I must be blind, as I do not see any; are there any maxwell based gpu workhorses, or any in the pipeline?
I am referring to a maxwell based gpu that would compare to at least a titan/ 780 in terms of cuda cores (2k+); I know of the geforce 750 and 750 ti, but I honestly do not know what nvidia expects me to do with a gpu with (only) 640 cuda cores
GTX 750 Ti is probably as good as it gets, currently. According to some rumors, GTX 880 might hit retail in 1-2 months.
I just got a 750 Ti and it’s awesome! Well, it’s better than my old 460.
The rumours about the GTX880 don’t scream out massive performance though, instead it looks to be about great efficiency and the memory model shift. The memory bandwidth estimate alone of 224GB/s (compared to the 336GB/s of the 780 Ti) could be a killer for math-light but bandwidth-heavy apps.
With no official roadmap this is gut speculation, but I am working on the assumption that the 880GTX is purely going to be a graphics card with no real benefit over the current gen for CUDA (maybe less power usage) but that big-maxwell (probably 2nd quarter next year) will provide the compute speed jump. This is what I have planned for so if anyone has any information to the contrary please let me know :)
GM107 got a large L2 cache increase to 2MB from GK107’s 256KB, I bet GM204(GTX 880/GTX 870) will have a large L2 cache increase over GK104’s 512KB to offset smaller 256bit memory bus and 224GB/s bandwidth.
Don’t count out the GM204 until you seen it properly benchmarked, it might surprise you. Don’t forget it’s second generation Maxwell also, there’s more improvement that we don’t know about yet compared to first generation Maxwell GM107.
If you want FP64 performance, then you have to wait for Big Maxwell.
That is a bummer, as I was hoping for an increase in bandwidth with the nex-gen Maxwell.
So it seems the GTX 780ti will still be the king for another 6 months at least.
What perplexes me is that the GTX 780ti significantly outperforms the GTX Titan Black both in compute (at least in my testing 32 bit) and in gaming;
On paper the Titan Black should be more powerful, so that has been a mystery. (Assuming 32 bit, as I know the Titan Black would outperform in the 64 bit realm).
Titan Black and the 780Ti both have 2880 Kepler cores. So it comes down to clock speed, and the 780 Ti is clocked ~10% higher so it’s going to be faster in your 32-bit calculations. An overclocked Titan Black should match the 780 Ti.
For all of us that compute, there’s really nothing to upgrade to until big Maxwell is released next year. Gamers may have reason to upgrade, but I’m a bit skeptical at how the 880 will perform at 1440p or higher resolutions with significantly less memory bandwidth than GK110.
What do you mean by “big Maxwell”? The cards with 2 GPUs?
If not, wouldn’t 880 be their biggest Maxwell, and the rumors seem to suggest October’ish release?
That is what I thought initially, but the difference in performance across a battery of tests(32 bit) is about 18%, while the difference in clock speed is about 8%. Also the GTX 780ti has better bandwidth tests as well.
This is in the same machine CUDA 5.5 Win7 64.
I would rather have a Titan Black because of the 64 bit Gflops, but I find it interesting that the 780ti has outperformed everything (K40, K6000, Titan) in the 32 bit realm in my limited set of tests.And those tests are using the WDDM driver as well, which is not as bad as it used to be.
I use Big Maxwell to describe the ‘full fat’ GM110 chip (or whatever they decide to call it) that mirrors the development of Kepler. The first Keplers were cutdown version designed to get reasonable performance on a smaller chip for getting the gaming card out the door (GTX600 series). The full-fat GK110 chips then went into teslas, and eventually a gaming equivalent (GTX700) when yields were high enough to justify the cost of a massive chip compared to the GK104.
Assuming they follow the same path here, the 880 will be a ‘GM104’, with the bigger chip version going through the Tesla line into mainstream next year when yields improve (GTX900?). Obviously if yields are not a problem or anything then they may follow a different path, or they may not even have a big maxwell in development (hard to believe since the success of the Titan).
I am holding out hope that if/when the GM110 arrives it has a 512 bit bus as well as a slight jump (+15%?) in compute cores, full speed double precision and 8GB of onboard memory. That is my hope for the Titan 2 anyway.
Re the discussion about the 780Ti vs the Titan Black, don’t forget the thermal effects on how far the chip up/downclocks. I imagine with the extra chip size the Titan throws out a chunk more heat that needs to be dissipated.