Benchmark System and Measurement Methodology
Tom’s Hardware Germany’s power consumption and performance benchmark system for consumer and workstation graphics cards was designed in cooperation with HAMEG (Rohde & Schwarz) to allow for very precise measurements. We put it together in several steps.
Only professional technology can handle the new challenges posed by AMD’s PowerTune and Nvidia’s Boost technologies and their massive load fluctuations. This is why we’re using a 500 MHz digital multi-channel oscilloscope, the HAMEG HMO 3054. It also lets us save all the measurement data in one place and brings a handy remote control function to the table.
We measure the currents with three calibrated DC current clamp probes (HAMEG HZO50). Two of them, 3.3 and 12 V, take their readings at a custom-made riser card, which can reliably pass PCIe 3.0 signals, and one of them at a specially-modified PCIe power cable. All voltages are measured at the single-rail power supply, which we slightly modified to allow better access.
The voltages are measured directly with a modified connection to a single-rail power supply. Our time resolution is now a mere 2 ms, which can measure and log all load transients incurred by AMD’s Power Tune and Nvidia’s GPU Boost. In order to keep the volume of data manageable, we limit the duration of a test run to two minutes.
| Test Method | Contact-free DC Measurement at PCIe Slot (Using a Riser Card) Contact-free DC Measurement at External Auxiliary Power Supply Cable Voltage Measurement at Power Supply |
|---|---|
| Test Equipment | 1 x HAMEG HMO 3054, 500 MHz Digital Multi-channel Oscilloscope 3 x HAMEG HZO50 Current Probes (1 mA - 30 A, 100 kHz, DC) 4 x HAMEG HZ355 (10:1 Probes, 500 MHz) 1 x HAMEG HMC 8012 Digital Multimeter with Storage Function |
| Test Bench | Microcool Banchetto 101 |
| Power Supply (PSU) | 1200 W, be quiet! Dark Power Pro 10 |
Power Consumption At Idle
Let’s first take a look at a normal desktop environment without any applications running in the foreground before and after modifying Gigabyte's GeForce GTX Titan Black.
The total power consumption is basically identical. However, distribution between the PCI Express slot and auxiliary connectors changes a little. This phenomenon can often be seen with Nvidia graphics cards at idle if the GPU temperature falls under 30 degrees Celsius. The additional 100 mW are probably due to the additional fans and can generally be ignored.
If you find the following two (very detailed) charts difficult to read, focus primarily on the bold red line, which indicates total average power consumption.


It’s a lot more important to observe how the GeForce GTX Titan Black GHz Edition fares under load before and after the modification. It goes without saying that we collected those numbers using the same high resolution. The results, on the next page, are certainly interesting.
- A GeForce GTX Titan Black You Modify Yourself
- The Gigabyte WindForce 600 Graphics Card Cooler
- Upgrading The Gigabyte GeForce GTX Titan Black
- Dimensions And Pictures: The Upgraded Gigabyte GeForce GTX Titan Black
- Power Consumption: Test Methodology And Idle Measurements
- Power Consumption: Gaming And Full Load Measurements
- Temperatures And Noise
- Performance
- Gigabyte Gets Its WindForce Cooler Right


If you bothered reading the first page you'd know why.
If you bothered reading the first page you'd know why.
"Nvidia doesn’t allow its partners to sell the GeForce GTX Titan Black with proprietary cooling. However, Gigabyte now offers a GHz Edition of the card that comes bundled with its WindForce solution, which you can install on the overclocked board yourself."
This one right? whats the difference between you install it yourself and Gigabyte take the initiative pre-factory installed? or Warranty Void?
If you bothered reading the first page you'd know why.
"Nvidia doesn’t allow its partners to sell the GeForce GTX Titan Black with proprietary cooling. However, Gigabyte now offers a GHz Edition of the card that comes bundled with its WindForce solution, which you can install on the overclocked board yourself."
This one right? whats the difference between you install it yourself and Gigabyte take the initiative pre-factory installed? or Warranty Void?
Ok, for your better understanding:
Nvidia doesn’t allow its partners to sell the GeForce GTX Titan Black with factory-installed proprietary cooling.
Igor Wallossek, I wonder if you could put up a graph for 3D rendering? If you use Blender's BMW scene by Mike Pan (a popular benchmark scene), make sure you properly set the tile size!
Ok, for your better understanding:
Nvidia doesn’t allow its partners to sell the GeForce GTX Titan Black with factory-installed proprietary cooling.
Silly question probably, but why does nVidia allow only EVGA to break this rule, with their hydro copper signature edition you mentioned? Is it just because it's a water cooled model? Do you think nVidia specially signs off on the design?
I'm genuinely curious.
that's what it looks like to me.
> Igor Wallossek, I wonder if you could put up a graph for 3D rendering? If you use
> Blender's BMW scene by Mike Pan (a popular benchmark scene), make sure you
> properly set the tile size!
Arion Bench 2.5.0 would be a better test, because it scales perfectly with
multiple GPUs.
Or the AE CUDA test my friend has created, but it's pretty intense, maybe
takes too long with just one card (about 20 minutes with a single 780Ti).
Ian.
Arion Bench 2.5.0 would be a better test, because it scales perfectly with
multiple GPUs.
Or the AE CUDA test my friend has created, but it's pretty intense, maybe
takes too long with just one card (about 20 minutes with a single 780Ti).
Ian.
I agree. The BMW scene is not the best CUDA benchmark. I just didn't want them to mess it up if they decided to use it. I heard some people complaining about this benchmark, although I don't know if they were right or wrong.
My Titan Black renders the BMW in just over 24 seconds!
For comparison, an Intel Core 2 Duo @ 2.33 GHz took 16 minutes!
Have you run a Titan Black on that AE CUDA test? If so, I am curious to see the results!
wolverine96 writes:
> I agree. The BMW scene is not the best CUDA benchmark. ...
I've tried it numerous times with various setups, it just seems to behave a
bit weird IMO.
> My Titan Black renders the BMW in just over 24 seconds!
Main problem I find is I can't work out how to make it use all available GPUs.
Is that possible? One of my 580s does it in about 43s, but my system has 4 of
them, so it's a bit moot really. Mind you, I'm using an older version of
Blender (2.61), stuck with it to ensure consistent CPU-based testing.
And as you say, it also involves some CPU stuff (scene setup doesn't use the
GPU).
> Have you run a Titan Black on that AE CUDA test? If so, I am curious to see
> the results!
Alas no, atm I don't have access to anything newer than some top-end 3GB GTX
580s (MSI LX, 832MHz); my system has 4 of them. Final version of the test file
takes 14m 48s to render in AE using 16bpc and 8 samples (ie. just average
quality), so on a Titan Black I'm guessing it would take maybe 25 mins? Hard
to say. Would certainly be interesting to find out. Note the 'max' quality
setting would be 32bpc and 10 samples (likewise, for the full animation, avg
quality is 1080p @ 25Hz, max quality is 50Hz).
I'll sort out the test readme, download archive, web page, etc., next week,
but need to talk to C.A. first about some things. Anyway, here's the rendered
image in original Targa format (just one frame, the default test frame 96, the
last frame in the main animation sequence):
http://www.sgidepot.co.uk/misc/cuda.101_Frame96.tga
Here's the file converted to BMP and SGI IRIS/RGB:
http://www.sgidepot.co.uk/misc/cuda.101_Frame96.bmp
http://www.sgidepot.co.uk/misc/cuda.101_Frame96.rgb
and for those who don't mind losing a bit of quality, here's a 100% JPEG:
http://www.sgidepot.co.uk/misc/cuda.101_Frame96.jpg
The full 4 second animation takes hours to compute even at average quality and
is thus intended more as a stress test for those interested in checking that
their system can handle long renders or other GPU tasks without falling over
(I've seen many people asking for a test like this on forums). I suspect at max
quality the whole sequence would take about a week to crunch on my system.
Also interesting for exploring power consumption & energy cost issues for
different GPU configs (load draw on my system during the render is around 920W).
Ian.
Your system with 4 GTX 580's is much faster than mine! (Two GTX 580's is about as fast as one GTX Titan Black.) I guess the only time mine would be faster is if the scene used more than 3GB of RAM. I actually was planning on getting 2 GTX 580's, but then I discovered the Titan Black.
Is that Cycles in those images you posted?
By the way, Blender 2.71 is coming out very soon. In the past 10 versions, there have been some major performance gains for Cycles. I think it's like 30-50% faster in some cases.
> Did you say you are having trouble getting multiple GPU's to work? I only use one
> GPU, but here's a very informative link. More specifically, see this section.
Thanks!! My goof, looks like V2.61 doesn't have the Compute Panel. Will try
the newer 2.70a in a moment... (downloading now)
> Your system with 4 GTX 580's is much faster than mine! ...
Yup, though I suspect your power bill is less likely to make your eyeballs explode.
> ... I guess the only time mine would be faster is if the scene used more than
> 3GB of RAM. ...
I had been hoping we'd see 6GB 780Tis, but seems like that's been killed off. Shame.
> I actually was planning on getting 2 GTX 580's, but then I discovered the Titan Black.
The real advantage of multiple 580s is just low upfront cost. Standard 580s are pretty
cheap (I have four 1.5GB 797MHz models which cost about 400 UKP total), if one's ok
with the VRAM limit. 3GB 580s cost a bit more, but not much more (I've bought/sold
nine Palit 3GB 580s in the past year). The MSI LXs though can be a tad pricey, depends
on luck really I guess. I got mine (five total) for good prices though overall, and they do
oc like crazy (1GHz+ is possible).
> Is that Cycles in those images you posted?
No, it's the RayTrace3D renderer within After Effects.
> By the way, Blender 2.71 is coming out very soon. In the past 10 versions,
> there have been some major performance gains for Cycles. I think it's like 30-50%
> faster in some cases.
Good that they keep boosting it, but a nightmare for benchmarking consistency.
Ok, download done, quick test...
Cycles does the BMW in 11.56s (blimey!), tile size 240x135. Just curious btw,
you mentioned using 512x512 tile size, but surely it'd be optimal to use an even
divisor of the image dimensions in both X and Y? What do you get if you try
a tile size of 240x135?
Ian.