Video Card Core Overclocking Scaling Analysis (AMD and Nvidia)

[guskline]

Wonderful read Termie.

 

[Termie]

That's what I meant, since these SKUs are bandwidth starved, starting with the vram OC as baseline makes their scaling % much lower..

Look at many reviews (and from forum goer results), the scaling above stock for core/vram combined OC is stellar.

FYI, I can give you the benchmarks for HD7870 at reference 1000/4800 speeds in a few games:

3DMark Fire Strike: 5400
Metro2033: 25
Just Cause 2: 68
Hitman: 38

Compare that to ~15% overclocks on Core/VRAM (1150/1400):
3DMark Fire Strike: 6020, 11.5% faster
Metro 2033: 29.3, 17% faster
Just Cause 2: 76.2, 12% faster
Hitman: 44.9, 18% faster

Average: 14.6%

Note that I concluded my review by saying a combined 15% overclock would yield 15% improvement, but let me be clear about one thing - my goal is to isolate the effects of core and VRAM overclocking separately, to inform overclockers of their differential effects on performance.

 

[wand3r3r]

It is interesting to see the effect of overclocking the two parts in isolation. Often times video cards do not oc proportionally, this could make it easier to decide whether it's paying more for e.g. a lightning or equivalent which has better memory / voltage control for the core etc.

I think all 3 aspects are important, core, memory, and combined. I don't think I've seen reviews (for sure not lately) focusing on them like this.

I would love to see the 7950 investigated too as it has great core overclocks and the wider bus so may not be so memory limited. Actually I'd love to see the whole range of cards put through the test.

Considering that the best way to overclock is in isolation (oc memory with the core at stock, then vice versa) as you mentioned, this type of investigation tells you which cards it pays to concentrate on squeaking out the last bit of performance on the core or memory, or both.

 

[Termie]

Ok, lots of calls for an HD7950 added to the test. I might need to pick one up the next time I see a good sale...

 

[Elfear]

Thanks for putting that together Termie. Interesting results. I hope you get a chance to test the 7950 as well.

 

[Silverforce11]

FYI, I can give you the benchmarks for HD7870 at reference 1000/4800 speeds in a few games:

3DMark Fire Strike: 5400
Metro2033: 25
Just Cause 2: 68
Hitman: 38

Compare that to ~15% overclocks on Core/VRAM (1150/1400):
3DMark Fire Strike: 6020, 11.5% faster
Metro 2033: 29.3, 17% faster
Just Cause 2: 76.2, 12% faster
Hitman: 44.9, 18% faster

Average: 14.6%

Note that I concluded my review by saying a combined 15% overclock would yield 15% improvement, but let me be clear about one thing - my goal is to isolate the effects of core and VRAM overclocking separately, to inform overclockers of their differential effects on performance.

These 256bit cards on 28nm are very bandwidth starved so to get a good OC, you need to boost vram as well, which you clearly show here. Almost perfect scaling for OC.

 

[Termie]

These 256bit cards on 28nm are very bandwidth starved so to get a good OC, you need to boost vram as well, which you clearly show here. Almost perfect scaling for OC.

Yes, this is very true. The HD7870 in particular seems to really gain from a big VRAM overclock. In the next generation, either the bus on midrange cards will need to be wider, or GDDR5 will have to be even faster than it is now.

 

[Janek1]

interesting