Beware of GPU accleration (not games)

I just thought I'd throw this in for fun.
Is gpu acceleration really as good as they make it out to be? I'm talking mostly about video decoding here. Yes, it uses less cpu power, and the gpu can probably do it more efficient and faster than the cpu, but here's the catch.
The gpu draws more power unless you have an energy efficient card. And most high-end cards are real power suckers. For example, a 9600GT can eat up to as much as 100W. That's a lot.

So I made a test. I decoded a 720p H264 movie using the cpu and the gpu.
And in average, the gpu used ~4W more than when using the cpu.
There you go. Don't believe in everything they say. Don't blindly rely on gpu acceleration.

How exactly did you compare and measure? With what software and hardware?

Same clip, same scenes, with and without CUDA (using CoreAVC) and looking at my energy meter.
Player was Zoom Player, using AMD Athlon II X2 250 (3 GHz) and Geforce 9600GT.
CPU was running at 800 MHz, consuming around 6W according to ASRock IES.

Energy meter connected to the wall outlet?

What's the watt reading using CPU and GPU? Also, how long did they take to decode the same video?

And what hardware are you using? (CPU and GPU)

It wasn't that precise a reading. I simply took a couple of seconds of measurement.
One was with the CUDA option enabled and one without. With CUDA, I can clearly see that cpu use is lower, albeit not that much.
And yeah, the energy reader is connected to the wall outlet. It actually measures a whole lot of stuff, where the monitor, computer and a lot of other stuff on my desk is connected, but I took relative measurements.
~227W (CPU) against ~231W (GPU).

Unfortunately, I only have a 65W TDP rated processor. I'd have loved to get a 45W, but they were essentially sold out. But if I had one, the number might have been even lower.

I have a 5.5W TDP CPU :D.

I think the more important question is, how long did it take to decode the video?

If it takes the GPU only half the time it would take on the CPU, that means it will have to consume more than twice the power to make it actually consume more electricity (energy), because, of course, energy = power * time.

You can't look it that way, because it's a video. So what I meant is that I basically let the video play a few seconds while watching the energy meter.
So consider that since it measures Watts (ie joules per second), then when I get a higher wattage reading when using CUDA, it means that even if the GPU completes faster, it still consumes more power doing it.

Hehe, a 5.5W TDP CPU can't be good for performance, though.
Even though this CPU is a little old (well, old architecture, really; newer cpus are too expensive), it packs plenty of power, even at a mere 800 MHz.

Ah, so it's decoding for real time playback. Not converting a video to another format (that's the impression I got when I hear "video decoding").

Then I agree. It may not help if the CPU is fast enough. In most cases where GPU video acceleration is designed for, though, the CPU is not strong enough, and would introduce skipping. GPU acceleration enables cheap computers with low-end processors to play high resolution videos with inexpensive low-end cards (a 8400 can be had for ~$30, and I think ATI cards are even better for video accleration).

My 5.5W TDP CPU is an Ultra Low Voltage Core 2 Solo 1.4ghz. Roughly equivalent to a Pentium 4 at 2.8ghz. The dual core version (with same frequency) is 9W TDP.

Yes, I would agree it's helpful in such situations. But you keep hearing about how gpu acceleration is something "good" to have (or at least that's the impression you get), and that's basically what my point was - don't take it for granted. Gpu decoding is not necessarily better (doesn't mean it can't be, though).

Core 2 Solo :(
I hate single core cpus...

Yeah, Vista runs a bit sluggish on this laptop. Linux runs just as smooth as on my desktop, though :D.

Otherwise... the single core is not as bad as I thought. Except compilations take longer since I can no longer compile with 2 threads. I don't do any heavy work on my laptop, though. For that I can connect back to my desktop with a C2D overclocked to 3ghz. On the other hand, it's loaded with benefits. 8 hr battery life, and the laptop is cool to the touch all the time, and the fan doesn't start until you put the CPU at 100% for a minute or 2. I even forgot to turn it off once (since it was too cool), and put it in my backpack. It was only slightly warm when I took it out 3 hrs later, with 70% of battery left. Lucky it has a SSD. The harddrive would probably have been killed since I don't treat my backpack nicely at all.

Oh yes, I have a similar ... netbook. But I hate the lack of processor power.
But it sounds as if you need to upgrade to Win7 ;)
Lucky me. I get it for free, though :)

Haha. Mine came with Vista Basic, so no free upgrade to W7 :(.

Lucky me, I get Linux for free :D.

A ULV C2S is about half way between Atom and "normal" processors.

Ha, mine came with XP.
But Atom is seriously crap. It's not even 64-bit. It's really infuriating.

Ah, so school thing?

I can get Windows Server 2008 for free, but will have to tune it to a non-server OS. I heard I can get W7 for free if I join IEEE as a student member, too, due to a contract between MS and IEEE. I will stick to Linux for now, though. The only thing I would get Windows for is battery life. My laptop can run for ~10 hrs on Windows, but "only" ~8 on Linux. I think the Linux drivers for my hardware are not as optimized for efficiency yet, since the hardware is pretty new, so they probably just tried to make it work first.

Yes, school thing. Lots of software... FREE.
And you get 10 hours on Vista? I wonder what you would get on Win7, since it's actually even more efficient on battery than XP.

That's pretty cool, Elysia. I think we tend to accept vendor claims pretty much on face value. I wonder how far from the truth many of those claims would turn out to be if people actually put them to the test?

I get the feeling you're being sarcastic here...

The GPU can be used for AI, as a raw vector math CPU, physics, and just about every math intensive operation out there. The main advantage to the GPU is it can access its local on-board video memory which means it can use its own bus which is independent of the speed of the system bus. So if you upload all the data to the GPU and allow it to operate on it, it will be faster than the equivalent operation on the CPU simply due to memory access speeds. Plus this means the CPU can twiddle its thumbs or do whatever it wants to while the GPU works on something else.

Your test is highly dependent on how that software uses the card to do the work, the model and manuf. of your video card, the internal memory interface (128/256/512 bit?), the amount of video memory that is onboard, and the speed of the video card memory bus.

Quote:

---

And you get 10 hours on Vista? I wonder what you would get on Win7, since it's actually even more efficient on battery than XP.

---

Well, Vista eats ~6W for the whole system (including LCD with LED backlight and wifi, calculated from battery reading), so I'm not sure how much lower it can get. Linux eats ~8W.

It's just a regular 13" 60Wh battery. The long battery life comes from low consumption, not big battery.

Quote:

---

Originally Posted by Bubba

The GPU can be used for AI, as a raw vector math CPU, physics, and just about every math intensive operation out there. The main advantage to the GPU is it can access its local on-board video memory which means it can use its own bus which is independent of the speed of the system bus. So if you upload all the data to the GPU and allow it to operate on it, it will be faster than the equivalent operation on the CPU simply due to memory access speeds. Plus this means the CPU can twiddle its thumbs or do whatever it wants to while the GPU works on something else.

Your test is highly dependent on how that software uses the card to do the work, the model and manuf. of your video card, the internal memory interface (128/256/512 bit?), the amount of video memory that is onboard, and the speed of the video card memory bus.

---

Of course. It was never meant to be an extensive test. Take it with a grain of salt and know that gpu acceleration is not always better. For games, obviously it is, though.

Quote:

---

Originally Posted by cyberfish

Well, Vista eats ~6W for the whole system (including LCD with LED backlight and wifi, calculated from battery reading), so I'm not sure how much lower it can get. Linux eats ~8W.

It's just a regular 13" 60Wh battery. The long battery life comes from low consumption, not big battery.

---

I believe you'll see those figures scaling when the system in idle on Win7.
I haven't used Vista, so I can't use it as comparison, though. But I know that Win7 beats XP in battery performance.

Quote:

---

Originally Posted by Elysia

I just thought I'd throw this in for fun.
Is gpu acceleration really as good as they make it out to be?

---

It can be. The GPU can turn a compute bound task into an I/O bound task. They have awesome floating point horsepower (single precision only in most cases), but they have crap memory bandwidth. So if you are doing things cryptanalysis, weather modelling, FFT's etc where you are doing a lot of operations on a (relatively) small amount of data they are great. In my experience though, unless you are doing at least 12 floating point operations per byte of data you will not see a performance increase.

They can be and they can not be. My message is: don't take it for granted that GPU acceleration really is better. It has its advantages and disadvantages. Do not forget.

Oh yeah, I totally agree. They aren't a magic wand.