Does bandwidth matter on HD 3800?

[Spirou]

Yes, i strongly believe it does. Some say it doesn't.

In theory there must be some dependancy between GPU speed and memory speed, and setting up different clock speeds should show which of both is the limiting factor.

Taking in account that it's difficult to determine how different clock settings influence game performance due to the sheer amount of cross dependencies, i simply tested different clock settings using Ati Tray Tools' internal benching facility with my Sapphire HD 3870 Dual Slot.

Common clock-settings sorted by result:

GPU/Mem (MHz) Result, Comment
-----------------------------------------------
904.50/0999 = 14219,20 very high GPU, low mem
776.25/1125 = 15891,41 reference GPU, ref mem
904.50/1125 = 16063,80 very high GPU, ref mem
796.50/1152 = 16274,39 S.Toxic GPU & mem
796.50/1197 = 16773,16 typical OC GPU & mem
823.50/1197 = 16910,54 typical OC GPU & mem
850.50/1197 = 17018,82 S.Atomic GPU & mem
695.25/1341 = 17370,40 low GPU, extreme mem
850.50/1251 = 17661,97 high GPU, high mem
904.50/1251 = 17887,53 very high GPU, high mem
776.25/1341 = 18068,95 reference GPU, extreme mem
850.50/1296 = 18105,02 Atomic GPU, very high mem
904.50/1296 = 18488,85 very high GPU, very high mem

904.50/1305 = 18560,28 very high GPU + 0.0, mem + 9
904.50/1323 = 18758,06 very high GPU + 0.0, mem + 27
918.00/1323 = 18851,14 very high GPU + 13.5, mem + 27
918.00/1332 = 18921,22 very high GPU + 13.5, mem + 36
931.50/1332 = 19003,25 very high GPU + 27.0, mem + 36
945.00/1332 = 19077,39 very high GPU + 40.5, mem + 36
931.50/1341 = 19107,60 very high GPU + 27.0, mem + 45
945.00/1341 = 19188,50 very high GPU + 40.5, mem + 45
945.00/1350 = 19290.98 very high GPU + 40.5, mem + 54
958.50/1350 = unstable very high GPU + 54.0, mem + 54

Highest stable test so far:

I also tested some extreme combinations to find an optimum ratio between GPU and memory clocks which can be described as a sweet spot. It's also clear to see, where limitation steps in. Optimum ratio seems to be between 9/13 (reference clock ratio) and 2/3 on first sight, but even at 776 MHz GPU an extreme memory overclock shows how important bandwidth is for the HD 3800 Series.

Hope this tests can help to decide which overclock should work best, or wether or not a given model can be expected to be a good choice. My advice: stay away from GDDR3 or models labeled as turbo or oc, which IMHO are not as fast as you would expect.

Depending on heatsink (mine did all this tests at 45 degrees) and different voltage regulators all HD 3870 should reach nearly 900 MHz and due to the specs of GDDR4 all should reach 1250 MHz without voltmodding.

Another session using FurMark v1.3, results by GPU clock:

GPU/Mem (MHz) Result, Comment
-----------------------------------------------
695.25/1341 = 2850 low GPU, extreme mem
776.25/1125 = 3181 reference GPU, ref mem
776.25/1341 = 3181 reference GPU, extreme mem
796.50/1152 = 3263 Toxic GPU & mem
796.50/1197 = 3262 typical OC GPU & mem
823.50/1197 = 3371 typical OC GPU & mem
850.50/1197 = 3479 Atomic GPU & mem
850.50/1251 = 3483 high GPU, high mem
850.50/1296 = 3483 Atomic GPU, very high mem
904.50/0999 = 3660 very high GPU, low mem
904.50/1125 = 3695 very high GPU, ref mem
904.50/1251 = 3702 very high GPU, high mem
904.50/1296 = 3701 very high GPU, very high mem

904.50/1305 = 3703 very high GPU + 0.0, mem + 9
904.50/1323 = 3703 very high GPU + 0.0, mem + 27
918.00/1323 = 3755 very high GPU + 13.5, mem + 27
918.00/1332 = 3758 very high GPU + 13.5, mem + 36
931.50/1332 = 3815 very high GPU + 27.0, mem + 36
931.50/1341 = 3812 very high GPU + 27.0, mem + 45
945.00/1332 = 3868 very high GPU + 40.5, mem + 36
945.00/1341 = 3867 very high GPU + 40.5, mem + 45
945.00/1350 = 3866 very high GPU + 40.5, mem + 54
958.50/1350 = ---- very high GPU + 54.0, mem + 54

As you see, there is no difference between various mem clocks. All results depend on GPU clock only. FurBench seems to use data from internal buffers in GPU, which is not accurate, because gaming performance is always a mix of calculation and data throughput.

Assumingly there'd be no difference in FurMark results for 3870 and 3850 set to same GPU clocks.

I will add another series using Canyon Flight soon. Anyone with a HD 3850 who could confirm some of the tests?

Comment if you like, or add different clock speeds if you wish. I will add them to this list.

[Darknova]

I believe bandwidth is always important in whatever component you are testing.

CPU
Mobo
RAM
Graphics card

Always important. Each component requires talking to one or more of the others, the more bandwidth the faster it can talk and the more data it can get down in one go.

Of course there will always be a point of diminishing returns, but bandwidth is still consequential.

[Wile E]

Nice write up.

[Darknova]

Originally Posted by Spirou

Optimum ratio seems to be between 9/13 (reference clock ratio) and 2/3 on first sight, but even at 776 MHz GPU an extreme memory overclock shows how important bandwidth is for the HD 3800 Series.

My advice: stay away from GDDR3 or models labeled as turbo or oc, which IMHO are not as fast as you would expect.

Depending on heatsink (mine did all this tests at 45 degrees) and different voltage regulators all HD 3870 should reach nearly 900 MHz and due to the specs of GDDR4 all should reach 1250 MHz without voltmodding.

My advice: Stay the f*** away from Sapphire's single slot POS. Mine won't do 1250mhz on the RAM (my old Powercolor did it without breaking a sweat) and the GPU will barely do 810Mhz (again, my Powercolor did 830Mhz). This is even with an Accelero S1 and turbo module.
This card is just poor build quality.

Oh, and I always said that OC'ing my RAM to 1250 on my powercolor did more for the performance than overclocking the GPU to 810mhz, but would anyone listen? nooo...

[HTC]

Originally Posted by Darknova

My advice: Stay the f*** away from Sapphire's single slot POS. Mine won't do 1250mhz on the RAM (my old Powercolor did it without breaking a sweat) and the GPU will barely do 810Mhz (again, my Powercolor did 830Mhz). This is even with an Accelero S1 and turbo module.
This card is just poor build quality.

Oh, and I always said that OC'ing my RAM to 1250 on my powercolor did more for the performance than overclocking the GPU to 810mhz, but would anyone listen? nooo...

Why? Why? WHY didn't you tell me this BEFORE i bought mine?

I've been able to bench it @ 904 / 1251 but with heavy artifacts, but can't repeat it because it crashes the PC

[Darknova]

Originally Posted by HTC

Why? Why? WHY didn't you tell me this BEFORE i bought mine?

I've been able to bench it @ 904 / 1251 but with heavy artifacts, but can't repeat it

Because I originally had a Powercolor and it's only recently it died and I had to get it replaced?

[HTC]

Originally Posted by Darknova

Because I originally had a Powercolor and it's only recently it died and I had to get it replaced?

Excuses ... Bah!!!

[Darknova]

Originally Posted by HTC

Excuses ... Bah!!!

Dont lol. I'm pissed off enough as it is that I got this crappy card as a "replacement" feel like I've been ripped off...not that the place I got it from will do anything about it "oh but we can only give you one of a lesser or same value" "even if I pay?" "yes" ....pathetic.

Bah, I'm still undecided whether to just cut my losses and buy a brand new HIS card, but this close to the 4870 I'm not sure...

[HTC]

Originally Posted by Darknova

Dont lol. I'm pissed off enough as it is that I got this crappy card as a "replacement" feel like I've been ripped off...not that the place I got it from will do anything about it "oh but we can only give you one of a lesser or same value" "even if I pay?" "yes" ....pathetic.

Bah, I'm still undecided whether to just cut my losses and buy a brand new HIS card, but this close to the 4870 I'm not sure...

I hear you, dude and i understand your anger!

I suppose i can try and sell mine. Can't you try and do the same?

[Darknova]

Originally Posted by HTC

I hear you, dude and i understand your anger!

I suppose i can try and sell mine. Can't you try and do the same?

Probably, but I'd be out a graphics card unless I bought one first :P

[niko084]

Nice and you are correct it does matter, but its minimal when compared directly to the gpu freq.
Obviously until you hit certain points when you need the memory bandwith though.

[HTC]

Originally Posted by Darknova

Probably, but I'd be out a graphics card unless I bought one first :P

Duh

[Spirou]

Sad to hear that.

I had a look at nearly all different models in february and made some efforts to describe their differences, listing the sku's which Sapphire did not like at all, and posted in lots of forums, mostly in german. Think of the different prices for single slot (11122-03) and dual slot (11122-05). In germany most reviews stated HD 3870 would have gone to extremely low prices and most dealers sold them with pictures showing the reference design, but delivered the single slot. They sold it for 170,- Euro in January, later in february for 150,- while the dual slot was at 200,- at the same time.

Don't know what happened to the new dual slot which is in production since february as a replacement for the BBA cards with ATI reference heatsink. Sapphire doesn't give much information. Some make a real bad story from what could be one of the finest.

I had lots of trouble to get the model i wanted, and bought it for 200,- (ligh retail) in march.

The SKU is important. Look for 11122-05 if you go for Sapphire. It has the best heatsink ever built and RAM gets cooled too. I hope it is still in production. Should be easier in US and Asia than in Europe where it shall not be sold at all by decision from Sapphire, which noone understands. For non-Sapphire look at the voltage regulators before you buy. I think the GPUs are the same on all models. Maybe the RAM is different.

[Darknova]

Spirou, I've been looking at the HIS IceQ3 Turbo, the one with the 850Mhz clock speed. I can more than likely OC the RAM to 1250Mhz and then I'll be set. Slightly better overclock than my powercolor did, but it means buying a new card...

[P4-630]

Originally Posted by Spirou

Yes, i strongly believe it does. Some say it doesn't.

In theory there must be some dependancy between GPU speed and memory speed, and setting up different clock speeds should show which of both is the limiting factor.

Taking in account that it's difficult to determine how different clock settings influence game performance due to the sheer amount of cross dependencies, i simply tested different clock settings using Ati Tray Tools' internal benching facility with my Sapphire HD 3870 Dual Slot.

Common clock-settings sorted by result:

GPU/Mem (MHz) Result, Comment
-----------------------------------------------
904.50/0999 = 14219,20 very high GPU, low mem
776.25/1125 = 15891,41 reference GPU, ref mem
904.50/1125 = 16063,80 very high GPU, ref mem
796.50/1152 = 16274,39 S.Toxic GPU & mem
796.50/1197 = 16773,16 typical OC GPU & mem
823.50/1197 = 16910,54 typical OC GPU & mem
850.50/1197 = 17018,82 S.Atomic GPU & mem
695.25/1341 = 17370,40 low GPU, extreme mem
850.50/1251 = 17661,97 high GPU, high mem
904.50/1251 = 17887,53 very high GPU, high mem
776.25/1341 = 18068,95 reference GPU, extreme mem
850.50/1296 = 18105,02 Atomic GPU, very high mem
904.50/1296 = 18488,85 very high GPU, very high mem

904.50/1305 = 18560,28 very high GPU + 0.0, mem + 9
904.50/1323 = 18758,06 very high GPU + 0.0, mem + 27
918.00/1323 = 18851,14 very high GPU + 13.5, mem + 27
918.00/1332 = 18921,22 very high GPU + 13.5, mem + 36
931.50/1332 = 19003,25 very high GPU + 27.0, mem + 36
945.00/1332 = 19077,39 very high GPU + 40.5, mem + 36
931.50/1341 = 19107,60 very high GPU + 27.0, mem + 45
945.00/1341 = 19188,50 very high GPU + 40.5, mem + 45
945.00/1350 = 19290.98 very high GPU + 40.5, mem + 54
958.50/1350 = unstable very high GPU + 54.0, mem + 54

Highest stable test so far: Attachment 14433

I also tested some extreme combinations to find an optimum ratio between GPU and memory clocks which can be described as a sweet spot. It's also clear to see, where limitation steps in. Optimum ratio seems to be between 9/13 (reference clock ratio) and 2/3 on first sight, but even at 776 MHz GPU an extreme memory overclock shows how important bandwidth is for the HD 3800 Series.

Hope this tests can help to decide which overclock should work best, or wether or not a given model can be expected to be a good choice. My advice: stay away from GDDR3 or models labeled as turbo or oc, which IMHO are not as fast as you would expect.

Depending on heatsink (mine did all this tests at 45 degrees) and different voltage regulators all HD 3870 should reach nearly 900 MHz and due to the specs of GDDR4 all should reach 1250 MHz without voltmodding.

Comment if you like, or add different clocks if you wish. I will add them to this list.

What about pixel and texture fillrates?
And what are your FurMark scores?

I have a single 3870 with gddr3, overclocked it outperforms most overclocked 3870 gddr4 cards.

[Spirou]

thx for the suggestion. I ran the same series of clock speeds using FurMark now and added my results. Looks like a characteristic of FurMark to scale on GPU clock only.

I don't believe in fillrates, because every card on the market gives higher fillrates than needed. At one GPixel/s a scene can be rendered in 1920*1440 with 100 fps 3.6 times overlayed (1.000.000.000/276480000=3.6).

You'd have to overlay each pixel more than 40 times to reach the limits of an HD 3870. I really don't know, what Nvidia wants to do with higher fillrates. Sounds like marketing bollocks to me.

[NinkobEi]

Originally Posted by Darknova

My advice: Stay the f*** away from Sapphire's single slot POS. Mine won't do 1250mhz on the RAM (my old Powercolor did it without breaking a sweat) and the GPU will barely do 810Mhz (again, my Powercolor did 830Mhz). This is even with an Accelero S1 and turbo module.
This card is just poor build quality.

Oh, and I always said that OC'ing my RAM to 1250 on my powercolor did more for the performance than overclocking the GPU to 810mhz, but would anyone listen? nooo...

My HIS wont do higher than 1150 on RAM. I'd kill for 1250!

[Dr. Spankenstein]

Awesome write-up! Now the problem remains: can you keep the RAM cool enough to run at these speeds without artifacting? Heat is the biggest factor.

Thanks for taking the time to put this together. Very informative!

[btarunr]

Originally Posted by Darknova

My advice: Stay the f*** away from Sapphire's single slot POS.

Is it the one in 'A' or cards with Vapour-X coolers (Atomic/Toxic) or both you're talking of? A friend bought the Toxic, he's not able to push the core beyond 820/1100 without artiacts.

[HTC]

Originally Posted by btarunr

Is it the one in 'A' or cards with Vapour-X coolers (Atomic/Toxic) or both you're talking of? A friend bought the Toxic, he's not able to push the core beyond 820/1100 without artiacts.

I had the first one: managed to sell it and got an Asus one instead! Hope this one is better!

[P4-630]

Originally Posted by Dr. Spankenstein

Awesome write-up! Now the problem remains: can you keep the RAM cool enough to run at these speeds without artifacting? Heat is the biggest factor.

Thanks for taking the time to put this together. Very informative!

To keep your gddr cool don't let a fan BLOW hot air up to your card!
Unfortunately most gpu coolers are designed to blow on the pcb...

Even gddr with heatsinks won't keep your gddr cool when hot air passed through a gpu heatsink that blows warm air on the memory and pcb.

Also a heatsink based on a 1 big copper base plate that covers the gpu and the ram, the gpu will heat up the copper plate and this will heat up the ram, and it does NOT make your ram cooler!

From experience I would say that the best cooling solution for a videocard is to SUCK the heat AWAY from the gpu and from the pcb, this way the pcb and ram stays MUCH cooler

[btarunr]

Originally Posted by HTC

I had the first one: managed to sell it and got an Asus one instead! Hope this one is better!

Better keep some RAM-sinks handy...



...at least for the ones behind.

[HTC]

Originally Posted by btarunr

Better keep some RAM-sinks handy...



...at least for the ones behind.

Already bought an accelero S1 Rev 2, Enzotech RAM sinks and 2 NF-P12 fans for it: both the cooler and the RAM sinks haven't arrived, yet ...

[btarunr]

Originally Posted by HTC

Already bought an accelero S1 Rev 2, Enzotech RAM sinks and 2 NF-P12 fans for it: both the cooler and the RAM sinks haven't arrived, yet ...

If you had better cooling at hand, why did you sell that Sapphire card in the first place

You could've used Accelero S1 on that card because IIRC, all that pulls back that card is poor cooling.

[HTC]

Originally Posted by btarunr

If you had better cooling at hand, why did you sell that Sapphire card in the first place

You could've used Accelero S1 on that card because IIRC, all that pulls back that card is poor cooling.

I'm still wondering about that ...