Any hd7870 AND hd7970 card owners want to help w/ experiment?

[tviceman]

1050/1500:

1050/1000:

Hmmm very much in line with what I was expecting. PM incoming!

 

[james1701]

980X and 7970's here, but I dont have the game.

 

[footballrunner800]

1050/1750

1050/1600 -

 

[tviceman]

Very interesting! The performance continues to scale up, supporting my theory that the extreme AA setting is very bandwidth heavy. Now to find a few gtx680 users....

 

[Bryf50]

Alright here's my 7870 results.

1000/4800

1000/5400

1000/5800

 

[Termie]

Alright here's my 7870 results.

1000/4800

1000/5400

1000/5800

Color me impressed. You're getting 28% scaling for your memory overclock, which is exactly the same as the HD7970's memory overclock scaling above. I think tviceman's theory has some backing here, because the memory overclock scaling for the HD7000 series is typically closer to 20%.

 

[BoFox]

Also, I should note that both my 670 and 7870 exhibit 22% scaling with memory overclocking. It seems from the 7970 benches above that the 7970 scales at 30%, but that is an incorrect conclusion. That's with extreme bandwidth constraint. If you instead overclock the memory, you'll find that its memory scaling is actually lower than the the 7870 or 670/680, because it is fundamentally less bandwidth constrained due to its wider bus. I've seen data from other users that demonstrated approximately 18-20% memory scaling on the HD7900 series.

Honestly if you want to run this test, we need to be a bit more scientific about the approach. I'm really not sure what the OP's hypothesis was and whether the data tend to support it or refute it.

Yeah, that's just for Sleeping Dogs - which is why it was just a "shot in the barrel talk".. Of course, it's much less than 30% overall - and I'd go further to say that it's less than 20% for the 384-bit bus, overall.

How many games did you test your cards to get the "22%" figure for both your 670 and 7870? I'd say that the 680 has maybe 1-2% more scaling, due to having more GPU power while not having any more bandwidth.

 

[BoFox]

I want to see how the hd7970 and hd7870 perform against a gtx670/680 in this game with the highest anti-aliasing settings when all the cards have similar memory bandwidth. Some people argue that GK104's compute abilities suck, but whether or not true I don't think that is the case of what is causing the larger performance drop with Kepler (vs. Tahiti) when maxing out the AA options in this game. I think it's mostly memory bandwidth related (a noticeable area of weakness that pretty much the entire Kepler lineup suffers from).


Soooo, for the sake of curiosity and discussion of memory bandwidth / compute functionality and this game, could you run the game's benchmark with your hd7870 at stock core clocks and maxed out memory overclocks? It appears your 7870 can get to at least 1400mhz vram, which is 179.2 gb/s bandwidth, 93.3% the bandwidth of an hd7970 @ 1000mhz vram or a gtx670/680 @ 1500mhz.

One "shortcoming" that we should not overlook is that when overclocking GDDR5 memory to the MAX, there are roughly 2 intervals where the timings seem to be relaxed (which is why there is a sudden drop in performance when going past a certain MHz threshold at very different points).
(see Termie's awesome thread on this: http://forums.anandtech.com/showthread.php?t=2301435 )
So it helps to keep that in mind.

 

[boxleitnerb]

Geforce GTX Titan:
The GPU clock was fixed, no variations

967/3004


967/2005


As we lose only 10% fps by having 33% less memory bandwidth, I think it is safe to say that we're not bandwidth bottlenecked at stock speeds.

 

[BoFox]

The term "bandwidth bottlenecked" is a loosely defined term by many. For some engineers, even as low as 5% is considered a bottleneck. For some, the word "bottleneck" is so taboo that anything less than say, 50+% is considered blasphemy. Hehe!

10% performance for 33% bandwidth is almost exactly:

3% perf. for every 10% bandwidth, or simply 30% ratio.

It's not quite a baby-bottle bottleneck, but more like a Budweiser glass bottleneck. Cheers with your Titan!

 

[boxleitnerb]

Well we don't know how the performance scales with the bandwidth. Could be very good from 1000 to 1200 MHz, mediocre from 1200 to 1350 and not at all from 1350 onwards. But I'm too lazy to run the benchmark 10 times

 

[tviceman]

Geforce GTX Titan:
As we lose only 10% fps by having 33% less memory bandwidth, I think it is safe to say that we're not bandwidth bottlenecked at stock speeds.

Not entirely, at least not at 1080p anyways. But even the hd7970GE running at stock speeds showed a small increase (4.7%) in performance bumping the vram up to 7000mhz (16.7% increase in bandwidth). I'm willing to bet that when you get your backplate and (if you are willing) overclock your vram you'll see better performance scaling than what the hd7970GE exhibited in this benchmark. And if not at 1080p, definitely at 1440p. But that is a whole other slew of benchmarks and I figured 1080p was the easiest way to get able participants.

I am going run a few benchmarks with my gtx670 @ 1200mhz (according to my calculations, 1344 cores @ 1200mhz is the same shading power as a gtx680 1536 cores @ 1050mhz) with various ram speed settings and post results. I'm guessing at 1200mhz (simulating a 680) my gtx670 will show pretty darn good performance scaling with the vram speed moving up and down.

 

[blackened23]

MSI lightning GTX 680 at various speeds including stock 680 speeds, SLI DISABLED IN ALL TESTS :

Test 1: Reference 680 speeds, 1045mhz core, 6000mhz VRAM, SLI disabled

Test 2: Stock lightning core (1202mhz), 6000mhzVRAM, no SLI

Test 3: Reference 680 core (1045mhz), 7050mhz VRAM OC, no SLI

Test 4: Stock MSI Lightning speed with LN2 BIOS, no SLI, 1202mhz, 7000mhz VRAM

 

[tviceman]

Interesting scores there. 10% scaling with core OC'd and vram at stock. Vram overclock alone produced poor results, but overclocking both the core and vram gave you >20% increase. Very interesting! It appears at stock speeds, the gtx680 isn't bandwidth limited in this game at 1080p, but when the core is overclocked you got perfect scaling when also overclocking the vram as well.

Now I'm going to ask the few people who have participated if they can run their tests at 1440p (or 1600p) if possible.

 

[Termie]

Interesting scores there. 10% scaling with core OC'd and vram at stock. Vram overclock alone produced poor results, but overclocking both the core and vram gave you >20% increase. Very interesting! It appears at stock speeds, the gtx680 isn't bandwidth limited in this game at 1080p, but when the core is overclocked you got perfect scaling when also overclocking the vram as well.

Now I'm going to ask the few people who have participated if they can run their tests at 1440p (or 1600p) if possible.

The 680 results are shocking. At stock core speed, his vram scaling is 30%, just like the HD7970 and HD7870. With the oc'd core, though, his vram scaling is astronomical at 71%.

And the overall scaling from stock to fully overclocked is a higher than theoretical 22%. The overclock is only 15% on the core and 17% on the memory. The only way for that to happen is if at stock, the core and memory were bottlenecking different sections of the benchmark.

Fascinating. I think tviceman was on to something with this benchmark as to its bandwidth load, but something unexpected has been found in GTX680 performance.

 

[tviceman]

I want to know how it's possible that scaling is greater than 100% with both core and memory overclock, but with just the memory overclock scaling is rather poor. So weird.

 

[Dravonic]

Since I can't PM you before I get 25 posts, I'll just answer this here:

What is the native resolution of your monitor? Do you have a 1440p or 1600p screen?

Nope, unfortunately. 1080p is all I can do.

 

[UaVaj]

great review from both sides.

goes to solidify. when AA are on and resoultion are hugh. memory bandwidth and vram are crucial. granted the gpu can continue to push.