Which graphics card for my system - 770 v 7970

[ubersonic]

If you think it's a delusion then you clearly don't understand how graphics cards work.

I do, I'm just combining that understanding with experience and history in order to form a reasoned prediction.

You also don't seem to know what delusion means either.

I do hence my correct usage of the word but for ref: "An idiosyncratic belief or impression that is firmly maintained despite being contradicted by what is generally accepted as reality" see.

The bottom line is that in the future, you will have to turn down more settings for a 6XX card than you would on a 79XX to get a playable frame rate, this is how the situation is now, so on what planet is it a delusion?

No you will not, you seem to think that just because the scenario exists now at insane resolutions that it will translate to standard resolutions down the line but this is not how it works it's how you would expect it to work, and that is the delusion. In reality however, history has shown that whenever you have GPU's with similar GPU performance that by the time you would have to turn down the effects due to memory throughput you will have already had to turn them down due to the GPU running out of grunt.

(NB: that's historically correct for GPU's of similar power when one has decent memory throughput and the other better, not when one of them has crippled throughput a la the GDDR3 version of a GDDR5 card.)

 

[ubersonic]

If the 256bit and the 384 bit have their memory running at the same speed, so that their total bandwidth both = 250gb for example, would the wider bus of the 384bit card still give it a performance advantage or would that not mean anything now as the actual bandwidth is the same?

It wouldn't mean anything anymore, that's why AMD/Nvidia give the GB/s values for the cards as most people cannot work it out form the bus/MHz.

For ref the calc is: bus width / 8 * memory clock * 2 = MB/s

I.E

384 / 8 * 2500 * 2 = 240,000
256 / 8 * 3500 * 2 = 224,000

 

[bru]

I wouldn't bother ubersonic.

In my opinion some people seem to see the world their way only.

No amount of reasoned discussion will get them to change their mind.

It is obsessive, almost extremist in nature.


And yes there are probably speeeeling mistakes in this post as well, no need to correct them

 

[oweneades]

If the 256bit and the 384 bit have their memory running at the same speed, so that their total bandwidth both = 250gb for example, would the wider bus of the 384bit card still give it a performance advantage or would that not mean anything now as the actual bandwidth is the same?

At a high level it would be reasonable to assume that performance would be largely similar (assuming identical workloads).

However, deep down at the technical level there are likely a multitude of factors that come into play that could affect performance in different ways. For example; how the GPU/card handles the throughput through the memory controllers.
The drivers will also have a heavy involvement in this as well so just because the hardware has x, y or z doesn't mean it can use it - best example being the T/Flops throughout touted for the high end GPU's.

 

[spoffle]

I do, I'm just combining that understanding with experience and history in order to form a reasoned prediction.

You've already shown you don't have an understanding.

I do hence my correct usage of the word but for ref: "An idiosyncratic belief or impression that is firmly maintained despite being contradicted by what is generally accepted as reality" see.

You're just making statements and claiming they are fact. You haven't actually contradicted anything other than "no you're wrong".

That's not a contradiction, you're disagreeing about something you don't understand.

No you will not, you seem to think that just because the scenario exists now at insane resolutions that it will translate to standard resolutions down the line but this is not how it works it's how you would expect it to work, and that is the delusion. In reality however, history has shown that whenever you have GPU's with similar GPU performance that by the time you would have to turn down the effects due to memory throughput you will have already had to turn them down due to the GPU running out of grunt.

The issue exists now at mainstream resolutions. As per the above, you're arguing authoritatively about something you clearly don't understand.

The GTX670 and 680 currently have a greater fall off of FPS when enabling anti-aliasing at 1920x1080 than cards with greater memory bandwidth, ie, 79XX cards.

Which means if you theoretically could overclock a 680's or 670's VRAM to 8Ghz, giving a memory bandwidth of 256BG/s, then they would both have a lesser FPS falloff than a 7950 at stock clocks with its 240GB/s memory bandwidth.

That aside, the point you clearly don't seem to get is that I am not talking about maxed out, yet you keep insisting I am and then claiming I'm deluded?

The point is that you will have to turn settings down MORE. It's that simple.

So stop sperging about people being delusional and quit with the strawman arguments.

(NB: that's historically correct for GPU's of similar power when one has decent memory throughput and the other better, not when one of them has crippled throughput a la the GDDR3 version of a GDDR5 card.)

Your example of the GTX580 shows you don't understand this topic.

 

[spoffle]

I wouldn't bother ubersonic.

In my opinion some people seem to see the world their way only.

No amount of reasoned discussion will get them to change their mind.

It is obsessive, almost extremist in nature.


And yes there are probably speeeeling mistakes in this post as well, no need to correct them

So you're ignoring his posts sperging about stuff with authority that he doesn't actually understand? If you think that's fine, it's your issue not mine.

 

[oweneades]

The GTX670 and 680 currently have a greater fall off of FPS when enabling anti-aliasing at 1920x1080 than cards with greater memory bandwidth, ie, 79XX cards.

Genuine question - Has any major site (or any site) actually tested this? It would be quite interesting to see the extent to which it becomes apparant.

EDIT: I know theoretically we could work it out from general reviews, I am just wondering if it has been specifically tested.

I know from my own testing in Portrait Eyefinity/Surround that the GK104 GPU's (670 & 680 tested) drop back from the 7970/780 quite noticably (at least since the 12.11 drivers) compared to their performance delta at 1080p.

More noticable is a recent test I did of BF at 3600x1920 ultra Settings with 0x MSAA, my 780 overclocked is basically as fast as 670SLI* (I don't have the results for the 670SLI with 4xMSAA which is annoying).

*Asus Direct CUII 670's @ 1050-1100/6000

I wish I still had the 670's to do some further cross comparisons with the 780 (I no longer have the 7970 either).

 

[spoffle]

I don't know of any sites that have tested for it specifically, however Rusty0611 did some tests comparing 2x 7950s to 2x680s that demonstrated this, and there are a few sites that have shown increasing AA amounts at the same res giving a sharper falloff on the 670 and 680 than on the 7950 and 7970.

Your 780 benchmarks are pretty telling too! As no way does a 780 have twice the actual power of a 670. This is something I commented on when it became apparent that the 770 would still be using a 256bit bus, as it's going to show an unusually large disparity between the 780 and 770 as time goes on.

I'm curious as to when you did your tests for the 670s and 7970s as it looks like they were before the 12.11 drivers.

 

[oweneades]

I don't know of any sites that have tested for it specifically, however Rusty0611 did some tests comparing 2x 7950s to 2x680s that demonstrated this, and there are a few sites that have shown increasing AA amounts at the same res giving a sharper falloff on the 670 and 680 than on the 7950 and 7970.

Your 780 benchmarks are pretty telling too! As no way does a 780 have twice the actual power of a 670. This is something I commented on when it became apparent that the 770 would still be using a 256bit bus, as it's going to show an unusually large disparity between the 780 and 770 as time goes on.

I'm curious as to when you did your tests for the 670s and 7970s as it looks like they were before the 12.11 drivers.

Ah yes I remember Rusty's comparison now. I recall that the 7950's (overclocked) were noticably faster at 5760x1080.

7970Ghz were done on Cat 13.2 Beta 6 (I couldn't use anything newer due to DP tearing)

670SLI & 680 were done at launch (or not long after). Driver improvements would make a small difference but not anything major (certainly nothing like the 12.11 driver from AMD).

CPU for all tests bar the 780 (as shown) was a 2500K @ 4.4Ghz.

Average wise the 680 OC (1280/6800 IIRC - it was a while ago) puts in a reasonable show, but its minimums are rubbish compared to the 7970Ghz OC (1250/6400).

TBH I was suprised when I threw the 670 SLI results into the graph initially. I mean a 780 does have 77% more SP and 50% more memory bandwidth (and 50% more ROPS) - but to end up basically 100% faster* leads to the obvious conclusion that the GK104 cards are being held back somewhat.

*780 overclocked.

 

[spoffle]

Oh wait there, is that only a single 7970 in your results?

 

[oweneades]

Oh wait there, is that only a single 7970 in your results?

Yes, a VTX 7970 running 1050/6000 (listed as Ghz) and overclocked to 1250/6400 (Ghz OC).

All of the results are single card bar the 670 SLI.

Edit: For clarification the 6970 right at the bottom is a nice reference point but can be ignored (in case this is causing confusion).

 

[imginy]

Genuine question - Has any major site (or any site) actually tested this? It would be quite interesting to see the extent to which it becomes apparant.

There is a piece here but it's tested on old console ports that chew on your memory bandwidth, most games wouldn't be so dramatic.

http://www.tomshardware.com/reviews/geforce-gtx-660-ti-memory-bandwidth-anti-aliasing,3283-13.html

 

[spoffle]

*facepalm* I saw SLi 670s and assumed it was 7970 crossfire, that's why I asked about whether it was on 12.11 drivers with the minimums being so close.

 

[oweneades]

*facepalm* I saw SLi 670s and assumed it was 7970 crossfire, that's why I asked about whether it was on 12.11 drivers with the minimums being so close.

I would love it if my single 780 was faster than 7970 CF! Money well spent then....

 

[spoffle]

There is a piece here but it's tested on old console ports that chew on your memory bandwidth, most games wouldn't be so dramatic.

http://www.tomshardware.com/reviews/geforce-gtx-660-ti-memory-bandwidth-anti-aliasing,3283-13.html

There isn't really a matchup there, firstly as "console ports" typically don't exist, what you mean is poorly optimised games, and secondly memory bandwidth is chewed up by things such as AA and resolution over poor optimisations.

Some games simply use more memory bandwidth than others due to what's going on on screen and so on.

 

[spoffle]

I would love it if my single 780 was faster than 7970 CF! Money well spent then....

I'm surprised at how close the 7970 is to the 780 in an nVidia sponsored game too.

 

[LtMatt]

Good results and thanks for posting owen always interesting to see.

 

[teppic]

Aside from the memory issues, all the new engines are going to be very well optimised for AMD hardware so it'll be interesting to see how things look in 6-12 months on these cards.

 

[Marine-RX179]

If the 256bit and the 384 bit have their memory running at the same speed, so that their total bandwidth both = 250gb for example, would the wider bus of the 384bit card still give it a performance advantage or would that not mean anything now as the actual bandwidth is the same?

Yes. However, the issue at the moment is that the 770's memory bandwidth cannot reach 7970's memory bandwidth level, even if the memory was overclocked from 1753 (7000MHz) to 2000MHz (8000MHz).

GTX770 (memory at stock clock):
256/8 *1753 *4 = 224,384 (224GB/sec)

GTX770 (memory overclocked to 2000MHz/2050MHz):
256/8 * 2000 *4 = 256,000 (256GB/sec)
256/8 * 2050 *4 = 262,400,000 (262GB/sec)

HD7970 (memory at stock clock)
384/8 * 1375 *4 = 264,000 (264GB/sec)

HD7970 GHz Edition (memory at stock clock)
384/8 * 1500 *4 = 288,000 (288GB/sec)

HD7970 (memory overclocked to 1700MHz)
384/8 * 1700 *4 = 326,400 (326GB/sec)

HD7970 (memory overclocked to 1800MHz)
384/8 * 1800 *4 = 345,600 (345GB/sec)


There's no going around the fact that the 256-bit is clearly holding back the memory bandwidth.

 

[spoffle]

Good results and thanks for posting owen always interesting to see.

Seconded!

Yes. However, the issue at the moment is that the 770's memory bandwidth cannot reach 7970's memory bandwidth level, even if the memory was overclocked from 1753 (7000MHz) to 2000MHz (8000MHz).

GTX770 (memory at stock clock):
256/8 *1753 *4 = 224,384 (224GB/sec)

GTX770 (memory overclocked to 2000MHz):
256/8 * 2000 *4 = 256,000 (256GB/sec)

HD7970 (memory at stock clock)
384/8 * 1375 *4 = 264,000 (264GB/sec)

HD7970 GHz Edition (memory at stock clock)
384/8 * 1500 *4 = 288,000 (288GB/sec)

HD7970 (memory overclocked to 1700MHz)
384/8 * 1700 *4 = 326,400 (326GB/sec)


There's no going around the fact that the 256-bit is clearly holding back the memory bandwidth.

All this aside, I really can't get my head around why so many people are rabidly recommending a card with lesser specs for more money.

It's mental.