AMD Radeon 7950 Flashed to 7970 with Unigine Benchies [ocuk]

[Grooveriding]

Gibbo at OCUK has flashed a 7950 with a 7970 bios and has some benchmark results. Shaders did not unlock but the card overclocked like a monster the same way 7970s have been clocking. :biggrin:

http://forums.overclockers.co.uk/showthread.php?t=18366789

Hi there

First things first, this is all the benchmarks you are getting, so be grateful for this. I've done this to simply see how they clock and how the Heaven scores compete to the NVIDIA GTX 580 3GB and ATI 7970 3GB.

First of all the card tested is a stock version from one of the AIB's, the cooler is rather cut down and not that powerful, as the same AIB also has OC editions which come with a much better cooler. However even still with this stock average performing cooler the card managed the following:-

1. Using ATI OverDrive it maxed out - NDAMHz Core & NDAMHz memory.
2. Using Asus 7970 BIOS it managed - NDAMHz Core & NDAMHz memory.

Flashing the card was the same as flashing a 7970, very easy, unfortunately it DID NOT unlock any additional shaders, the card had the stock 1792 shaders still, though these cards do have a FULL 32 ROPS.

He since has edited out the clocks the 7950 attained, but they can be found here http://wccftech.com/amd-radeon-hd79...chmarks-shows-promising-overclocking-results/

AMD 7950 Maxed OC on Asus OC BIOS (1180MHz Core / 7200MHz Memory)


OCed 7950

Compared to an OCed 7970

For reference an overclocked 580 @ 900 core

This may be part of the 7950 delay, it's looking as good as the 7970 +/- 5% when you overclock it for $450.

 

[railven]

Hmmm, maybe I want to get an Eye6 HD 7950...

Curse you tax return and HD 7950 delay! I got a Corsair Carbide 500r waiting to be filled with new computer part goodness!

 

[Lepton87]

Looks like 5850/6950 again. The difference in shader count is 12.5%, if performance difference at the same clocks is really on the order of 5% then it looks like Tahiti is bottlenecked someplace else, probably by ROPs just like Evergreen and Cayman.

 

[Piano Man]

Looks like 5850/6950 again. The difference in shader count is 12.5%, if performance difference at the same clocks is really on the order of 5% then it looks like Tahiti is bottlenecked someplace else, probably by ROPs just like Evergreen and Cayman.

Yea, except this costs $200 more.

 

[LOL_Wut_Axel]

Shading power on these cards is abundant. That's why I said you shouldn't expect in even the most extreme cases a higher-than-10% difference clock-for-clock.

Now, if the 1.5GB version was under $400...

 

[Lepton87]

Well after some thought I came to the conclusion that this is not a good benchmark to compare those two cards, it uses unreasonable amounts of tessellation and 7970 and 7950 have exactly the same tessellation performance so in the real world the performance difference might be bigger.

 

[Silverforce11]

Absolutely no point paying that much extra for the 7970. Unless you don't OC, ever.

Edit: Lepton, Tessellation in those benchmarks are set at Normal, not Extreme. The new architecture is hardly bottlenecked by tessellation as seen in so many dx11 game benches. I agree its ROP starved. All that extra bandwidth and shading power and they didn't up the ROPs.

 

[LOL_Wut_Axel]

Yea, except this costs $200 more.

It's also ~40% faster than the HD 5870 and ~65% faster than the HD 5850, stock. It'll also either be similar or somewhat faster than the GTX 580 while costing less. Doesn't look bad when you see what it's competing against.

 

[Lepton87]

Absolutely no point paying that much extra for the 7970. Unless you don't OC, ever.

Edit: Lepton, Tessellation in those benchmarks are set at Normal, not Extreme. The new architecture is hardly bottlenecked by tessellation as seen in so many dx11 game benches. I agree its ROP starved. All that extra bandwidth and shading power and they didn't up the ROPs.

I didn't notice tessellation factor was set to normal, in that case ROP bottleneck seems even more likely. I don't get why AMD always skimp on ROP units... Its previous two architectures were bottlenecked by ROPs as evidenced by relatively small performance difference between 5870 and 6870 and huge performance deficit of 5830.

 

[WMD]

Well after some thought I came to the conclusion that this is not a good benchmark to compare those two cards, it uses unreasonable amounts of tessellation and 7970 and 7950 have exactly the same tessellation performance so in the real world the performance difference might be bigger.

Exactly my thoughts. I would like to see avp or stalker cop benchmarks. Those are much more shader intensive.

 

[Silverforce11]

Exactly my thoughts. I would like to see avp or stalker cop benchmarks. Those are much more shader intensive.

There's been benches in a wide variety of games, on cards OC to the same speed.

5850 vs 5870 = ~3-5% difference
6950 vs 6970 = ~3% difference

The shader counts differed by >10%.

Here we're seeing an identical trend.

 

[Tempered81]

Absolutely no point paying that much extra for the 7970. Unless you don't OC, ever.

More robust cooler, capable of dissipating higher levels of heat. Binned ASIC with more enabled compute units and texture units. More robust power circuitry on the PCB delivering better power to the GPU allowing for higher core clocks and higher memory clocks.

I believe it to be worth the extra $100 given we are talking about the top card. If it were 6850 versus 6870 or 560 vs 560ti or even 6950 versus 6970 (because of the shader unlock) I wouldn't say the same. But IMO in the case of the 7970 it's absolutely worth it. Just like a reference 580 PCB, ASIC, and HSF is more robust than a 570.

Cheers

 

[nanaki333]

Absolutely no point paying that much extra for the 7970. Unless you don't OC, ever.

i don't OC. ever. that's why i went with twin 6970 xfire and now a couple 7970 xfire. i just want my stuff to work without much tweaking. it's mostly due to laziness

 

[Piano Man]

It's also ~40% faster than the HD 5870 and ~65% faster than the HD 5850, stock. It'll also either be similar or somewhat faster than the GTX 580 while costing less. Doesn't look bad when you see what it's competing against.

Its debuting at 200 more than what the 5850 debuted at. Its the 2nd tier card that can be as fast as the top tier with OC'ing. Same with the 5850 back in the day and the 6950 as well. Those were 250-290 on debut.

 

[3DVagabond]

More robust cooler, capable of dissipating higher levels of heat. Binned ASIC with more enabled compute units and texture units. More robust power circuitry on the PCB delivering better power to the GPU allowing for higher core clocks and higher memory clocks.

I believe it to be worth the extra $100 given we are talking about the top card. If it were 6850 versus 6870 or 560 vs 560ti or even 6950 versus 6970 (because of the shader unlock) I wouldn't say the same. But IMO in the case of the 7970 it's absolutely worth it. Just like a reference 580 PCB, ASIC, and HSF is more robust than a 570.

Cheers

Excellent points! Unfortunately the typical poster only compares fps/$ and has no clue what the costs and value of the items you mention are. Add to that the **50's are (supposedly) salvaged chips (read as "of no value for their intended purpose") and the price difference is easily more than justified. In any other industry, seconds, which are what these chips are, are only sold for a fraction of the perfect spec parts.

 

[Riek]

I didn't notice tessellation factor was set to normal, in that case ROP bottleneck seems even more likely. I don't get why AMD always skimp on ROP units... Its previous two architectures were bottlenecked by ROPs as evidenced by relatively small performance difference between 5870 and 6870 and huge performance deficit of 5830.

ROP performance has been increased by so much i doubt it is an issue..

 

[AtenRa]

ROP performance has been increased by so much i doubt it is an issue..

I agree

 

[NumericalMethods]

ROP performance has been increased by so much i doubt it is an issue..

Yep. The fact that memory overclocks are yeilding such large returns in performance suggests that there is plenty of ROP performance and they are currently somewhat bandwidth bottlenecked in their performance.

 

[Grooveriding]

I really expected shaders to make more difference in GCN. The way these new 7XXX series scale perfectly with overclocking - % increase in core = % increase in FPS. That is something we didn't see with 5XXX or 6XXX series, figured it would play out differently.

 

[Joseph F]

I guess the 7970 firmware has a higher default voltage.
That would explain the nice OC on these.

 

[Concillian]

Looks like 5850/6950 again. The difference in shader count is 12.5%, if performance difference at the same clocks is really on the order of 5% then it looks like Tahiti is bottlenecked someplace else, probably by ROPs just like Evergreen and Cayman.

I think Tahiti is setup in modules that include 1 ROP / 4 TMU and a bunch of shaders, right? Disabling shaders would disable ROP and TMU too.

Does this suggest that memory bandwidth is an issue? That's kinda boggling given the bandwidth of the 7950 / 7970 is so much higher than anything we've seen previously, especially overclocked.

 

[Bobisuruncle54]

Yep. The fact that memory overclocks are yeilding such large returns in performance suggests that there is plenty of ROP performance and they are currently somewhat bandwidth bottlenecked in their performance.

So essentially the 7970 could use a memory bus wider than 384 bits, but 512bit would have probably been too expensive and eliminated the bottleneck to the point where some of the extra bandwidth would be wasted?

Could AMD use any faster GDDR5, does it exist?

 

[Arkadrel]

If the ROPs arnt cutting it, AMD needs to fix its ratio of ROPs to TMU and Shaders.
So theres more ROPs pr TMU/Shader than there is now.

Im kinda suprised at the bandwidth thingy too, is that really whats holding the cards back?

I mean the 7950 comes with 240GB/s memory bandwidth.

1180MHz Core /7200MHz Memory

Nuts.... thats what? 346 GB/s memory bandwidth?

I cant imagine that its a bottleneck at those speeds....
nor at the 240GB/s or so it comes with at stock.

I mean a 580 only has like 192GB/s memory bandwidth, and the 7950 isnt really *that* much faster than the 580, to warrent needing that much more memory bandwidth, is it?

 

[LOL_Wut_Axel]

Its debuting at 200 more than what the 5850 debuted at. Its the 2nd tier card that can be as fast as the top tier with OC'ing. Same with the 5850 back in the day and the 6950 as well. Those were 250-290 on debut.

So? AMD is not a charity, and they probably got tired of making low margins on Enthusiast cards. Like all companies, they need to price relative to what the competition has, and if you look at the HD 7950 the pricing is better than the GTX 580. When Kepler finally arrives, those prices will probably fall.

That's not to say it's not expensive, but looking at the market it IS justified.

 

[Lepton87]

I think Tahiti is setup in modules that include 1 ROP / 4 TMU and a bunch of shaders, right? Disabling shaders would disable ROP and TMU too.

Does this suggest that memory bandwidth is an issue? That's kinda boggling given the bandwidth of the 7950 / 7970 is so much higher than anything we've seen previously, especially overclocked.

You're partially right, disabling SPs also disables TMUs but ROPs are now fully decoupled even from memory controller.