Recommend a graphics card for video enhancement (Upscaling)

[MalickT]

I currently have Radeon RX 560. It runs 1280MHz on overclocking mode. It has H265 and H264 support.

I use Topaz Video AI Enhance to upscale my 1080p videos to 4K. Im impressed with the result but it takes 7 days for 2h video. About 3 second a frame!

I do not play any games. I only want faster processing so it is the GPU clock frequency that matters, right?

I remember that my card cost only 170$. I have looked for cards that cost around 500-800$ but still the specs say about 1400Mhz GPU. Only 200Mhz faster than my current card. Im confused.

Im willing to pay 500-800$ for a card that can do the job for 1-2 days instead of 7 days.

Any recommendations?

[QBhd]

clock speeds mean nothing in today's processor world. IPC (instructions per clock) is the name of the game today. And there is no real value associated to that. Also there almost is no $500 cards anymore. I could sell my current Sapphire Nitro+ RX 5700XT and probably make a profit on it even now!

QB

[Asmodian]

Originally Posted by MalickT

I remember that my card cost only 170$. I have looked for cards that cost around 500-800$ but still the specs say about 1400Mhz GPU. Only 200Mhz faster than my current card. Im confused.

GPUs are VERY parallel. Most of the difference between models is the number of "cores" not the speed of each one.

The RX 560 has 896 or 1024 shaders (depending on model).
The 6800 XT has 4608 shaders, so it would offer a huge increase in speed without any changes in clock speed.

You cannot blindly compare shaders across architectures, they can be very different, but within the same architecture they are what makes one GPU faster than another.

[RanmaCanada]

Cryptocurrency mining and Covid have made current and past video cards extremely difficult to buy at retail, let alone to buy period. Even if we recommend you a card, would you be willing to pay over twice the retail price of it?

[MalickT]

So pretty much the more expensive, the better/faster?

[benwaggoner]

Originally Posted by MalickT

So pretty much the more expensive, the better/faster?

For current gen cards. There are a lot of prior gen cards around at high prices that may be worse for your application.

It really depends on how the software is leveraging the GPU. Could be VRAM bound. Could have a maximum number of cores it will use. Could have big perf improvements between generations, or not.

Does the documentation or support info make any recommendations?

[ognirats]

VSgan > topaz

[Asmodian]

Originally Posted by MalickT

So pretty much the more expensive, the better/faster?

Pay attention to the number of shaders and clock speed. Also, the memory bandwidth can be very important, but I don't know if that bottlenecks Topaz.

Comparing GPUs based on paper specs is somewhat complicated. You really need benchmarks of your specific application, but shaders and memory bandwidth are the most important things to look at if want to judge if it is worth looking for benchmarks.

[Blue_MiSfit]

This really belongs in the PC Hardware forum, not the HEVC encoding forum Moving.

[MalickT]

Yes, hardware topic but involves H265 encoding.

What about this card: https://www.zotac.com/us/product/gra...-1650-oc-gddr6

If I could get average 2 days per video instead of 7 days per video, it would be great...

[Asmodian]

The 1650 is not a significant upgrade over the RX 560. It only has 896 CUDA cores, so it doesn't have any extra cores (not that you can compare directly, but it would need more to reach 100+% faster). I don't know if Topaz is faster on CUDA, it was CUDA originally so the 1650 might offer a more significant boost than expected, but in general a 1650 is only a bit faster than a RX 560.

[MalickT]

How about Palit GeForce RTX 3060 StormX - https://www.palit.com/palit/vgapro.p...9-190AF&tab=ov

?

[Asmodian]

Yes, that would be a lot faster. 3584 CUDA cores and 360 GB/s of memory bandwidth.

[MalickT]

OR what if I buy a 800$ high end CPU with 8 or more cores? Or CPU will never be a match for GPU with H265 support?

[excellentswordfight]

Originally Posted by MalickT

OR what if I buy a 800$ high end CPU with 8 or more cores? Or CPU will never be a match for GPU with H265 support?

Well, we need to know if the bottleneck is the image processing or the video encoding, and if the encoding is using hw-encoding at all. Cause the encoding cannot be faster then the speed of the image processing (upscaling), so if thats done at 1fps, then the upper limit of the encoding would be 1fps.

I'm not to familiar with Video Enhance AI, I did a bit of googling but Topaz dont specify much at all unfortunately, so making GPU recommendations seems a bit hard; does it scales with tflops? Does it prefer CUDA? Does it utilize tensor cores? Memory bandwith intensive etc. I would direct my question to their forum instead.

If it's the video encoding that is the bottleneck, you would need to know if it does it in sofware (then speed would be CPU-dependant) or in hardware. If its in hardware the price of your GPU wont do much in terms of speed, it would more be a question of AMD vs Nvidia or the generation of the card.

[MalickT]

The bottleneck is both image processing (enhancement) and encoding.

For testing, I used a 20 second 1080p clip of 30MB size. I used no compression (lossless) when encoding. Enhanced the video into 4K

using GPU (AMD Radeon RX 560):
Time - 18 minutes
File size - 2.5GB

using CPU (AMD Ryzen 3 3200G):
Time - 4 hours!
File size - 350MB

There was no difference in video quality.

I have heard that using CPU for encoding will always get you better quality and smaller size but GPU encoding is faster but not so good quality and larger size.

So I am still in between two options: buy CPU (AMD Ryzen 9 5950X) or buy GPU (Palit GeForce RTX 3060 StormX)

[Asmodian]

That does not sound correct to me. Upscaling is very likely the slow step, encoding is relatively quick. Topaz quotes 0.4 sec for SD to HD on a GTX 1080, HD to 4K will a lot slower on a RX560.

There is a very big difference between lossless compression and no compression. What framerate was your test clip? I think you had something else different between those encodes. If you used no compression you cannot have gotten a 350MB file.

Topaz uses the GPU for upscaling, but I have no idea how it encodes the video. If it uses software encoding a faster CPU might speed it up, but it would never be faster than the upscaling, so I don't think a faster CPU is a good idea unless you already had a faster GPU.

The 3200G does have a GPU as well. Are you sure you weren't using the 3200G as the GPU for upscaling during this test? It would be very slow, even compared to a 560.

Originally Posted by MalickT

I have heard that using CPU for encoding will always get you better quality and smaller size but GPU encoding is faster but not so good quality and larger size.

This is the difference between using software video encoding on the CPU or hardware video encoding on the GPU. Hardware encoding is very fast, but generally larger for the same quality. However, this difference in size would be much less that 2.5GB v.s. 350MB you saw.

I wouldn't worry about video encoding until you can do the processing faster. A 3200G would be terrible for CPU video encoding, but the hardware encoder on the 3060 is pretty good, so that would be a great option.

[MalickT]

I got my new CPU today (AMD Ryzen 9 5950X). A little dissapointed, speed is 5 sec/frame. (Previous CPU was 20 sec/frame) So my current GPU is still faster than my new CPU (GPU does it 2 sec/frame)

For some reason the CPU shows about 35% busy when upscaling the video. I was expecting the CPU would be 100% busy. I have to take a look into it. Why isn´t Topaz not taking full advantage of my CPU...

[mastrboy]

Originally Posted by MalickT

I got my new CPU today (AMD Ryzen 9 5950X). A little dissapointed, speed is 5 sec/frame. (Previous CPU was 20 sec/frame) So my current GPU is still faster than my new CPU (GPU does it 2 sec/frame)

For some reason the CPU shows about 35% busy when upscaling the video. I was expecting the CPU would be 100% busy. I have to take a look into it. Why isn´t Topaz not taking full advantage of my CPU...

Unfortunately not all workloads scale linearly with cores, since you are capping out at 35% you could try splitting the video into 3 segments and running 3 instances of the processing software, that should be able to utilize 100% and shorten total rendering time.

[Asmodian]

Even bringing it to 100% usage, with perfect scaling, would only bring it down to 1.75 sec.

You really need a faster GPU.