part X: The Blunt & The Rich Feature
- Thread starter Jawed
- Start date
If you just reduce the width of the texel data paths and filter units, I don't think it would be that much cheaper. Splittable filter units might make sense. You'd also have to cut a bit from the texture address units.
It's useful if you want some unfiltered samples, but writing a shader with a mix of Sample() and Load() + bilinear/trilinear filter seems awkward if all you really want is filtered samples.
Sure, but even with past experiences (which may differ a lot for different people) it remains a gamble.
So apparently R580+ never came to fruition (or never existed in the first place). That would explain why X1950 had the same core clock as X1900.
Both actually. If IHVs made hardware around current games, and game developers optimize around current hardware, we'd be stuck in a chicken and egg situation. Someone has to take the step ahead, and it's usually the IHVs.
AFAIK the filter is the most expensive part of the texture unit. Which would also explain why even the 9700 could sample 128bit textures, but not filter them.
Sure, but wanting unfiltered samples is not uncommon.
I honestly can't add anything to the technical discussion... but this I really wanted to address. A lot of comments on the last few pages of the thread have talked about gamers on a budget much the same was as you would someone who lives in a foriegn country who's customs and culture you really don't understand in the slightest. Speaking as one of those people on a budget, comments like the quoted text are infuriating.
I don't wait for things to hit budget bin before buying them unless I'm not particularly interested in them. I get plenty of stuff at or near their launches, even if I'm only capable of running at medium or lower settings. Nothing's been so ugly at those settings that the fact that it'd look considerably better on a better system gets in my way of playing them. I'm basically offered a choice between buying a new video card, or buying 10 games. Between playing a game on lower quality settings or not playing a game at all, the choice is usually pretty clear. But god, I don't wait for the games to be discounted. It's just that I can choose between playing fewer games with prettier graphics or playing more games with tolerable but reduced graphics. As a gamer, I'd rather actually play the games.
And as far as life on a video card goes... I'm still running a 9700pro. So my video card has been in my system for over 4 years (got it around the launch of the 9800pro). The only game thus far that I've been interested in that I can't play is BioShock. Eventhing else is playable, even if I am forced to reduce the graphic quality. I am going to upgrade, and hopefully soon, but it's not something I can go into easily, and it definitely has to last. My current setup has served me well enough at this point that I haven't missed anything too serious except some of the eye candy. I can live without that though. I can't live without the games themselves.
Course, it does help that I'm not horribly fond of (most) FPS games, and that DX9 remained the standard for a relatively long period of time. SM3 is required with BioShock, but that's honestly the first time I've fallen behind minimum spec on something I'm even remotely interested in this entire time.
Last edited by a moderator:
Silent_Buddha
Legend
So you talk for the whole world? Nice
And you do? Nice.
I believe, like everyone else here, he was just stating his opinion.
Obviously some of us disagree on the whole budget/longevity gamer segment and we'll just leave it at that.
Regards,
SB
Well, the fact that you own a Crystal Ball, yet have refused to help me with some loto numbers, for the 10 million Euro jackpot that was on the table this weekend is infuriating to me!Now that may have no connection to the topic, but I'm still very furious about that...may your copy of Bioshock refuse to activate itself or something
) J/KSure it is as it's a personal statement and not speaking for everyone.
People try to categorize markets all the time, but there's always going to be a significant minority that acts differently. In many cases I doubt if there's even a majority.
I'm usually not with the majority when it comes to purchasing behavior, but I'm definitely in the camp of someone who doesn't upgrade often and buys games when they're new. Can't remember the last time I bought an old game. I usually just look at some reviews and try to determine if my current system can handle a new game and provide a decent experience.
Is piling on SPs and then requiring an uber smart compiler? (Hello Itanium) Which GPU is more elegant? The one that does more with fewer SPs or the one with smaller and simpler SPs but with many of them going to waste?
This is much like the arguments in CPU land. Which is more elegant, an in-order CPU with simple instruction issue, lots of threads, and hardware scouting, or an out-of-order CPU with uber complex ROBs, LS queues, branch predictors, etc? One has slow serial performance, but you can build alot of em on one chip, the other has great serial performance, but chews up alot of silicon, harder to test and verify, etc
I have never understood why you think the G8x design is brute force. I've always felt the exact opposite. ATI's chips have far higher theoretical ALU power on paper, but have trouble keeping up with the NVidia's chips when fed complex, general purpose ALU workloads. An 800SP chip should stomp all over a 240SP (or "480SP equivalent") chip if it was "elegant" IMHO, or atleast, efficient.
The ATI design IMHO depends too much on moving decisions from runtime to the driver, and there are simply limits to optimizations that can be performed on static code.
I think you need to look at some more shader intensive apps to see if thats really the case. i.e. Its not for no reason that HD 4850 is some 30+% faster in Vantage Extreme mode.
Mintmaster
Veteran
Heh, DC I don't think you realized that the post Jawed directed you to is almost a year old.
Anyway, FWIW I agree with you. NVidia destroyed ATI in the perf/mm2 department when R600 came out, and still edged ahead when RV670 was out, despite having far lower ALU count.
Only with the amazing efficiency leap of RV770 did ATI's strategy really show merit. It's not elegant, but when their quad of 5x1D processors takes half the space of NVidia's 8x1D processors (I know, 25% bigger process and double the clock speed), you have to start thinking that the 'elegant' solution is a bit too smart for its own good.
By the way, the 800 SP chip does stomp all over the 240 SP chip in tests where SP performance is the limiting factor.
Anyway, FWIW I agree with you. NVidia destroyed ATI in the perf/mm2 department when R600 came out, and still edged ahead when RV670 was out, despite having far lower ALU count.
Only with the amazing efficiency leap of RV770 did ATI's strategy really show merit. It's not elegant, but when their quad of 5x1D processors takes half the space of NVidia's 8x1D processors (I know, 25% bigger process and double the clock speed), you have to start thinking that the 'elegant' solution is a bit too smart for its own good.
By the way, the 800 SP chip does stomp all over the 240 SP chip in tests where SP performance is the limiting factor.
I don't really trust 3DMark as a target because driver devs tend to tweak compiler heuristics towards those workloads (I'm not saying driver detection, I'm saying, you compile, look at the output, and then go back and adjust your heuristics so that the output is better. This is not guaranteed to avoid regressions on other workloads. I do compiler work on a regular basis, and this is what we do, compile popular apps, and tweak)
Last time around, when the R600 was released, we went through the same arguments, and I remember Jawed discounting heavy shader workloads that had a few texture fetches in them. But even if we take the benchmarks on face value, the R770 should be winning by even high margins.
The proof is in the pudding. Take a hetereogenous collection of shader workloads, that are not ROP bound, and show me an R770 beating a GT200 by close to paper spec margin. On paper, it has a big theoretical advantage, but in the real world, it doesn't pan out. So, either utilization is low, or they made a poor decision in spending too many trannies on ALUs and not enough on TMUs to balance out the demands of the workloads.
In any case, I do not find a hardware design that depends on complex software to be "more elegant", just different. In another thread, we're having a discussion about software cache management (CUDA/CELL LS style) vs HW managed cache. Again, what's more elegant? It's like a debate over static vs dynamic typing. There is no "correct" answer.
Jawed's pronouncements are a factor of his personal aesthetics. I personally think NVidia's design is more elegant, from compiler point of view, it is much easier to optimize for, simpler to understand. Those are my aesthetics. I'm not passing them off as objective
We'll see what happens in another cycle. There was tons of FUD last cycle about NVidia's future prospects before the G80 came out too, and yet the G8x dominated for a long time. I don't think NV to where they are today by being too stupid (NV3x aside), so there must be a reason behind the decisions made for GT200 that we're not seeing yet. NVidia loves high margin chips, and they clearly know how the yield calculus works out.
Last time around, when the R600 was released, we went through the same arguments, and I remember Jawed discounting heavy shader workloads that had a few texture fetches in them. But even if we take the benchmarks on face value, the R770 should be winning by even high margins.
The proof is in the pudding. Take a hetereogenous collection of shader workloads, that are not ROP bound, and show me an R770 beating a GT200 by close to paper spec margin. On paper, it has a big theoretical advantage, but in the real world, it doesn't pan out. So, either utilization is low, or they made a poor decision in spending too many trannies on ALUs and not enough on TMUs to balance out the demands of the workloads.
In any case, I do not find a hardware design that depends on complex software to be "more elegant", just different. In another thread, we're having a discussion about software cache management (CUDA/CELL LS style) vs HW managed cache. Again, what's more elegant? It's like a debate over static vs dynamic typing. There is no "correct" answer.
Jawed's pronouncements are a factor of his personal aesthetics. I personally think NVidia's design is more elegant, from compiler point of view, it is much easier to optimize for, simpler to understand. Those are my aesthetics. I'm not passing them off as objective
We'll see what happens in another cycle. There was tons of FUD last cycle about NVidia's future prospects before the G80 came out too, and yet the G8x dominated for a long time. I don't think NV to where they are today by being too stupid (NV3x aside), so there must be a reason behind the decisions made for GT200 that we're not seeing yet. NVidia loves high margin chips, and they clearly know how the yield calculus works out.
Mint, I'm just saying putting "smarts" on the chip vs putting them in the compiler appeals to different people's aesthetics. Software developers tend to like hardware that is easier to deal with (less pathological bottlenecks, less bookkeeping, e.g. cache is better than manual cache management, etc), and HW devs tend to like simpler HW that is easier to design/test and move logic off chip when possible. Different people favor different approaches, just like the recent trend towards simple in-order CPU cores.
I'm not saying the R770 is a bad chip, I'm just saying "blunt" vs "elegant" is subjective. The clock issue is another design debate, high clock vs high work per clock. I remember there was a time when people were salivating over Fast14 and ATI running way ahead of NV in ALU clocks. There, custom high speed ALU logic was deemed "elegant"
I'm not overly impressed with the GT200. It's hard to follow on the G80. Maybe a Folding@Home war with hand-optimized CUDA vs CAL code on GT200 vs RV770 would settle things.
I'm not saying the R770 is a bad chip, I'm just saying "blunt" vs "elegant" is subjective. The clock issue is another design debate, high clock vs high work per clock. I remember there was a time when people were salivating over Fast14 and ATI running way ahead of NV in ALU clocks. There, custom high speed ALU logic was deemed "elegant"
I'm not overly impressed with the GT200. It's hard to follow on the G80. Maybe a Folding@Home war with hand-optimized CUDA vs CAL code on GT200 vs RV770 would settle things.
Actually, its farily easy to see what other effects our Vantage improvements had - generally they benefitted all DX10 apps.
Some of us are still salivating, hoping, and waiting for that.