GALAXY GTX 660 Ti GC OC vs. OC GTX 670 & HD 7950

[Final8ty]

http://www.hardocp.com/article/2012/...vs_670_hd_7950

[raghu78]

http://www.hardocp.com/article/2012/...vs_670_hd_7950

the Batman AC 1920 x 1200 MSAA 4X results prove how badly the GTX 660 Ti is going to fare in bandwidth constrained cases . People who still compare the GTX 660 Ti with the HD 7950 are just clueless.

[BababooeyHTJ]

I'm surprised that both of those cards are able to hit higher clocks than any of the Kepler cards that I have owned. That said, that review seems to be a pretty good representation of how the cards perform at settings that most of us would actually use.

the Batman AC 1920 x 1200 MSAA 4X results prove how badly the GTX 660 Ti is going to fare in bandwidth constrained cases . People who still compare the GTX 660 Ti with the HD 7950 are just clueless.

That drop seems pretty excessive to me. Makes me wonder if crc checks are kicking in with that extreme memory overclock.

[raghu78]

I'm surprised that both of those cards are able to hit higher clocks than any of the Kepler cards that I have owned. That said, that review seems to be a pretty good representation of how the cards perform at settings that most of us would actually use.

That drop seems pretty excessive to me. Makes me wonder if crc checks are kicking in with that extreme memory overclock.

http://www.hardocp.com/article/2012/...cking_review/2

No you can check the Galaxy GTX 660 Ti overclocking review. there is a consistent perf increase with the memory overclocking alone applied first. So its bandwidth and ROP decrease which are hurting MSAA performance.

[SKYMTL]

That drop seems pretty excessive to me. Makes me wonder if crc checks are kicking in with that extreme memory overclock.

I believe so. I have seen this issue in Batman, Shogun and Wargame. Seems some titles / settings will cause error checking to kick in while others won't. That graph clearly illustrates it....even though the reviewer chalked it up to something completely different. MVSS in one scene vs tessellation in another? Honestly, all scenes in Batman: AC have a crazy amount of dynamic shadows so the MVSS levels are pretty even throughout.

[raghu78]

I believe so. I have seen this issue in Batman, Shogun and Wargame. Seems some titles / settings will cause error checking to kick in while others won't. That graph clearly illustrates it....even though the reviewer chalked it up to something completely different. MVSS in one scene vs tessellation in another? Honestly, all scenes in Batman: AC have a crazy amount of dynamic shadows so the MVSS levels are pretty even throughout.

http://www.hardwarecanucks.com/forum...-review-5.html

your own review shows the GTX 660 Ti faster than HD 7970 Ghz with FXAA at 1080p and on par with the HD 7950 with MSAA. the GTX 670 performs very well. with MSAA iton par with HD 7970 Ghz at 1080p. This is completely at stock clock. no error correction. its just a bandwidth / ROP issue. MSAA needs ROP and bandwidth performance. GTX 660 falls flat on its face with MSAA in Batman AC. there are other titles like Merro 2033 where the case is even worse . the GTX 660 Ti falls behind GTX 580 with DOF and AAA because DOF uses a compute shader which is very bandwidth intensive.

http://www.guru3d.com/article/palit-...ream-review/15

[BababooeyHTJ]

There is only a 25% difference in memory bandwith between the 660ti and 670 and that is only at the same memory clocks. The 660ti in this review has is running a much higher memory clock than the 670. There is nowhere near enough difference in memory bandwith to explain that much of a drop in performance. It has to be something else. Thats a pretty excessive memory clock, I don't know how much testing that they've done at that clock speed.

We're talking about a 50% difference in minimums and a 30% difference in average which according to that graph evens up toward the end. Something is up.

[raghu78]

There is only a 25% difference in memory bandwith between the 660ti and 670 and that is only at the same memory clocks. The 660ti in this review has is running a much higher memory clock than the 670. There is nowhere near enough difference in memory bandwith to explain that much of a drop in performance. It has to be something else. Thats a pretty excessive memory clock, I don't know how much testing that they've done at that clock speed.

We're talking about a 50% difference in minimums and a 30% difference in average which according to that graph evens up toward the end. Something is up.

The GTX 660 Ti has a 25% reduction in ROP. Don't forget about that. Its 24 ROP in GTX 660 Ti vs 32 in GTX 670. MSAA requires ROP performance. The ROP reduction combined with bandwidth hits MSAA badly.

[SKYMTL]

http://www.hardwarecanucks.com/forum...-review-5.html

your own review shows the GTX 660 Ti faster than HD 7970 Ghz with FXAA at 1080p and on par with the HD 7950 with MSAA. the GTX 670 performs very well. with MSAA iton par with HD 7970 Ghz at 1080p. This is completely at stock clock. no error correction. its just a bandwidth / ROP issue. MSAA needs ROP and bandwidth performance. GTX 660 falls flat on its face with MSAA in Batman AC. there are other titles like Merro 2033 where the case is even worse . the GTX 660 Ti falls behind GTX 580 with DOF and AAA because DOF uses a compute shader which is very bandwidth intensive.

http://www.guru3d.com/article/palit-...ream-review/15

I wasn't referring to overall performance, just that drop. As I said, when GDDR5 error checking / correction enables, performance can be castrated very quickly and sometimes only for a limited time. Back in the post-GDDR5 days, memory would start displaying artifacts when pushed too far. Due to this feature in GDDR5, the only way to see if you are pushing memory too far is to check performance over time.

[Carbon_Unit]

I believe so. I have seen this issue in Batman, Shogun and Wargame. Seems some titles / settings will cause error checking to kick in while others won't.

That's something I had not considered, although it seems obvious when I think about it. Thanks for the info.

[[XC] Oj101]

raghu78 lay off it dude. You seem to have a vendetta against SKYMTL.

[BababooeyHTJ]

I wasn't referring to overall performance, just that drop. As I said, when GDDR5 error checking / correction enables, performance can be castrated very quickly and sometimes only for a limited time. Back in the post-GDDR5 days, memory would start displaying artifacts when pushed too far. Due to this feature in GDDR5, the only way to see if you are pushing memory too far is to check performance over time.

Yup, and if its kicking in there you have to wonder where else in the test it may have.

[Frontl1ne]

If the GTX 660 Ti is so bandwidth limited, why does it perform better than the 7950 in Batman at 2560x1600 4xAA ?

[BababooeyHTJ]

If the GTX 660 Ti is so bandwidth limited, why does it perform better than the 7950 in Batman at 2560x1600 4xAA ?

because Nvidia cards tend to perform better in that game. It doesn't look like memory bandwith is an issue in that game, at those settings. Look at how closely it compares to the 670. That may change with AA.

[Lanek]

There is only a 25% difference in memory bandwith between the 660ti and 670 and that is only at the same memory clocks. The 660ti in this review has is running a much higher memory clock than the 670. There is nowhere near enough difference in memory bandwith to explain that much of a drop in performance. It has to be something else. Thats a pretty excessive memory clock, I don't know how much testing that they've done at that clock speed.

We're talking about a 50% difference in minimums and a 30% difference in average which according to that graph evens up toward the end. Something is up.

If we follow Anandtech, who try to see what happend with the bus, i think it is not only crc, but more how the bus is working.. asymetric bus have never been a good idea . The card is good but not if you pass over 1080P or use high amount of MSAA.

[DilTech]

If we follow Anandtech, who try to see what happend with the bus, i think it is not only crc, but more how the bus is working.. asymetric bus have never been a good idea . The card is good but not if you pass over 1080P or use high amount of MSAA.

That galaxy card [H] used is the 3gb 660ti, so all memory buses have an identical amount of ram attached to them. As such, that isn't the problem here at all.

[raghu78]

If the GTX 660 Ti is so bandwidth limited, why does it perform better than the 7950 in Batman at 2560x1600 4xAA ?

According to the review GTX 660 Ti is faster than HD 7950 at 2560 x 1600 FXAA. FXAA is a shader based post process AA technique and is not bandwidth intensive at all. At 1920x 1200 4x MSAA in Batman AC the GTX 660 Ti falls behind badly. MSAA is extremely bandwidth intensive and thats one area where the GTX 660 Ti is clearly short.

[behrouz]

just look here :

660Ti :
Gpu Frequency : 1298 MHz
Memory : 1902MHz ( 7.6 GHz )
7950 :
Gpu Frequency : 1195 MHz
Memory : 1700MHz ( 6.8 GHz )

660Ti is faster than 7950 (FXAA mode) but for MSAA mode despite higher clock and bandwidth than AMD's card , 7950 is faster 660Ti ? it's weird.if MSAA needs bandwidth , then 660Ti should pass 7950.

[raghu78]

just look here :

660Ti :
Gpu Frequency : 1298 MHz
Memory : 1902MHz ( 7.6 GHz )
7950 :
Gpu Frequency : 1195 MHz
Memory : 1700MHz ( 6.8 GHz )

660Ti is faster than 7950 (FXAA mode) but for MSAA mode despite higher clock and bandwidth than AMD's card , 7950 is faster 660Ti ? it's weird.if MSAA needs bandwidth , then 660Ti should pass 7950.

Bandwidth = memory bus width x memory speed.

GTX 660 Ti has 192 bit memory bus while HD 7950 has 384 bit memory bus. so GTX 660 ti has much lesser bandwidth even if its memory runs at higher speed.

[LightSpeed]

Well the rest of the games show what I thought as well, 7950 overclocked would surely be faster than the 660Ti and seems to be even faster than an overclocked 670. Batman always favoured Nvidia GPU's anyway and the 7950 still manages to somewhat tie the 660Ti in that game (considering 670 fps in the case of no error). With the current prices, its a great deal.