Memory overclocking does nothing for the 7850

[SickBeast]

I haven't tested this extensively yet, but:

with my memory at 4400mhz, I get 29.2fps in Heaven at 1280x1024 maxed out with 8xAA

- with my memory at 5000mhz, I get 30.5fps at the same settings

- with my memory at 5200mhz, I get 31fps at the same settings

- with my memory at 5400mhz, I get 31.4fps at the same settings

- with my memory at 5600mhz, I get 31fps at the same settings

-----------------------------------------------------------------------------

Crysis:

5000mhz - 47.7fps
5200mhz - 48.2fps
5400mhz - 48.7fps
5600mhz - 49fps

-----------------------------------------------------------------------------

I will confirm this with more benchmarks when I can.

For now I am going to run at 5400mhz as it's faster and will probably keep the card a little cooler.

 

[tigersty1e]

gddr5 memory has error detection.

bench at 5000, then 5100, then 5200, etc.

fps will gradually move up until a certain point.

 

[SickBeast]

gddr5 memory has error detection.

bench at 5000, then 5100, then 5200, etc.

fps will gradually move up until a certain point.

Well the difference between 4400mhz and 5600mhz is 6%, meanwhile the memory is clocked 22% faster.

 

[tigersty1e]

bench your card at 4800, 4900, 5000, 5100, 5200, 5300, 5400, 5500, and 5600.

write down your fps and look at the trend.

 

[SpeedyGonzales]

I think you can not expect linear scalability from RAM. Even GPUs do not always do that under perfect conditions.

 

[tigersty1e]

The scenario laid out to us by AMD is that overclockers who have reached the limits of their card’s memory bus will now see the impact of this as a drop in performance due to retransmissions, rather than crashing or graphical corruption. This means assessing an overclock will require monitoring the performance of a card, along with continuing to look for traditional signs as those will still indicate problems in memory chips and the memory controller itself.

http://www.anandtech.com/show/2841/12

 

[SickBeast]

I updated the OP with more results.

 

[KingFatty]

When the memory speed is lowered, does that affect the temperature or overclockability of the GPU core at all?

 

[SickBeast]

I updated the OP with Crysis benchmarks.

 

[Tuna-Fish]

7850 has exactly the same B/W as the 7870, which has 45% more flops (at stock). When overclocked to the powertune limit, the 7850 still has (just a bit) lower computational throughput than the 7870.

Giving it more bandwidth really *shouldn't* do anything. It has plenty.

 

[SickBeast]

It will be interesting to see if these cards can be flashed to a 7870. So far either no one has the guts to do it or else it just doesn't work.

 

[Meaker10]

Cores will be laser cut.

 

[Stuka87]

Cores will be laser cut.

This.

As I recall AMD said there will be no more unlocking of GPU's. So a 7950 will not flash to a 7970 (and they don't) and a 7850 will not flash to a 7870.

 

[3DVagabond]

Well the difference between 4400mhz and 5600mhz is 6%, meanwhile the memory is clocked 22% faster.

This does show that memory O/C'ing is doing something. For it to increase "more or less" linearly your card would have to be completely bottle necked by the memory. What are your GPU clocks? Is this with your card @ ~1200MHz or stock?

When the memory speed is lowered, does that affect the temperature or overclockability of the GPU core at all?

Typically O/C'ing memory to the max will reduce the O/C'ing capability of the GPU. It's recommended to first O/C the core and then the memory. The core will typically make a bigger difference in performance than the memory.

This.

As I recall AMD said there will be no more unlocking of GPU's. So a 7950 will not flash to a 7970 (and they don't) and a 7850 will not flash to a 7870.

This could very well be true, for now. It doesn't mean that either AMD or the board partner can't castrate the chips with software in the future. With the kind of money they are getting for chips this gen they can likely afford to sell the XT chips for the Pro chip price and still make plenty of money.

It was really effective marketing with the 6000 series. When things get tough marketing usually reaches back into their bag of tricks and recycle something that worked. Rarely does anyone actually think of something new, and when they do, it was likely an accident that worked out well.

 

[SickBeast]

When the memory speed is lowered, does that affect the temperature or overclockability of the GPU core at all?

No, unfortunately.

 

[tuffluck]

personally, i see a pretty significant boost in FPS when i increase memory speed on my 7850. however, i haven't really heard of anyone going past 6000, so i don't either. that's not to say it wouldn't continue giving me added performance however.

 

[toyota]

personally, i see a pretty significant boost in FPS when i increase memory speed on my 7850. however, i haven't really heard of anyone going past 6000, so i don't either. that's not to say it wouldn't continue giving me added performance however.

do you have any data to back that up? you really have to do a consistent benchmark of some sort otherwise you are just guessing.

 

[Upgrade_Itch]

personally, i see a pretty significant boost in FPS when i increase memory speed on my 7850. however, i haven't really heard of anyone going past 6000, so i don't either. that's not to say it wouldn't continue giving me added performance however.

I've yet to see ANY videocard I've owned get much of a performance boost by increasing the memory frequency. Its all about the CORE, the overclocked memory just adds heat.

 

[toyota]

I've yet to see ANY videocard I've owned get much of a performance boost by increasing the memory frequency. Its all about the CORE, the overclocked memory just adds heat.

it depends on the card and the actual game being tested. the good example I have is with my 8600gt several years ago. I could oc the core by 20% and get 20% more performance in a game like FEAR. I could oc the memory by 20% and also get 20% more performance in that game even if I left the core stock. so that's a 40% increase from overclocking both in total in FEAR. so yeah the 8600gt was pretty starved for memory bandwidth in that case for the memory to scale by itself just like the core.

 

[Meaker10]

One thing to consider is memory clocks usually aide minimum fps.

 

[toyota]

One thing to consider is memory clocks usually aide minimum fps.

I am not sure how you came to that conclusion but I have never really noticed that while testing on any card.

 

[Jaydip]

Memory clocks generally help texture heavy games.

 

[aaksheytalwar]

1. VRAM doesn't scale linearly. A 20%+ increase should ideally get you 5-10% more performance consistently.

2. Your Crysis results seem pretty reasonable for memory I guess.

 

[toyota]

1. VRAM doesn't scale linearly. A 20%+ increase should ideally get you 5-10% more performance consistently.

2. Your Crysis results seem pretty reasonable for memory I guess.

again it varies from card to card and game to game. I already gave an example of nearly perfect memory scaling that I have seen.

 

[Kenmitch]

Memory only needs to be fast enough to keep the GPU fed. If the GPU doesnt have enough umph then xtra bandwidth is wasted.