Right now I have a GTX 660ti. I haven't upgraded in awhile so now it is time t do so. I currently have a 750 watt power supply with two 6 pin connectors for the graphics card. What card would be a nice upgrade from a 660ti. Currently working with a $200 budget. Thanks for the suggestions!
30
Comments
$200 is roughly 270x / 960 territory. The 270x is slower than the card you already have and the 960 while slightly faster isn't that much of a difference to warrant spending $200.
For a worthwhile upgrade you should be looking at cards from the 280x up on AMD side and 970 up on Nvidia, none of which fall in your budget.
Thanks. Not looking to keep my current pc long term so hence why I do not want to dump a lot of money on upgrades.
here is my power supply specs...
http://www.newegg.com/Product/Product.aspx?Item=N82E16817139021
I was thinking about purchasing this card. Seems like a good bump from my 660 ti.
http://www.newegg.com/Product/Product.aspx?Item=N82E16814487093
Really because I was thinking the 660ti would be out of date by now. I always choose Nvidia cards.
Thanks so much for that. I might not need a card after all. I could just pocket that extra $200 towards a brand new pc sometime next year.
That's got 4 PCI-E 6+2 pin connectors. You're fine on the power supply and can buy whatever single-GPU card you want without having to worry about power delivery.
It's an upgrade, yes. But not really enough of an upgrade to justify the bother. A GTX 660 Ti has 1344 shaders, while a GTX 960 has only 1024. The GTX 660 Ti also has 3 memory channels, compared to 2 for the GTX 960. The GTX 960 is able to use its resources more efficiently, which is why it is still an upgrade. But it's not a very big upgrade, and that's a waste of money. If you're not willing to get at least a Radeon R9 290 or a GeForce GTX 970, I wouldn't upgrade.
Architectural differences can be a big deal. Kepler botched the architecture for their warp schedulers, with only four warp schedulers for six groups of shaders. That's likely the biggest thing Maxwell fixed, keeping four warp schedulers but dropping to four groups of shaders per SMM, which allows more SMMs in the same die space/power consumption and thus better performance.
Enough with the stupid FUD about drivers. I've had a Radeon HD 5850 for 5 1/2 years and the nearest thing to driver problems I've had with it is that it's undocumented that if you plug in multiple monitors, the card doesn't clock down as far at idle. For comparison, the contemporary Nvidia Fermi cards didn't clock down at all at idle if multiple monitors were plugged in.
GTX 960 is maybe 20 - 30% faster than GTX 660 TI:
http://www.tomshardware.com/charts/2015-vga-charts/compare,3673.html?prod%5B7372%5D=on&prod%5B7248%5D=on
EDIT: When looking at the link, look especially at "20 index at 1080p", or "21 index 2160p". Those give you the averages measured /EDIT
GTX 960 costs a lot of dollars for a small performance increase. I'd suggest either keeping your current card, or if you upgrade get at least something like R9 280x that gives 40%-50% performance gains for about 250$
http://www.tomshardware.com/charts/2015-vga-charts/compare,3669.html?prod%5B7284%5D=on&prod%5B7248%5D=on
If you need upgrade and can find the money, something like GTX 970 would also be a good solution. It costs about 350$ so it's expensive as hell, but it would offer about 100% speed increase compared to your current card:
http://www.tomshardware.com/charts/2015-vga-charts/compare,3669.html?prod%5B7477%5D=on&prod%5B7248%5D=on
Try to go for big upgrades when upgrading your GPU if possible. It's a lot cheaper in the long run to pay for big upgrade every 5 years than it's to pay for a small upgrade every 2 years.
While it generally holds true - going from 1344 Kepler shaders to 1024 Maxwell shaders is definitely not a downgrade. And Maxwell redoes the memory architecture as well, so just counting memory channels isn't a good indicator.
Heck, the 780Ti has 2,880 and the 980 only has 2,048, and also dropped from a a 384bit memory bus to a 256bit bus, and down from 240 texture units to 128, but the 980 handily bests the 780Ti in pretty much any meaningful benchmark you care to use. Nearly every number on a spec sheet would point to the 780Ti being the faster card, except it's not, and by a decent margin.
Not saying the rest of your analysis is wrong (or right) - but you probably ought to update your old thumbrule of just counting shader units. Maxwell pretty well throws that out the window.
Counting shaders is pretty good as a way to get a ballpark comparison for two GPUs of the same architecture. Such a ballpark comparison might well be off by 20%. But it's not going to be off by a factor of 3. If you count memory capacity (as some people who don't know much about GPUs do), you can easily be off in performance by a lot worse than a factor of three.
When going from the GTX 780 Ti to the GTX 980, there are really two mitigating factors that make a naive shader count off. One is clock speed, as it goes from a boost clock of 928 MHz to 1216 MHz. That's a 31% difference right there. The other is that Maxwell's warp schedulers are more efficient, as Kepler botched that part of the architecture, as I referred to in the post below the one you quoted. In this particular case the architectures are similar enough that it's a closer comparison to say going from 15 SMXes at 928 MHz to 16 SMMs at 1216 MHz, which would actually overstate the GTX 980's advantage. But comparing the number of compute units of different architectures usually gives you nonsense (for example, the Radeon R9 290X has 44 of them); going from Kepler to Maxwell just happens to be a rare exception due to other architectural similarities.
Yup. That was my point.
Because you were trying to comparing a 660Ti to a 960 using all that flawed stuff.
Thanks for clearing that up.
Shader count was a very accurate comparison for nvidia gpus in the past few years, simply because the architectures of previous card generations were so similar, and because fermi and kepler were used for 2 generations each. 400 and 500 series is all fermi with incremental updates and fixes and so is 600 and 700 all kepler. What made comparison even more accurate was a single chip being binned a long way down the stack into multiple gpus.
And between fermi and kepler there wasn't that much architectural(and other) difference that would increase the margin of error for comparison. What's actually pretty funny is that there isn't that much actual difference between the last kepler and maxwell, however like Quizzical wrote earlier in the thread, because of some of the issues and problems with kepler, maxwell brings a lot raw power for less cores by simple optimizing and fixing various parts of the kepler architecture, thus making it look like it's a unprecedented revolutionary advancement