ATI HD4800 GPU/MEM clock performance scaling

[Cooper]

GPU clocks 600-800 MHz
Memory clocks 2000-4000 MHz
8 games
Plenty of graphs

Source
Google translator

[Monkeywoman]

this makes overclocking a 4850 more worth while. BTW i like your :banana::banana::banana::banana::banana: cooper

[Cooper]

BTW article also contains system's power consumption tests with different gpu/mem clock speeds combinations and performance comparison of heavily overclocked 4850 Vs 4870 on default clocks.
Huge work done here. Kudos to author

[Hornet331]

hmm so stock 4870 is still a good margin faster then a heavily oced 4850. Good to know.

[lm358]

Excellent comparison. It looks like the extra bandwidth of the 4870 certainly helps, but obviously comes at a cost... let's see what happens to prices after the 4870x2 launches.

[cegras]

hmm so stock 4870 is still a good margin faster then a heavily oced 4850. Good to know.

Surprisingly to see how starved for bandwidth both the RV670 and RV770 cores are.

[Tonucci]

Good work

Interesting... Very bandwidth starved core. Once you add more bandwidth, the cards bottleneck is removed allowing the GPU to really fly. this causes an big increase in power consumption aswell.

800/4000 = 174W

800/2000 = 121W

[OCam]

Is it just me or was there no mention of the extra $$ you will need to throw into your 4850 to get it to reach 800/2500. And most 4870's seem to be able to get to 830-50 stock volts/ with stock cooler.

just my 2 cents.

[JohnJohn]

Incredible how at even 600 2000 the rv770 is 5-20 frames botlenecked by memory compared to 600 3500. It seems that 3600 is the sweet spot to be, because further than that there is no significant gain overall

[eXa]

Wow, im impressed. Why dont we see more test like this!

I bet 4000 and more on memory is beneficial as soon as u start Vmod the card to get 900mhz+ on it.

[lm358]

However, one thing that I'm not sure has been considered is memory latency. I think they have only used a 4870 with GDDR5 memory. This probably has much higher latencies than GDDR3 at 2000 or whatever which could make the results worse than expected when running GDDR5 at 2000.

[demonkevy666]

However, one thing that I'm not sure has been considered is memory latency. I think they have only used a 4870 with GDDR5 memory. This probably has much higher latencies than GDDR3 at 2000 or whatever which could make the results worse than expected when running GDDR5 at 2000.

it's barely noticeable, still say the same about gddr4.

looks like they need to get some samsung 6ghz gddr5 & maybe 2gbs of it on each card just for good measure.

[largon]

I fail too see the point in comparing a HD4870 with underclocked memory to anything.
HD4850 at a given freq is way faster than HD4870 at the same clock (eg. 800/2000).

[Boissez]

I fail too see the point in comparing a HD4870 with underclocked memory to anything.
HD4850 at a given freq is way faster than HD4870 at the same clock (eg. 800/2000).

Yup, seems that GDDR5 is about 20% slower than GDDR3 clock for clock.

[MAS]

Excellent comparison. It looks like the extra bandwidth of the 4870 certainly helps

it helps mostly in high-res

[strange|ife]

"This map is the indisputable advantage over younger brother in the face of video DDR5 and has not yet disclosed upper potential nucleus, in the absence of specialized tools to break this accelerator. "- from end of article

lol! this guy is serious!

great review though, lot of effort and detail.

[Cooper]

"This map is the indisputable advantage over younger brother in the face of video DDR5 and has not yet disclosed upper potential nucleus, in the absence of specialized tools to break this accelerator. "- from end of article

lol! this guy is serious!

great review though, lot of effort and detail.

it's google translator. Original article is in Russian.

[STaRGaZeR]

I fail too see the point in comparing a HD4870 with underclocked memory to anything.
HD4850 at a given freq is way faster than HD4870 at the same clock (eg. 800/2000).

Latency. But we don't game at 500MHz GDDR5, do we?

[trans am]

"This map is the indisputable advantage over younger brother in the face of video DDR5 and has not yet disclosed upper potential nucleus, in the absence of specialized tools to break this accelerator. "- from end of article

lol! this guy is serious!

great review though, lot of effort and detail.

Dude is that the guy from blood sport? this avatar made me piss my pants! wow! its a good one! sorry ot.

[Helmore]

Latency. But we don't game at 500MHz GDDR5, do we?

Can it even go that low on all boards? 2000 MHz. effective is a 1000 MHz. data clock...
Although you were probably referring to the command clock, which is half of the data clock.

[STaRGaZeR]

Can it even go that low on all boards? 2000 MHz. effective is a 1000 MHz. data clock...
Although you were probably referring to the command clock, which is half of the data clock.

Yes. The value you can change in CCC, AMD GPU Clock Tool etc. is the command clock, so that's what the review is refering to.

[Kensei]

it's google translator. Original article is in Russian.

and english version cooming soon, as soon as I'll get all the source.

Now I've translated only first part, till pics.

[Manicdan]

just looking at the charts, its really helpful until they set the scale to show 90%+ instead of starting at 0%, its just annoying.

seems the clock scaling is quite linear, and anything above 3000mhz on ram requires the core at 800 to notice much.

[saaya]

BTW i like your :banana::banana::banana::banana::banana: cooper

hmmmm quite tempted to add that to my signature

very nice article! thanks artem!

[Iconyu]

Dude is that the guy from blood sport? this avatar made me piss my pants! wow! its a good one! sorry ot.

Bloodsport?! Is this comment a trap?