Video Card--I want a MINOR upgrade. How to assess?

Robert_341

Commendable
Jan 10, 2017
34
0
1,530
I found this excellent website for initial comparison:
http://www.videocardbenchmark.net/high_end_gpus.html

but what it doesn't tell me is how fast a gpu will encode or transcode a video. ROUGHLY speaking...is it fair to assume that a card benchmarked twice as fast as another will encode ROUGHLY twice as fast?

Also...a card I would really like does NOT list the number of CUDA cores or stream processors but touts how its good for Virtual Reality. Does that automatically mean it will also be good at transcoding...or not so much and its meant for frame speeds and shading and such only? I know its all related.

Is there any website that shows encoding speed relative to "real time?" It seems like such a common sense standard.... but the info is hard to find.

Thanks
 
Solution
Performance is relative. A Rx480 will hang with a gtx1060 all day long yet doesn't use a single Cuda core. If the software doesn't really use them to advantage either, then even the vaunted power of a pascal titan isn't much better off than a GT710 in that regard. It's well known that building a pc around 1 specific game is the height of stupidity (spoken aloud or not) so building around 1 specific software can't be any different. The best builds take everything into consideration
Most transcoding is still done on CPU and that's why there is no benchmarks for it. I know AMD makes their own tool, and Nvidia might, but it's limited and no uses them when good open source and other programs that aren't limited exist, but handbrake encoding is CPU as far as I know still.
 

Karadjgne

Titan
Ambassador
Doubtful you'll find any at all. There's a huge difference between games and encoding. And not what ppl automatically assume. The same 5 or 6 games have reached recognition status as challenging in a certain gpu area, so many use those games to show relative fps etc. It's a quantifiable number. Encoding is a lot more limited. There's no recognition to base tests on. Encode what, using what software, on what gpu with what cpu. There's just no recognised standard since there are so many variables that can change entire results. Using Sony Vegas with an AMD card might net great results, using an nvidia card less so. Swap that around in handbrake and the results might be totally opposite. So what wins? All it would tell you is use AMD for SV and use nvidia for handbrake. Still no quantifiable numbers
 

Robert_341

Commendable
Jan 10, 2017
34
0
1,530
Thanks getochkn: I'm no expert, but my own experience is software encoders have an "option" to use CUDA if your card has it....or cpu encoding by default. I use Xilisoft, a free transcoder, that is fast and acceptable. Everytime I have tried handbreak, it takes significantly longer for the same encode. "Maybe" that makes the final product better quality, but I haven't noticed that.

Anyway, I'm about to build a VR system and want to upgrade my old system to do transcoding only and the only cheap way to do that is to upgrade the video card. My current card has 96 Cuda Cores and transcode about 125% of real time. I'm thinking a 400 Cuda Core card should lower the transcode time to 30% or so? but even 50% would be worth this $50 upgrade. BUT: I fear I might spend the $50 and get a card that transcodes no faster than my current card? I can't find anything in writing to give me comfort.
 

Robert_341

Commendable
Jan 10, 2017
34
0
1,530
thanks Karadjgne--I see the "challenge" but the industry provides guidelines of "benchmarks" as I liniked to. Of Note: I'm curious using exactly the same everything....if I swap out a 96 Cuda Core card for a 400 Cuda Core card .... how much improvement can I reasonable expect? Now: in fact...the 400 Cuda Core has faster and more memory and bandwidth and clock speeds and so forth. The two cards will be run on the same everything else. So...what is reasonable to expect? People seem willing to advise that getting a GE Force 1080 card will improve the transcoding time. After two years.... no one will comment on my lesser wanted upgrade of X4 Cude Cores.

What would you guess?
 

Karadjgne

Titan
Ambassador
Here's my guess, it's flawed logic, but it's how I think. Take an plain Intel pentium. 2 cores. Take an fx 8350, 8 cores. By deductive reasoning, that 8 core should blow the pentium away in every test. But it's not so. Clock for clock, the pentium is faster, every time. IPC doesn't lie. So it'd stand to reason 400 Cuda cores would be much faster than 96 at encoding. But is it so? How much does bandwidth affect the results? Does the software actually use all those Cuda cores or is it limited to 100 or less? Does the gpu just use the Cuda cores as flash memory while shunting the bulk of actual encoding on the processor? How much does system ram play into transmission of data to-from the gpu-cpu? How much, if any, amount of Cuda cores relate to stream processors and gflops. Are Cuda cores stronger or weaker vrs them?

I don't have the answers, don't know anyone who does. What I do know is that for VR, you are looking at an Rx480 (490 if it ever gets here) or a 1070 / 1080. How that translates backwards for encoding, I dunno, there's far to many unknowns. All I can suggest is if you have to jump, play the frog and do it. Trying to pin down specific answers where there are so many 'ifs' is a fools errand. Maybe a stronger nvidia card will be better, maybe it won't, maybe amd is the way forward relying on gflops over Cuda. Maybe it isn't. And just how much is anyone's guess and probably more accurate than mine.
 
I've done some video transcoding, using Nvidia Nvenc and CUDA solutions. Due to various limitations in the selection of programs that utilize hardware acceleration at all, and issues with the ones that do, I've found that it's still better to do it in software and have your CPU handle it. This means my recommendation at this time is to get the most powerful CPU you can, not invest in a videocard for this purpose. My understanding is AMD cards use OpenCL for this kind of thing and support there is even more limited at this time.
 

Robert_341

Commendable
Jan 10, 2017
34
0
1,530
Grandmaster: thanks. I didn't realize that CUDA cores could be so irrelevant in determining transcoding speed. The fact that comparative information is about impossible to find.... makes me think this is very true. Spending money "just" to find out in this INFORMATION AGE...especially when the info is about computers.... is a bit frustrating?

Don't Listen: I've read the articles and yes they all conclude that cpu encoding is superior to gpu or even quick sync on the Intell chips. I've looked at the demonstrated still shots to show the difference .... and I just don't see much difference there, or a difference to "care" about? That could be something that becomes noticeable over time though. I think I'll do a quick 5 minute video with and without the gpu on my current machine....but I''ll need to send it to my big screen cause everything looks good on my small monitors which is where I did the test last time.

Just not enough time in the day.................and never enough money to spend: just to find out?
 

Karadjgne

Titan
Ambassador
It's a numbers game. And most of it is BS. A 10% increase is comparatively nothing gained. If encoding a video used to take exactly 1 hour, now it's done in 54 minutes. But since every encode is totally different, the next one could be 1 hr and 6 minutes or with that 10% increase, 1 hr and seconds. You hit that encode button, walk away, smoke a cigarette, get a drink, hit the bathroom, returning an hour later. Can you say honestly that there's any change having sped the process up by 6 minutes on something taking 60? Or shaving 3 minutes of something that used to take a half hour? Benchmarks will make a big deal over a 10% difference, that's huge performance gains. In reality. Not so much. When the monitor is locked at 60Hz, would there be a difference between a gpu capable of 100 fps and a cheaper mid grade capable of 90 fps at the same settings? Benchmark says huge disparity, reality, they are the same.
You could spend thousands on an x99 setup, just to get that benchmark 10% increase and that sounds like a huge gain. Reality is a pc costing a fraction of that will do the same job. You'd need 50% or better increase to make the cost justifiable.

So it all boils down to getting equipment to do the job. Which is a VR capable gpu.
 

RobCrezz

Expert
Ambassador
More cuda cores will almost certainly improve the performance. Bear in mind that it can depend on the architecture of the card, not all cuda cores are equal.

1000cuda cores on the current Pascal generation will likely be faster than 1000 cuda cores on the previous Maxwell generation.

If your card only has 96 cores, its either very old or very low end (or both).
 

Karadjgne

Titan
Ambassador
Performance is relative. A Rx480 will hang with a gtx1060 all day long yet doesn't use a single Cuda core. If the software doesn't really use them to advantage either, then even the vaunted power of a pascal titan isn't much better off than a GT710 in that regard. It's well known that building a pc around 1 specific game is the height of stupidity (spoken aloud or not) so building around 1 specific software can't be any different. The best builds take everything into consideration
 
Solution

RobCrezz

Expert
Ambassador


I respectfully disagree.

If the machine is going to mainly used for one specific piece of software, then building the specs to suit is a good plan.
 

RobCrezz

Expert
Ambassador


And how can you build a system to suit unknown new version or different software?