[Computerbase] GTX960 & R9 380: 2GB vs. 4GB -- Conclusion: Get a 4GB card

[Goatsecks]

Yup, I like to point out the hypocrisy of certain team members who ignore awful price/performance and VRAM limits but only when it comes to their brand. I can see how some of them are stuck between a rock and a hard place between $160-180 GTX960 2GB is VRAM gimped -- cannot be objectively recommended -- and $210 R9 380X smashes 960 4GB in price/performance. I guess that means NV has a gigantic gap of no card worth buying between $150->$280 space. It's no wonder this thread created so much controversy among the loyalists of a certain team. Notice how no one is denying how 380 4GB is a better buy against R9 380 2GB from the objective members of this forum. Just saying. :sneaky:

The 380x is indeed, objectively, a better card.

Couldn't agree more about the VRAM limits as well as the hypocrisy, well said. I am curious about any analysis regarding the 4GB VRAM headroom on the fury lineup, with each one of them being push as 4K cards.

However, we do not need pages and pages of analysis to know that the fury lineup is, mostly, not the best 4K solution. Similarly, we do not need pages and pages of analysis to know that the 960, especially the 2GB version, is grossly overpriced. It was obvious on release and it is obvious now.

To end on a more positive note: I think original article you posts is fantastic. It would also be genuinely helpful to see tech sites write some guides as to how much VRAM usage you can expect from popular games with various settings / resolutions.

 

[ultima_trev]

I've seen cases where AMD cards handle lower memory constraints better than nVidia but this article certainly paints AMD as being less efficient with how their GPU handles memory.

I personally wouldn't get anything less than a R9 380X (and I am hoping that future drivers will make it perform better than it does right now as 20% above GTX 960 is meh) although I am certainly glad I stretched the budget for a 390.

 

[Headfoot]

It was immediately apparent that the 960 was a turd as soon as you saw "128-bit Bus." All the magic pixie dust compression can't cancel out the fact that nVidia has never in the history of the x60 line, shipped one with a 128 bit bus. The lowest they got was 192.

Small bus = small chip = cheap chip. It's a cheap low end laptop chip they peddled as a midrange chip. Its not a real x60 chip. And who can blame them for playing it off like it was? People bought tons of them for more dollars than they were worth and nVidia made good margins. Stockholders are happy and the world keeps on turning.

 

[Seba]

The memory bus by itself is meaningless as a comparison element between cards. You have to also consider the memory speed (determined by memory clock and memory type). Bus width correlated with memory speed gives you memory bandwidth and this is what you should compare.

Some values for memory bandwidth:

GTX 960 - 112 GB/s
GTX 970 - does not compute
GTX 980 - 224 GB/s
GTX 980 Ti - 336 GB/s

R9 380 - 182 GB/s
R9 390 - 384 GB/s
R9 Fury - 512 GB/s
R9 Fury X - 512 GB/s

 

[xorbe]

970 should be 196 GB/s (224-bit effective)

 

[Seba]

970 should be 196 GB/s (224-bit effective)

If you ignore the last 0.5GB, with only 28 GB/s.

 

[Joepublic2]

yes the 960 2GB has a lot less frame times problems than the 285/380 2GB, it's pretty impressive

Probably down to superior texture compression vs. the 285-380. The driver might be smarter (probably is knowing AMD's drivers) about which data to shuttle across the pcie bus at what time to minimize processing stalls. They both still pretty much suck if stutters annoy you, though.

 

[moonbogg]

Oh god the VRAM! Its a pain in the ass, truly, when you buy rebranded mid range cards marketed as high end cards and you run out of Vram 6 months later in your favorite game. Yeah, that sucks. Yeah, that was me with 670's and BF4. 2GB is so 2012 and that 960 is a piece of unholy garbage.

 

[Joepublic2]

If you ignore the last 0.5GB, with only 28 GB/s.

Still has a much, much lower latency than going over the PCIe bus for the same data. If the driver is smart it'll keep data there that's latency sensitive but doesn't require large amounts of bandwidth i.e. large shader programs.

 

[Headfoot]

The memory bus by itself is meaningless as a comparison element between cards.

No.

It is a meaningful comparison in terms of how much die space must be allocated to memory and thus the relative size of the die, therefore it is relevant to where it falls in a product lineup which is divided primarily by die size, memory size and heatsink design as cost drivers.

Using 128-bit w/ GDDR5 quite obviously puts the die into the "budget" space, it is smaller to in order to require less memory, less die size and be cheaper. If busses were free they'd all be 512bit or higher.

Thus, it should have been obvious that any die in this generation using 128-bit, lower than the lowest midrange chip of last generation at 192 bit, would be a low end part. And its turned out to be a low end part indeed.