4850@750Mhz vs 4870

[keiths]

The question has been brought by people a few times about how much the performance of the 4870 is its memory bandwidth or GPU. The 4870 reviews give results for it, but haven't seen any results for a 4850 /w its GPU clocked to 750, so this is asking if some 4850 owners would bench something at 750 so that a rough comparison could be made.

[keninishna]

check the 700 MHz barrier smashed thread in the xtreme bios section. Someone scored over 30k in 3dmark 05 at 800 mhz on GPU the only problem is your gonna have to wait for some 4870 3dmark 05 benches.

[keiths]

Yes, it's already known that 4850s are clocking past 750. That's not the point. Again, I and others are interested in seeing the difference the 4870's memory bandwidth makes, doing so by clocking the 4850 core the same as the 4870. One can then better compare with the 4870 reviews. I say better, as the platforms will by chance be different.

[gundamit]

If you want to isolate the memory performance difference, wouldn't it be easier just to drop the core on the 4870 to 700 then measure against the a CCC Oc'ed 4850 at the same core clock?

[keiths]

No. The desire to is to be able to compare results from 4870 reviews and 750 is nothing to achieve; 4850 has been clocked to 900 already, with stock heatsink no less. That's sort of the issue, people have just been clocking to the max, not doing a comparison like this.

[aldamon]

I want to see this kind of comparison, too.

How starved for memory bandwidth is the 4850 at 800+ with max memory speed?

How does it compare to the 4870 overclocked to 800+ and max memory speed as well?

[oohms]

According to the expereview 4870 review, the 4870 is on average 25% faster.. but i agree, i would like to see just how memory limited these cards are

[gurusan]

Well my 4850 is only at 700/1100 at the moment because I'm waiting for some cooling things to come before I push it.

However, I did have it at 750/1100 briefly on 1.23v and did 2 crysis benches. First one is 1280x1024 all high, no AA. 2nd bench is 1680x1050 All very high, DX10, no AA....and the GPU was at 810mhz at the time as well :/ sorry but it's all i have. I don't really bother with 3dmark06 because 1) these 48xx series don't seem to do particularly well in 3dmark...and 2) because game benchmarks are what matter...screw synthetics :p

http://img511.imageshack.us/img511/5935/79763882mj6.png

http://img168.imageshack.us/img168/6305/32runkt4.png

[aldamon]

According to the expereview 4870 review, the 4870 is on average 25% faster.. but i agree, i would like to see just how memory limited these cards are

Yeah, but that 25% difference was at stock speeds. 750 is 20% higher than 625, so are we to assume that at stock speeds, the GDDR5 only gives the 4870 a 5% advantage, clock per clock? That's a pretty crappy value if so and would explain the BIOS lock. I'm curious to see if that gap widens with higher core speeds.

[keiths]

Just as helpful would be for those with a 4870 to underclock their gpus to 625 and bench something to see what performance results from the memory bandwidth.

[keiths]

Alright, I've gone through the 4870 review thread list(http://www.xtremesystems.org/forums/showthread.php?t=192421&highlight=4870+review,) grabbed all the comparable results, sorted, giving us the following:

crysis dx10 high noaa
gurusan e6750@3.6
1280x1024 33.78 48.52 58.68 4850@750
1680x1050 16.74 22.88 27.00 4850@810 very high

anand qx9770@3.2
1280x1024 ??.?? 45.50 ??.?? 4870
1680x1050 ??.?? 35.60 ??.?? 4870
1280x1024 ??.?? 37.40 ??.?? 4850
1680x1050 ??.?? 29.80 ??.?? 4850

pcper x6800@2.93
1280x1024 23.00 43.10 57.00 4870

hardwarecanucks qx9770@3.85
1280x1024 25.98 40.61 ??.?? 4870
1280x1024 23.11 34.27 ??.?? 4850

expreview qx9650@4.0
1680x1050 ??.?? 40.46 ??.?? 4870
1680x1050 ??.?? 32.58 ??.?? 4850

chw.net qx9650@3.0
1680x1050 20.90 38.00 ??.?? 4870
1680x1050 11.70 30.50 ??.?? 4850
very high
1680x1050 02.00 16.50 ??.?? 4870
1680x1050 02.90 13.10 ??.?? 4850

legitreviews qx9770@3.2
1280x1024 ??.?? 38.30 ??.?? 4870
1280x1024 ??.?? 34.80 ??.?? 4850

1280x1024 crysis dx10 high noaa
33.78 48.52 58.68 4850@750 gurusan e6750@3.6
??.?? 45.50 ??.?? 4870 anand qx9770@3.2
23.00 43.10 57.00 4870 pcper x6800@2.93
25.98 40.61 ??.?? 4870 hardwarecanucks qx9770@3.85
??.?? 38.30 ??.?? 4870 legitreviews qx9770@3.2
23.11 34.27 ??.?? 4850 hardwarecanucks qx9770@3.85
??.?? 37.40 ??.?? 4850 anand qx9770@3.2
??.?? 34.80 ??.?? 4850 legitreviews qx9770@3.2

1680x1050 crysis dx10 high noaa
??.?? 35.60 ??.?? 4870 anand qx9770@3.2
??.?? 40.46 ??.?? 4870 expreview qx9650@4.0
20.90 38.00 ??.?? 4870 chw.net qx9650@3.0
??.?? 29.80 ??.?? 4850 anand qx9770@3.2
??.?? 32.58 ??.?? 4850 expreview qx9650@4.0
11.70 30.50 ??.?? 4850 chw.net qx9650@3.0
22.20 36.70 ??.?? 260gtx chw.net qx9650@3.0

1680x1050 crysis dx10 very high noaa
16.74 22.88 27.00 4850@810 gurusan e6750@3.6
02.00 16.50 ??.?? 4870 chw.net qx9650@3.0
02.90 13.10 ??.?? 4850 chw.net qx9650@3.0
09.30 16.40 ??.?? 260gtx chw.net qx9650@3.0
??.?? 26.59 ??.?? 280gtx stelaras e8500@3.8 http://www.thelab.gr/showthread.php?t=52224&page=16
13.44 19.38 23.34 4870@625 dnottis q6600@3.6

Thanks to gurusan responding; could use some more data points and other game benches(hint, hint 4850/4870 owners:))

Hunt turned up a forum post of a 280gtx for 1680x1050 very high. 4850@900 should about match this.

[aldamon]

keiths, gurusan's second set of Crysis numbers is at Very High, not High.

[keiths]

Ah, missed that. There was only one or two reviews that benched at very high, so it'd be helpful if gurusan benched at high, but I'll look at tracking those results down again.

[Truckchase!]

here's a different take on how to tell what impact mem bandwidth is having...

1> benchmark 4850 @ stock speeds
2> oc the GPU and vMem by a static percentage (15%) and bench
3> OC just the GPU and Just the vMem by the same static percentage and benchmark each oc seperately.

Compare the results to see how large of an impact the vMem overclocking has on performance vs. the gpu. In all actuality, you need only to do step 1 and 3, but 2 is always fun to compare against to make sure there isn't an odd unknown variable.

Here's an example from when I did this on a 2900xt:
http://www.xtremesystems.org/forums/showpost.php?p=2381435&postcount=31

I'm willing to do this, as I have an unopened HIS 4850 sitting here, but last night I ordered a 4870, so I was just gonna return the '50. Anyone wanna buy a slightly used 4850 for $175 next week so I can do the 4850 benching this weekend? :D

edit: I'll do the testing on the new 4870 when I get it as well.... I've got an automated test suite that makes it pretty easy.

[keiths]

Lucked out. One review had 1680x1050 at very high noaa. Updated the listing.

[keiths]

My critique of what little results there are so far is the gpu is the significant factor and the memory difference not even registering, for crysis at least. Of course, this is just at 1280x1024 and 1680x1050. Need some higher res benches and aa.

[gurusan]

here's a random result that won't help you with any direct comparisons but figured I'd throw it in here:



http://img406.imageshack.us/img406/6613/87981222an9.jpg

[aldamon]

My critique of what little results there are so far is the gpu is the significant factor and the memory difference not even registering, for crysis at least. Of course, this is just at 1280x1024 and 1680x1050. Need some higher res benches.

I agree and if it doesn't register in Crysis, then where would it register? Nothing else touches Crysis.

[Truckchase!]

I agree and if it doesn't register in Crysis, then where would it register? Nothing else touches Crysis.

Crysis is all math. There isn't much strain put on the memory. Did you guys read my post? This is a pretty simple issue that one guy with a 4850 could solve....

[aldamon]

You shouldn't have to do that. It's a shame that the dozens of hardware sites, that had both cards in their possession, didn't run this kind of test so you wouldn't be put in that position. It's like they took for granted that the 4850 would never hit the core speed of the 4870 and just ignored the possibility the GDDR5 is all fluff.

[TheScavenger]

I am interested in seeing these comparisons as well.

[Hornet331]

hmm i think oblivion with texture pack would stress memory, thought it also would stress the core a lot.

[gurusan]

a simple 3dmark06 SM2 and SM3 run would work if I ran it at 1680x1050 8xAA 16xAF...no?

And how about UT3bench with 16xAF and 8xAA 1680x1050

[keiths]

Need to be able to compare, so need to choose from the games and settings the reviews use, unless you can find some results from forums with the same game and configuration like I happened to find with the 280gtx result to compare with your very high run.

Truckchase!, a mild overclock shows the difference between GPU and memory, but it's less likely to show where the performance drop off is at, the 4850@750 and 4870@625 allows to compare to any review, and satisfies that whatif itch of a question of medium vs high end card. Also of interest and in the same vain as the medium vs high end whatif is max overclock vs 260/280gtx competition/topdog.

[Truckchase!]

Need to be able to compare, so need to choose from the games and settings the reviews use, unless you can find some results from forums with the same game and configuration like I happened to find with the 280gtx result to compare with your very high run.

Truckchase!, a mild overclock shows the difference between GPU and memory, but it's less likely to show where the performance drop off is at, the 4850@750 and 4870@625 allows to compare to any review, and satisfies that whatif itch of a question of medium vs high end card. Also of interest and in the same vain as the medium vs high end whatif is max overclock vs 260/280gtx competition/topdog.

1. What performance drop off? GPU, memory, both? Be specific.
2. You can't compare to a review. They have different base hardware, different OS, different benching techniques, etc. On top of that I can't guarantee the skill of the tester. (nor should you)
3. Max overclock will vary on a card by card basis, but obviously will reach an average. (as I believe you're getting at) However, the point of this thread is not to compare to Nvidia products, but to compare a 4850 vs. A 4870.

My point is that the main difference between a 4850 and 70 is the memory bandwidth. (yes I realize the voltage circuity is better on a 70) Given this information, one can conduct tests to see if the 4850 is memory bandwidth limited and answer the majority 4850 vs. 4870 question, which is the premise of this thread. That being said, I think I'm going to unwrap my 4850 tomorrow and do some tests since I have the automated scripts to do so. The offer still stands if anyone wants to buy this card since I'm getting 70 from FXVideoCards on Wednesday.

Edit: Oh snap, I didn't see that it was you that opened the thread until now KeithS.... do you want to change the context? I can't provide Nvidia data.... I'll just say they're "for the most part, faster and more expensive" :D

[Truckchase!]

You shouldn't have to do that. It's a shame that the dozens of hardware sites, that had both cards in their possession, didn't run this kind of test so you wouldn't be put in that position. It's like they took for granted that the 4850 would never hit the core speed of the 4870 and just ignored the possibility the GDDR5 is all fluff.

I enjoy doing this stuff. Most reviews aren't very good. Quantity does not necessarily equal quality. :D I appreciate the concern tho....

[adpr_02]

Here is my 2C. I'm sure it's not much, but just shows that memory does help - at least in synthetics.

According to Legit Reviews, the HD4870 gets P8769 3dmarks.
Their processor is the newer Q9770.

My system, with the card clocked at 720/1050 and my q6600 at 3.6ghz I get P7850 marks.

The net processor performance, I am guessing is the same since, although theirs is 3.2ghz only, it's the newer version with higher performance/clock. Also, I do not think the CPU score has a very large impact on the overall score.

Seeing as how my clocks are so close to the 4870 (with the memory overclocked, even) and my score is so far away, I am guessing that the GDDR5 does have an impact, but it is not that large. I am sure that if I pushed my core to 750 I would jump my score to P8100.


Just my 2C

[aldamon]

Here is my 2C. I'm sure it's not much, but just shows that memory does help - at least in synthetics.

According to Legit Reviews, the HD4870 gets P8769 3dmarks.
Their processor is the newer Q9770.

My system, with the card clocked at 720/1050 and my q6600 at 3.6ghz I get P7850 marks.

The net processor performance, I am guessing is the same since, although theirs is 3.2ghz only, it's the newer version with higher performance/clock. Also, I do not think the CPU score has a very large impact on the overall score.

Seeing as how my clocks are so close to the 4870 (with the memory overclocked, even) and my score is so far away, I am guessing that the GDDR5 does have an impact, but it is not that large. I am sure that if I pushed my core to 750 I would jump my score to P8100.


Just my 2C

If you did overclock to 750 and hit P8100, you'd only be 8.25% off of the 4870's pace. Games are showing an even smaller difference.

[adpr_02]

I've got close to that. I looked at their score, and didn't realize they shwed the gpu and cpu score. So I get my cpu score close to theirs, and bench again, this time with my card at 735/1060.

I got 7121 marks on my gpu score
They got 7890 marks on their 4870 gpu score

My cpu score was 13295
Their cpu score was 13132

I noticed that cpu speed does not affect the gpu score at all (My gpu score was exactly identical at 3.6ghz and 3.8ghz.)....

Overall, I got 8050 marks
They got 8769

My gpu score jumped more than 100 marks by increasing the gpu frequency 15mhz (and memory 10mhz). At 750/1050 this card should score around 7250 gpu marks which would bring it close to the 4870.

I am guessing that this card clocked at 800/1200 should EASILY overtake the 4870. I'll try that when I get the right size variable resistors:D.

[keiths]

1. What performance drop off? GPU, memory, both? Be specific.

2. You can't compare to a review. They have different base hardware, different OS, different benching techniques, etc. On top of that I can't guarantee the skill of the tester. (nor should you)
3. Max overclock will vary on a card by card basis, but obviously will reach an average. (as I believe you're getting at) However, the point of this thread is not to compare to Nvidia products, but to compare a 4850 vs. A 4870.

My point is that the main difference between a 4850 and 70 is the memory bandwidth. (yes I realize the voltage circuity is better on a 70) Given this information, one can conduct tests to see if the 4850 is memory bandwidth limited and answer the majority 4850 vs. 4870 question, which is the premise of this thread. That being said, I think I'm going to unwrap my 4850 tomorrow and do some tests since I have the automated scripts to do so. The offer still stands if anyone wants to buy this card since I'm getting 70 from FXVideoCards on Wednesday.

Edit: Oh snap, I didn't see that it was you that opened the thread until now KeithS.... do you want to change the context? I can't provide Nvidia data.... I'll just say they're "for the most part, faster and more expensive" :D

1. Memory. I'm interested in seeing how big a benefit the 4870's memory gives.

2. You can, and we already have, roughly. In spite of different platforms, there are trends in the results.

3. Of course, and yes, average.

I had added it since gurusan provided an oc result. That a 4850 can reach 260 and 280 performance makes me interested in seeing those results as well.

[gurusan]

The 4850 can surpass the GTX 280 :D The guy with .2 fps more than me has a highly clocked 45nm quad @ 3.7ghz while I only have a 65nm dual core at 3.6ghz. Also you should notice his GTX 280 is overclocked quite a bit :p

http://img511.imageshack.us/img511/6281/83392671aj3.jpg

[Lightman]

The 4850 can surpass the GTX 280 :D The guy with .2 fps more than me has a highly clocked 45nm quad @ 3.7ghz while I only have a 65nm dual core at 3.6ghz. Also you should notice his GTX 280 is overclocked quite a bit :p



That is very impressive:shocked:!!

Congratz!!

[keiths]

Yep, what I expected from linear extrapolation of your 810 result. What's your cooling now gurusan? Is 1680x1050 your max/monitor resolution? I'd like to see where the 4850's memory peaters out. And where is that chart from?

[gurusan]

on watercooling. 27-28c idle 36-38c load.

And yeah 1680x1050 is my max resolution. The chart is from a private forum so I can't really share it here :p

[aldamon]

I'm ordering one of these bad boys when I get back from the beach.

[keiths]

What's your water cooling setup?

[gurusan]

Laing D4 --->XSPC Edge--->Maze4--->Alphacool 120.3

[zaraza]

in order to reach those speeds could i do it on AIR with a accellero s1 v2 with 2 120mm fans? Also the pencil volt mod, is it safe to do with an S1 on it. How high will i be able to clock with the pencil mod? I want to try it when i get my 4850 next week. Also what kind of pencil did you use for the mod?

[gurusan]

you should be able to do high 850-900mhz on air i would think. And I just used a #2 pencil :)

[zaraza]

Dang gurusan thats awsome! i really want to buy the 4850 now and try the mod, i was thinking about buying it from besbuy inorder to have the option to return it if it goes wrong. lol. but newegg gives 2gb of ddr2 for free with a visiontek. In the techpower article it said i wont get more than 1.4 on the vddc, is that true? Please stick with me and help me understand this. This is my first time considering to volt mod.

[zaraza]

the only problem i would have from overclocking, is that i have to flash the bios to get the speeds right?

[keiths]

Long ago I did a homemade water setup for my k6; square PCV block caulked right onto the CPU:) Still have the bilge pump I used(attwood v450, 450GPH@13.6v so ~= 397GPH/1503LPH@12v vs 1400LPH for the Liang D4,) and tubing. Also have an AC condenser/evaporator from a 90s Civic that I can use as a radiator, which is as tall as my tower case, slightly narrower, ie big for this use, which has a half and quarter inch inlet/outlets. Leaves coming up with a container to hold the pump, scrounge something up, and the water blocks.

[aldamon]

Can we get a list of expected overclock and maximum voltage per cooling option?

Is this about right?

Stock cooler with higher fan speed
1.3V? 800MHz?

Upgraded air
1.4V? 850MHz?

Water
1.5V+? 900MHz+?

[gurusan]

910mhz on 1.45v for me under water

[zaraza]

someone with a 4870 really needs to step it up and post comparisons to 4850 oc's

[dnottis]

910mhz on 1.45v for me under water

What block?

[gurusan]

What block?

maze4. 27c idle 36-38 load

[dnottis]

maze4. 27c idle 36-38 load

any problems with it fitting? no mods or anything right?

[dnottis]

Ive got a Q6600 65nm @ 3.6Ghz, 4870 @ stock I scored P89xx. Dont remember the exact number but was just under 9000.

I will run Crysis 1680 later, what settings are all the ones in the spreadsheet run at? 1680x1050, very high, no AA?

[jaredpace]

dudes, hunch: r700 will be slightly faster than two 4850's in crossfire. :yepp:
i checked some benchies, 4850 averages around 79-80% the performance of a 4870, i think the extra 125mhz adds about 17% while have gddr5 adds the next 3-4%.

if the plx bridge doesn't do anything super for increasing r700's performance over 2x hd4870 in old-fashioned traditional [two-pcb/two-pcie] x-fire, then we should expect R700 to be 4% faster than 4850 x-fire (which still eats up a GTX 280 for the performance crown). Nvidia loses come August! Wonder if they take the lead again with 55nm gt200?

edit: just realized i posted this in the 4850 vs. 4870 core speed comparison thread, but it still has relevant information. I agree with this guy 100%:

Yeah, but that 25% difference was at stock speeds. 750 is 20% higher than 625, so are we to assume that at stock speeds, the GDDR5 only gives the 4870 a 5% advantage, clock per clock? That's a pretty crappy value if so and would explain the BIOS lock. I'm curious to see if that gap widens with higher core speeds.

[TheScavenger]

Can we get a list of expected overclock and maximum voltage per cooling option?

Is this about right?

Stock cooler with higher fan speed
1.3V? 800MHz?

Upgraded air
1.4V? 850MHz?

Water
1.5V+? 900MHz+?

I second that idea. We need a chart!

[dnottis]

Q6600 G0 L737B @ 3.6Ghz
Asus P5K-E/WiFi (Bios 1012)
OCZ DDR2-8500 DDR2-1066 4096MB Memory
1/2" ID Masterkleer, Swiftech MCP-655, Thermochill PA120.2 Rad, D-Tek Fuzion, Swiftech Stealth
Thermaltake ToughPower 1000watt PSU
HIS HD4870 790/1000
X-Mystique 5.1 Digital Dolby Sound Card
Lian Li PC V-2000
Acer X222Wbd 22" LCD

http://3dxtreme.net/other/HIS%20HD4870/dx10-64bit.jpg

[keiths]

dnottis, if you would, do a run with your 4870 set with GPU@625 and memory@stock. This would contrast with the results of gurusan's 4850 set to GPU@750 and memory@stock, and the stock 4850s & 4870s. Though I think we already know what the outcome will be:)

[keiths]

http://www.xtremesystems.org/forums/showthread.php?t=191944&highlight=Accelero

Comparing gurusan's numbers with the above results, 4850@910 /w water cooling edges out a 4850@700 /w an Accelero at load, and destroys it at idle. I would definitely say water FTW:) To be fare though, kiwi was limited to ccc overclocking at the time.

[dnottis]

Guys, I have a waterblock coming for this very soon, in the meantime I installed a TT Duo Orb. Idle is 40's, load 60s. Great air cooler imo.

In the other thread they talk about not leaving the VRMs uncooled. Are the VRMs the grey things I dont have sinked or the smaller black things I do have sinked? TIA.

http://3dxtreme.net/other/HIS%20HD4870/TT%20Duo%20Orb/IMG_5286.JPGhttp://3dxtreme.net/other/HIS%20HD4870/TT%20Duo%20Orb/IMG_5292.JPGhttp://3dxtreme.net/other/HIS%20HD4870/TT%20Duo%20Orb/IMG_5293.JPG

http://3dxtreme.net/other/HIS%20HD4870/TT%20Duo%20Orb/IMG_5298.JPGhttp://3dxtreme.net/other/HIS%20HD4870/TT%20Duo%20Orb/IMG_5299.JPGhttp://3dxtreme.net/other/HIS%20HD4870/TT%20Duo%20Orb/IMG_5300.JPG

http://3dxtreme.net/other/HIS%20HD4870/TT%20Duo%20Orb/IMG_5301.JPG

[Calmatory]

I didn't read the whole thread, only the few first posts but anyway.

You can't really say how much card is held back by the memory speed or GPU speed. It depends on the game/app and settings used. High resolutions with AA/AF require more bandwidth than low resolutions without them. Thus one can't really give any kind of percentage estimates that would apply in all apps/games. It depends so much in the game/app. For example in Company of Heroes my 7300GT is very badly GPU bound, giving very linera FPS increase as I increase the GPU clock. Then in Trackmania Nations Forever the same applies for memory, but not for GPU. I can underclock the GPU from 500 to 400 without big FPS drop, but if I underclock the memory from 500 to 400, the performance drops by almost 20%. Oh, and ATITool's 3D Cube FPS seems to scale well with memory BW.

[dnottis]

dnottis, if you would, do a run with your 4870 set with GPU@625 and memory@stock. This would contrast with the results of gurusan's 4850 set to GPU@750 and memory@stock, and the stock 4850s & 4870s. Though I think we already know what the outcome will be:)

Here it is...
http://3dxtreme.net/other/HIS%20HD4870/625-900.jpg

[keiths]

That was unexpected. Matches gurusan's score @810, doesn't match chw.net's 4850 score, and far exceeds their 4870 score. They used a quad@3.0, but variance in platform has only had a minor effect on the results, while this is out of the park different. Something's not right, your underclocked 4870 is closing in on a 280gtx.

[dnottis]

Welp the CPU is at 3.6Ghz, ram at 1066Mhz. The CP showing the GPU speed was opened right after the run. It was definitely run at 625/900.

[ExodusC]

So we still have no definitive answer on how much (if at all) the GDDR5 memory helps the 4870's performance?

I'm assuming I should just go with a 4850 and an aftermarket heatsink and do some overclocking.

[adamsleath]

question:
does having faster memory make memory amount less critical?
ie i have this idea that having higher bandwidth will mean that vram memory is less likely to be maxed at high res/settings, but is this true or false?
if true would mean more vram on lower bandwidth cards might help :shrug:
or if false, it's the opposite of what im saying and high bandwidth needs more vram:eek::shrug::confused:

nv approach 1gig vram high bandwidth
ati approach...not yet 1gig vram and high bandwidth...

[Miwo]

Heres my Vantage run at 750/1000. This was on "High" preset, because i can't run the Extreme because I can't select 1920x1200 as a resolution (I'm on a 1080p TV)
For comparison purposes, Techreport's scores @ High on a QX 3.0GHz: http://techreport.com/articles.x/14990/14

http://techreport.com/r.x/radeon-hd-4870/3dm-gpu.gif
http://www.fileden.com/files/2006/6/20/80195/750core.jpg

Going off GPU score only,

4870 @ 750/900: 4886
4850 @ 750/1000: 4485 <--- My Score
4850 @ 625/1000: 3851

So.... in Vantage High,
80% higher bandwidth of the 4870 over my 4850 @ 750/1000 yields an 8.9% increase in score.
20% Increase in Core Speed (750MHz vs 625MHz) yields an increase of 16.4% in score.

Hope these results were a little helpful in shedding more light on this subject :). Obviously, certain scenarios will favor memory bandwidth, and some it may not matter so much. Will do some more testing when I get around to it.

[Calmatory]

It doesn't matter are you using 50 GB/s or 500GB/s VRAM if it runs out. It has to be swapped to the system RAM and then back when needed, and thats why it is so slow. Basically higher FSB might help if VRAM runs out, but does it matter if FPS is less than 2 anyway?

[zaraza]

Miwo can you ran crysis if you have it? Dang im in the same position, im curious to see if the ddr5 really benefits it that much. Heck 5-7 % according to http://www.ocforums.com/showthread.php?t=567799&page=3 is not much. Look at posts 65 and 66.

[Miwo]

Sorry, this is pretty much a fresh Vista64 install and i dont have any games installed yet. Just vantage for now

Here is Vantage High again @ 925/1000MHz
http://www.fileden.com/files/2006/6/20/80195/925core.jpg
Techreport for comparison:
http://techreport.com/r.x/radeon-hd-4870/3dm-gpu.gif

48% Core increase(925MHz) over Stock = 5141 GPU Vantage High (33%)
20% Core increase(750MHz) over Stock = 4485 GPU Vantage High (16.4%)
Default 625 / 1000 = 3851 GPU Vantage High (Baseline Techreport's Results, sorry didn't test stock :) )

Definitely see some diminishing returns here!!!. Will have to do vmem mod tomorrow and play around with the DDR3. My card is definitely dying for some bandwidth.

Temps are 48C under 1.55v using Accelero S1 + 120mm Delta TriBlade + 70mm YS Tech. Coolaboratory Liquid Pro. I think Power Circuitry is holding me back or something. ATITool artifact tester fails with 1 pixel incorrect (I can't see any yellow flakes), Vantage passes fine, and my temperatures are fine. 4870 does have an extra 6pin connector as well as beefier vregs. Still, all in all very fun card to play with for under $200. I'll probably see if I can run this around 1.45-1.5v @ 900MHz for 24/7 usage

[dnottis]

That was unexpected. Matches gurusan's score @810, doesn't match chw.net's 4850 score, and far exceeds their 4870 score. They used a quad@3.0, but variance in platform has only had a minor effect on the results, while this is out of the park different. Something's not right, your underclocked 4870 is closing in on a 280gtx.

Alright, I know why. In the screenshot Enable Overdrive is disabled (no check in the box) so actually it was running stock. Still a good showing for a stock HD4870. I'll rerun at 625/900 a bit later when I'm home.

[zaraza]

sweet, thanks Dnottis, I am really close to pulling the trigger on either card. Just no 4870's available and 4850s are on sale like crazy lol.

[dnottis]

sweet, thanks Dnottis, I am really close to pulling the trigger on either card. Just no 4870's available and 4850s are on sale like crazy lol.

Here is the corrected HD4870 @ 625/900 you guys wanted. The other pic (in previous post) is updated also -

http://3dxtreme.net/other/HIS%20HD4870/625-900.jpg

[keiths]

You must have goofed and copied the old picture onto your new post instead of the other way around; your new numbers are the same as the old numbers. 22.575 average.

[grimREEFER]

so, what do u guys reckon, a 4850@800mhz would be faster than a stock 4870?

[zaraza]

still tryin to figure that out lol. all we have the the crysis GPU test and Vantage. It looks like yes? but we dont know the effects that the bandwidth will have on the card.

[Truckchase!]

I didn't read the whole thread, only the few first posts but anyway.
.......


Read my posts.

[aldamon]

still tryin to figure that out lol. all we have the the crysis GPU test and Vantage. It looks like yes? but we dont know the effects that the bandwidth will have on the card.

I think the fact that everyone has to search and compare and bench to find significant performance differences between these cards speaks volumes. The circumstantial evidence is starting to pile up.

[keiths]

Lower, but still way over stock 4870 results from chw. I added your numbers anyway. Why the more results the better, to spot the trends and anomalies. Hopefully more people will respond to get a better feel.

[biohead]

dnottis how does 750/900 (or lower mem) compare to 750/1000 (or higher mem if possible). the main question is how bandwidth hungry these cores are.

[gurusan]

another member and I ran some CoH benches on another forum and compared results. Remember his memory bandwidth is x4 effective, mine is x2 effective.

These are our quick results.

http://img127.imageshack.us/img127/7107/36267797qz2.png

[dnottis]

You must have goofed and copied the old picture onto your new post instead of the other way around; your new numbers are the same as the old numbers. 22.575 average.

Refresh - your browser is pulling the jpg from the cache. The name is the same so the pic isnt getting updated on your browser ;)

[zaraza]

another member and I ran some CoH benches on another forum and compared results. Remember his memory bandwidth is x4 effective, mine is x2 effective.

These are our quick results.

http://img127.imageshack.us/img127/7107/36267797qz2.png

That is perfect!!! im gonna go ahead and order a 4850 for 150 AR if its still instock after class. I got an S1 so its not gonna be a problem

[gurusan]

this is only from 1 game. another game may tell a different story. also for reference it was the DX9 demo i think, 1680x1050, everything maxed and AA "enabled" (either enabled/disabled, dunno how many multisamples it set it to)

[jaredpace]

my guess is that having GDDR5 is only adding 3-7% performance in most benchmarks between the 4850 & 4870. since both use Rv770's then why not get 1 or 2 of the cheaper 4850's and biosmod to use Vgpu high enough for 900mhz core...

Granted you may need a fan fix/ aftermarket air/ or water cooling solution :)

[gurusan]

you can only turn the voltage up to 1.20 through the bios on the 4850

[Truckchase!]

Here you guys go... let me know if you need any help interpreting the graphs; I didn't polish them up yet.

http://www.xtremesystems.org/forums/showthread.php?t=193085

2nd post is memory/gpu bottleneck data.

Comments welcome.

[jaredpace]

you can only turn the voltage up to 1.20 through the bios on the 4850

what kinda volts do you need for 800mhz, and do you have to do a solder voltmod? or can something be done through software?

so you did a pencil mod? & 800mhz requires ~ 1.3 volts?

[gurusan]

800mhz you will probably need around 1.26-1.3

check the voltmod thread for more info

[zaraza]

are there any other game tests you can run? with your friend? I really like that graph. I am about to purchase a 4850, i want to pencil mod it too. I saw that on the pencil mod guide, you cant get higher than 1.4 volts. Is that true?

[kryptobs2000]

I'd be more interested in seeing benches at 1920x1200+ and with & w/o antialiasing. I'm sure the bandwidth would be much more helpful there, and that's what I game at. I can't imagine people buying these cards to play at low resolutions ( < 1650x1050) and w/o AA/AF.

[aldamon]

another member and I ran some CoH benches on another forum and compared results. Remember his memory bandwidth is x4 effective, mine is x2 effective.

Did you guys have the same CPU platform and overclock?

[S_G]

Wow, that is slick. Thanks for the data. Very glad I adopted the 4850s early. Seems like I was right about the 4870 not really being worth the extra cash.

[Theta]

Everybody has asked for high res differences... so here we go... Some raw data between overclocked 4850 Crossfire and overclocked 4870 Crossfire...

2560 x 1600 Call of Duty 4 (COD4)

4850 Crossfire @ 800/1100 = Avg 96.40 FPS

2008-07-01 22:07:13 - iw3sp
Frames: 11463 - Time: 118299ms - Avg: 96.898 - Min: 46 - Max: 156

2008-07-01 22:10:48 - iw3sp
Frames: 11616 - Time: 121112ms - Avg: 95.911 - Min: 52 - Max: 162

4870 Crossfire @ 790/1000 = Avg 105.68 FPS -- 9.63% Faster than 4850 CF

2008-07-01 22:26:05 - iw3sp
Frames: 12419 - Time: 117629ms - Avg: 105.577 - Min: 51 - Max: 163

2008-07-01 22:33:31 - iw3sp
Frames: 12506 - Time: 118234ms - Avg: 105.773 - Min: 54 - Max: 160

4870 Crossfire @ 790/1100 = Avg 109.14 FPS -- 13.22% Faster than 4850 CF

2008-07-01 23:47:49 - iw3sp
Frames: 12648 - Time: 116528ms - Avg: 108.540 - Min: 51 - Max: 198

2008-07-01 23:51:22 - iw3sp
Frames: 13063 - Time: 119035ms - Avg: 109.740 - Min: 45 - Max: 185

These are just rough numbers, but it goes to show a few more things regarding upper-end clock speed, etc.

[purecain]

interesting thread, good info m8... keep it coming....

[adamsleath]

interesting results, good work.

Everybody has asked for high res differences
with AA ?

109 vs 95 fps....4870cf vs 4850cf
this may be more relevant if comparing 39fps to 25fps(for hypothetical example) in some heavily graphics bound app/settings/AA/etc. situation., but that is my question.

ie4870cf may pull away more from 4850cf in some hypothetical grfx bound situation...yada.anyway thats wot id like to see.(bandwidth advantage/disadvantage)

[gurusan]

weeeeeeeEeeee!

http://img133.imageshack.us/img133/3079/59575007me4.png

[zaraza]

what speed is your card at for that pic, gurusan?

[Cooper]

http://www.ixbt.com/video3/rv770-2-part3-d.shtml#p18

[aldamon]

http://www.ixbt.com/video3/rv770-2-part3-d.shtml#p18

Indeed. Sub-$150 4850 FTW. :D

[Jim Morbid]

Interesting... Maybe 3 way crossfire 4850's against a pair of 4870's?

JM

[aldamon]

English translation posted:

http://www.digit-life.com/articles3/video/rv770-2-part1-p4.html

We have also analyzed the effect of memory bandwidth on performance. The charts include results of the HD 4870 operating at 625 MHz (like HD 4850), memory frequency being the same 900 (3600) MHz. The results were quite surprising. In most cases performance didn't depend on memory bandwidth, being limited by GPU. Thus, we retract words about HD 4850 being cut down much in terms of memory bandwidth, and that increasing memory frequency would strongly benefit its performance. Turns out we were wrong.

[regenade]

I have done a review on the Sapphire HD 4870 against the Palit HD 4850 clocked @ 800 core/1150 mem . Performance is almost 5-7% within each other :)

http://www.techenclave.com/reviews-and-previews/sapphire-hd-4870-review-115201.html

[pauldovi]

I have done a review on the Sapphire HD 4870 against the Palit HD 4850 clocked @ 800 core/1150 mem . Performance is almost 5-7% within each other :)

http://www.techenclave.com/reviews-and-previews/sapphire-hd-4870-review-115201.html

I started to read it but the ads were too much.

[binormalkilla]

Well if anyone needs data at 1920x1080 I have my cards installed, but they're yet to be OCed......I'm going to flash 1.20V to them and see how high that takes me. I'm on MCW60s so they should scale as much as the voltage will allow.

[drizzt5]

Nice, this makes me feel a lot better for resisting the urge to splurge money on the 4870.
4850 <3
Nice results gurusan, thanks for the comparison.

[SNiiPE_DoGG]

i think you really need to consider the comparison between the max OC of each card. most 4870's will do 840 on stock volts...