ATI HD4800 Review Thread

[SKYMTL]

Man, this thread has gone downhill fast.

Regarding all of these "supposed" benchmarks showing the HD4870 being close to the GTX 280 under load, I say bollocks. Either the testing procedures are way out of wack ("hey, let's run 3DMark at default settings with the system plugged into a Kill A Watt!") or they aren't reporting it properly.

The reality of the situation is that under load the GTX 280 consumes up to 20% (around 70W) more power than the HD4870 while being disgustingly loud. That is the truth whether people accept it or not.

[Syn.]

Man, this thread has gone downhill fast.

Regarding all of these "supposed" benchmarks showing the HD4870 being close to the GTX 280 under load, I say bollocks. Either the testing procedures are way out of wack ("hey, let's run 3DMark at default settings with the system plugged into a Kill A Watt!") or they aren't reporting it properly.

The reality of the situation is that under load the GTX 280 consumes up to 20% (around 70W) more power than the HD4870 while being disgustingly loud. That is the truth whether people accept it or not.

Its the latest eplage called fanboyism, spreading even to the far reaches of the XS. What ever you do please do not try and have a constructive argument with the ones that are already infected or your common sense and logic will suffer a painful death.

Once you become one of them people will always

[Hornet331]

The reality of the situation is that under load the GTX 280 consumes up to 20% (around 70W) more power than the HD4870 while being disgustingly loud. That is the truth whether people accept it or not.

what reality?

there are currently no users out there that own both, and can make an independent statement.

So i have to go after the data that is available. The available show that the powerconsumption is between 20-50W more then a 280gtx if you look around the web.

[SKYMTL]

what reality?

While I can't go into every one of the details, I trust my own results since we are using the proper equipment and tests to properly determine power consumption of the full system (Line conditioner, UPM power meter, etc.). In addition, since I personally spent weeks determining the 3D application which would put the most constant load on the GPU, I can am confident in any numbers we produce.

Many tests you see are highly influenced by the CPU power consumption fluctuating up and down. The trick to properly determining power consumption is to first see which application will properly load the GPU while letting the CPU sit as idle as possible. Then make sure that the program can put an almost constant load on the GPU and its memory with minimal load times in order to get an even testing field.

Finally, the test MUST be run for AT LEAT 45 minutes in order to determine a peak power consumption number. This is due to the fact that many power meters have a sampling rate of about 250ms to 750ms which means that peaks in the power draw may not be logged. That is why you need to have a very long test under constant load; so the power meter can pick up the peaks in consumption even if it misses it the first, second, third and so on time.

I also have to say that it is extremely important to use a line conditioner for power consumption tests. As many of us know, input voltage can fluctuate quite a bit from a household power outlet. This voltage fluctuation can have a pretty large impact on efficiency numbers of power supplies which in turn would influence the numbers generated by any GPU power consumption test.

That is my story, that is how I tested and I am sticking behind my 70W difference statement 100%.

[Miwo]

While I can't go into every one of the details, I trust my own results since we are using the proper equipment and tests to properly determine power consumption of the full system (Line conditioner, UPM power meter, etc.). In addition, since I personally spent weeks determining the 3D application which would put the most constant load on the GPU, I can am confident in any numbers we produce.

Many tests you see are highly influenced by the CPU power consumption fluctuating up and down. The trick to properly determining power consumption is to first see which application will properly load the GPU while letting the CPU sit as idle as possible. Then make sure that the program can put an almost constant load on the GPU and its memory with minimal load times in order to get an even testing field.

Finally, the test MUST be run for AT LEAT 45 minutes in order to determine a peak power consumption number. This is due to the fact that many power meters have a sampling rate of about 250ms to 750ms which means that peaks in the power draw may not be logged. That is why you need to have a very long test under constant load; so the power meter can pick up the peaks in consumption even if it misses it the first, second, third and so on time.

I also have to say that it is extremely important to use a line conditioner for power consumption tests. As many of us know, input voltage can fluctuate quite a bit from a household power outlet. This voltage fluctuation can have a pretty large impact on efficiency numbers of power supplies which in turn would influence the numbers generated by any GPU power consumption test.

That is my story, that is how I tested and I am sticking behind my 70W difference statement 100%.

Just curious, but which application did you find that best tests power consumption Sky?

When I tested the power consumption of my RV670, I used Prime95 Small FFT to find the load of CPU only. Afterwards I let it run on a single thread, then use the free core to run ATITool Artifact tool, which seems to put the highest temperatures on my video card. Its single threaded, maxes out the cpu and maxes out the gpu.

Then I subtract the difference between (Prime95(singlethread)+ATITool) and Prime95(dualthread) to get just the power consumption of just my gpu. I'm using a killwatt plugged into a Belkin PureAV Line Conditioner. Care to comment if this sounds like an accurate way to test just GPU consumption? I use SmallFFTs because system memory does not get loaded much

[SKYMTL]

Just curious, but which application did you find that best tests power consumption Sky?

When I tested the power consumption of my RV670, I used Prime95 Small FFT to find the load of CPU only. Afterwards I let it run on a single thread, then use the free core to run ATITool Artifact tool, which seems to put the highest temperatures on my video card. Its single threaded, maxes out the cpu and maxes out the gpu.

Then I subtract the difference between (Prime95(singlethread)+ATITool) and Prime95(dualthread) to get just the power consumption of just my gpu. I'm using a killwatt plugged into a Belkin PureAV Line Conditioner. Care to comment if this sounds like an accurate way to test just GPU consumption? I use SmallFFTs because system memory does not get loaded much

Well, the issue is that we have no way to accurately judge full system load BEFORE the PSU. When we use a Kill A Watt or any other power meter the problem is that you also have to take the efficiency factor of the PSU into consideration. This can be a HUGE factor. Here are a few examples:

If you are drawing 700W from the wall with a PSU which has 85% efficiency at a certain load, that means that the system is actaully drawing 595W from the PSU.

On the other hand....

If you are drawing 700W from the wall with a PSU which has 80% efficiency at a certain load, that means the system is drawing 560W from the PSU.

So that change of a mere 5% in efficiency resulted in a 35W difference in actual system power consumption. See what I am saying? It is next to impossible to determine how much power is being drawn by JUST the GPU. Even a clamp meter won't work due due to it not being able to judge the draw from the PCI-E slot.


For my GPU tests I use the following running on a loop for 1-2 hours:

3DMark06 Batch Size Test @ highest triangle count
2X AA
16X AF
Highest resolution (2650 x 1600 in my case)

If you have a smaller monitor, increase AA as needed as long as FPS stay above 25.

This ensures minimal CPU overhead while stressing both GPU core and memory. It also scales well with SLI or CF which ATITool DOES NOT do.

[Hornet331]

While I can't go into every one of the details, I trust my own results since we are using the proper equipment and tests to properly determine power consumption of the full system (Line conditioner, UPM power meter, etc.). In addition, since I personally spent weeks determining the 3D application which would put the most constant load on the GPU, I can am confident in any numbers we produce.

Many tests you see are highly influenced by the CPU power consumption fluctuating up and down. The trick to properly determining power consumption is to first see which application will properly load the GPU while letting the CPU sit as idle as possible. Then make sure that the program can put an almost constant load on the GPU and its memory with minimal load times in order to get an even testing field.

Finally, the test MUST be run for AT LEAT 45 minutes in order to determine a peak power consumption number. This is due to the fact that many power meters have a sampling rate of about 250ms to 750ms which means that peaks in the power draw may not be logged. That is why you need to have a very long test under constant load; so the power meter can pick up the peaks in consumption even if it misses it the first, second, third and so on time.

I also have to say that it is extremely important to use a line conditioner for power consumption tests. As many of us know, input voltage can fluctuate quite a bit from a household power outlet. This voltage fluctuation can have a pretty large impact on efficiency numbers of power supplies which in turn would influence the numbers generated by any GPU power consumption test.

That is my story, that is how I tested and I am sticking behind my 70W difference statement 100%.

i am not doubting your results, just the variety of results found on the web. If every result would be 50-70W higher then i say ok, but we have results raging from 20-70W. This is kinda hard to swallow.

Maybe your result is the absolut worse case, dependig on what load you used. Since this 2 architekture are quite different, different situations will give different results. E.g. I dare to say that a scene based on heavily texture usage, will produce more heat/power draw on a 280gtx then on a 4870 (80TMUs vs 40TMUs).

hornet you need to stop blindly questioning everything anyone says, SYMTL is a reviewer who is very trusted and clearly has a thorough methodology, his results from both the 280 and the 4870 show the 280 being 70w higher then you just need to accept it

blind following = brainless = not my style
even if hes trusted i have my doubts, as i have with everything. As long as there is such a hugh discrepancy i belive nothing, till i get the chance to test it myself, or far more data if available.

[SKYMTL]

Maybe your result is the absolut worse case, dependig on what load you used. Since this 2 architekture are quite different, different situations will give different results. E.g. I dare to say that a scene based on heavily texture usage, will produce more heat/power draw on a 280gtx then on a 4870 (80TMUs vs 40TMUs).

Actually, worst case would be the test I run PLUS something like Orthos Burn (CPU + Ram) test running in the background. My test is an almost exclusively 3D worst case test.

Then again, testing worst case scenarios is the best thing to do IMO. Every program out there has a different power consumption envelope so knowing the worst case means you know if the PSU you are using is enough. If we didn't publish worst case figures you may have someone whose PSU would be fine in one application and then load up something like WiC only to have their system crash due to the increased GPU load in that game nailing their PSU in the gonads.

[Chickenfeed]

That is my story, that is how I tested and I am sticking behind my 70W difference statement 100%.

Thank you for the detailed explanation SKY. That is the issue with reviews, people compare and reference them but seem to always forget that each review often use different hardware/drivers and even when all is equal, testing procedures can greatly differ. Your way of testing load seems to be a consistent and fairly accurate way all things considered. My personal thoughts on other some other reviewers load tests, showing smalls spreads on the 4870/280, is they simply did a short load test ( some mention running 1 of the 3D06 tests such as Canyon ) which like you said isn't enough to get a proper reading even when using a good test application.

I still personally think that the TDP on the GT200 cards is somewhat conservative currently based on most results I've seen, eg most games won't fully leverage the architecture resulting in lower power usage, where the R770s get a better workout. Regardless the 4870 will still use less power, theres no 2 ways about it.

[cstkl1]

E8500@ 500x8
Biostar TPower I45
Team Xtreem PC2-6400 3-3-3-8 4x1gb@ 1200mhz Cl5
Sapphire HD4870 Crossfired
BFG Ageia Physx
WD Raptor X
Windows Vista Ultimate 64bit

Dont see any bottleneck here..



3dmark06 default

3dmark06 1920x1200

3dmark06 1920x1200 8AA 16AF

[vengance_01]

wow great scores with AA and AF which is what counts. Everyone plays games with Filters applied.

[Spawne32]

wow great scores with AA and AF which is what counts. Everyone plays games with Filters applied.

not everyone, personally i cant stand playing games at anything less then 50fps

[SNiiPE_DoGG]

what reality?

there are currently no users out there that own both, and can make an independent statement.

So i have to go after the data that is available. The available show that the powerconsumption is between 20-50W more then a 280gtx if you look around the web.

hornet you need to stop blindly questioning everything anyone says, SYMTL is a reviewer who is very trusted and clearly has a thorough methodology, his results from both the 280 and the 4870 show the 280 being 70w higher then you just need to accept it

[spursindonesia]

Man, this thread has gone downhill fast.

Regarding all of these "supposed" benchmarks showing the HD4870 being close to the GTX 280 under load, I say bollocks. Either the testing procedures are way out of wack ("hey, let's run 3DMark at default settings with the system plugged into a Kill A Watt!") or they aren't reporting it properly.

The reality of the situation is that under load the GTX 280 consumes up to 20% (around 70W) more power than the HD4870 while being disgustingly loud. That is the truth whether people accept it or not.

Agreed SKYMTL, i just can't understand people that insist HD 4870 has similar power consumption under load compared to GTX 280, i believe both ATi and nVidia don't pull TDP value for each respective product out of their CEO rear. HD 4870 TDP is 156w while GTX 280 TDP is 236w for God sake, and those are official numbers i believe.

[Hornet331]

Agreed SKYMTL, i just can't understand people that insist HD 4870 has similar power consumption under load compared to GTX 280, i believe both ATi and nVidia don't pull TDP value for each respective product out of their CEO rear. HD 4870 TDP is 156w while GTX 280 TDP is 236w for God sake, and those are official numbers i believe.

People need to realise, what TDP is...

TDP is a value that is aimed at OEMs to give them a guide value what the cooling solution should be able to disipate under certain conditions. This conditions vary depending on the manufacturer specifications. For intel and amd you can check this specifications, for ati and NV not.

Its quite possible that NV was more conservative with there TDP this time then ATI. Just look at Intel vs Amd TDP, and how they have turned.

Also noone doubts that the 280gtx needs more then a 4870, but the question is how much more.

[Truckchase!]

Man, this thread has gone downhill fast.

Sometimes it seems like nearly all threads on this board (save the xmods section) are lost by page 2.

Can anyone speak to retail sightings of this board? I've checked two best buys and a Microcenter in MN and have had no luck....

[SKYMTL]

Sometimes it seems like nearly all threads on this board (save the xmods section) are lost by page 2.

Can anyone speak to retail sightings of this board? I've checked two best buys and a Microcenter in MN and have had no luck....

I actually sighted one at a store here in Montreal today.

[vengance_01]

I just want one 4870. Should be great for me at 1680x1050. Here is to hope that prices will fall closer to the 250 range.