I'm feeling hoaxed because there is no real innovation here that justifies the leap from 780 to 980.
I agree that the perf:watt ratio improvement is nice but it doesnt really interest gamers, performance that's the name of the game.
Gamers (and computer users in general) are going to have to get used to a slower cadence for improvement. I'm surprised most aren't already; we've been at the same TSMC process node for the last three generations now on nVidia's side (Kepler 1/GK104 doing the heavy lifting mainly, Kepler 2/GK110/210 introduction, Maxwell as 750Ti/GM107 and Maxwell as GTX 970 and 980/GM204). Consoles stuck us in DirectX9 for years and years.
Show me the people doing what nVidia isn't and I'll cede the point, but literally every company from integrated SoC makers to NAND makers to RAM makers to affiliated-fabs for companies like nVidia and AMD are -all- approaching things efficiency first. It is critical, because we're not going to get the kind of process node shrink march of progress that we were counting on before we understood how difficult it is to fab a single chip, or how expensive.
The economics that we had hoped for in accordance with the prediction of Moore's observation aren't holding true anymore. Transistor density is not doubling every two years. Performance that in previous generations has just been left on the table because a new process was coming soon and taping out a chip is incredibly involved so all resources had to be devoted to the new design instead of getting the most out of the old. Now, we've got companies actually spending time at a given node and working a great deal of additional performance out of it based on actual architectural refinements.
GM204 is still a 5+ billion transistor chip, enormous processing power on the order of, what, 5 TFLOPs, and yet they have it down to a 165W TDP. Compared to Kepler, already an extremely efficient design, that's down a solid two billion transistors from a GTX 780Ti, a card that draws 265W easy.
AMD is doing the same thing with Tonga, essentially tuning the crap out of their Tahiti XT architecture to improve efficiency. Thus far it's not been as impressive of a showing, and it would be a problem for them to deprecate their existing GCN 1.1+ parts because they can't really afford to quit making them yet (while nVidia has relied largely on die-harvested Tesla/Quadro intended chips for their top end product stack, AMD doesn't really have the same luxury as they are not well featured at all in the HPC markets apart from some linpack trick machines).
I don't see the problem. Amazing improvements to thread/warp scheduling and execution pipeline improvements bring the shader unit count down but efficiency dramatically up; an even more modular design than Kepler, with 128SMs per SMM as opposed to Kepler's 192 per SMX. Kepler had some odd situations because of its emphasis on efficiency and modularity, notably that the 650Ti and GTX 780 both had a peculiar situation where an academic difference in GPC count could occur based on which parts of the full chip had to be harvested; nVidia is good at running a full product stack from a relatively small chip base. AMD has no such luck at the moment and has tons of SKUs being actively fabbed right now.
Help me find the issue that nVidia isn't addressing but other companies are. Even the HPC side of things is virtually the same - look at what Intel is doing with an evolutionary approach, refining old P6 u-arch for ultra-parallelism for Knight's Landing. Efficiency efficiency efficiency, because believe it or not, efficiency has way more to do with performance than one might think if all one cares about is getting a vastly higher 3Dmark score one card generation to another
[Edited by Agreed, 9/25/2014 2:58:31 AM]
I'm feeling hoaxed because there is no real innovation here that justifies the leap from 780 to 980.
I agree that the perf:watt ratio improvement is nice but it doesnt really interest gamers, performance that's the name of the game.
Gamers (and computer users in general) are going to have to get used to a slower cadence for improvement. I'm surprised most aren't already; we've been at the same TSMC process node for the last three generations now on nVidia's side (Kepler 1/GK104 doing the heavy lifting mainly, Kepler 2/GK110/210 introduction, Maxwell as 750Ti/GM107 and Maxwell as GTX 970 and 980/GM204). Consoles stuck us in DirectX9 for years and years.
Show me the people doing what nVidia isn't and I'll cede the point, but literally every company from integrated SoC makers to NAND makers to RAM makers to affiliated-fabs for companies like nVidia and AMD are -all- approaching things efficiency first. It is critical, because we're not going to get the kind of process node shrink march of progress that we were counting on before we understood how difficult it is to fab a single chip, or how expensive.
The economics that we had hoped for in accordance with the prediction of Moore's observation aren't holding true anymore. Transistor density is not doubling every two years. Performance that in previous generations has just been left on the table because a new process was coming soon and taping out a chip is incredibly involved so all resources had to be devoted to the new design instead of getting the most out of the old. Now, we've got companies actually spending time at a given node and working a great deal of additional performance out of it based on actual architectural refinements.
GM204 is still a 5+ billion transistor chip, enormous processing power on the order of, what, 5 TFLOPs, and yet they have it down to a 165W TDP. Compared to Kepler, already an extremely efficient design, that's down a solid two billion transistors from a GTX 780Ti, a card that draws 265W easy.
AMD is doing the same thing with Tonga, essentially tuning the crap out of their Tahiti XT architecture to improve efficiency. Thus far it's not been as impressive of a showing, and it would be a problem for them to deprecate their existing GCN 1.1+ parts because they can't really afford to quit making them yet (while nVidia has relied largely on die-harvested Tesla/Quadro intended chips for their top end product stack, AMD doesn't really have the same luxury as they are not well featured at all in the HPC markets apart from some linpack trick machines).
I don't see the problem. Amazing improvements to thread/warp scheduling and execution pipeline improvements bring the shader unit count down but efficiency dramatically up; an even more modular design than Kepler, with 128SMs per SMM as opposed to Kepler's 192 per SMX. Kepler had some odd situations because of its emphasis on efficiency and modularity, notably that the 650Ti and GTX 780 both had a peculiar situation where an academic difference in GPC count could occur based on which parts of the full chip had to be harvested; nVidia is good at running a full product stack from a relatively small chip base. AMD has no such luck at the moment and has tons of SKUs being actively fabbed right now.
Help me find the issue that nVidia isn't addressing but other companies are. Even the HPC side of things is virtually the same - look at what Intel is doing with an evolutionary approach, refining old P6 u-arch for ultra-parallelism for Knight's Landing. Efficiency efficiency efficiency, because believe it or not, efficiency has way more to do with performance than one might think if all one cares about is getting a vastly higher 3Dmark score one card generation to another
[Edited by Agreed, 9/25/2014 2:58:31 AM]
You hit the nail right on the head. While yes the new cards are not as big an improvement over the previous generation, they are more efficient in using what they have. One of the best examples is when I look at my Brother's and my Gaming PC's. We built them a year apart. Both systems have the same configuration, but different processors. His machine uses the Intel i7 2700K and I went with it's equivalent from a year later a i7 3770K. They both have the same clock speed, cache, and chipset. My 3770K is not noticeably faster, but it has a TPD of only 77W versus the 2700K's 95W.