Best video card to use, and can the CPU also be used more to expedite processing?

Topaz Video AI General
1

Hi! I just purchased Topaz Video AI (v3.2.7) after finding it does the best job of upscaling content saved from UK (PAL 576-line format) DVDs, and with the least amount of issues (via “interlaced progressive” mode) to 1920x1080 w/pillarboxing…

My hardware:
AMD Ryzen 5950x w/64GB DDR4-2400 RAM
AMD 6700xt video card w/12GB RAM

Neither is overclocked.

I configured Topaz to run one process at a time, so it wouldn’t thin resources between multiple render jobs in the queue.

GPU usage is 99%.
CPU usage is 15~29%.
Both cite ffmpeg as using the CPU and GPU the most.

Would adding a 6950XT card and using both it and the 6700XT in Crossfire configuration improve rendering time?

Also, is there a way my 16-core Ryzen can be used more as well?

Side-question: I’ve yet to pick up blu-ray authoring software. The test file I rendered is 60GB using MKV. I’m new to this and am guessing that authorizing software will re-encode and compress into the file format that set-top players use?

Thanks!

2

No, 6700XT would slow down the process.

Switching to 3090, 4080, 4090 would increase the speed by a lot.

just look at the benchmark forum.

3

Video AI v3.2.X - User Benchmarking Results - Topaz Video AI / User Benchmarks - Topaz Community (topazlabs.com)

I had earlier. Didn’t find anything on Crossfire configurations, and looking for confirmation of Crossfire having any benefit of adding another AMD card next to the existing AMD 6700 since paying $600 for a second video card is bound to be a lot cheaper compared to the $1500~2000 for a new nVidia x090 (never mind the refurbished units for less will probably fry after a week, depending on what it was used before being returned for a refund – either way, graphics cards’ prices are through the stratosphere.)

Will re-read it; maybe the x070 is a comparable $600 cost but at a better framerate…

Just glanced; it looks like the 7900 is comparable to the 3060 with 1920x1080 source material. Will manually sift through a few more pages, wish there were an easier way to pull down criteria, but if my 6700 isn’t that much lower, then it’s not worth the time to add to or replace the GPU at all considering the prices right now. The 3090 and 4090 chipsets look great, as do the 3080 and 4080 models, but the costs ($1100+) are still higher than the hope of adding a second card in Crossfire configuration, even if the new card has to pare down its bus to match the slower card.

Still, would be sweet if ffmpeg could also use the local CPU as well as GPU for optimal throughput, unless that incurs too much of a bottleneck…

5

What is a rough estimate of speed improvements from a 1080TI to a 3090TI? Double, triple, 25%? Just curious of a rough estimate. Trying to justify an upgrade to main rig, and sidelining another GPU for TOPAZ. Thanks for any info as to potential speed improvements from 1080TI to 3090TI.

1 Like
6

You could have 4 GPU’s but VAI will only use one.

1 Like
7
1 Like
8

crossfire was used in Games and still is around, although not mentioned anymore by AMD anywhere and not a lot of games support it nowadays. Forcing Games to use it often results in garbage outputs, flickering, lags, etc… But it still exists…

BUT:

Crossfire is not a computational thing, its for gaming.

Having said that - you can simply use two cards in TVAI - don´t switch on crossfire on the driver level, so windows “sees” two cards and just select them in TVAI.

BUT:

You should use two very similar cards. Both are fed at the same time - if you have one from a different architecture (mixing AMD and NVIDIA for example or mixing GCN 2 with RDNA2) will result in enormous slowdowns or even errors or crashes. I just recently posted a few examples of running two FURY X - works fine…

Having said all that: At the moment, Tensor Core support is quite good as of today - so you might encounter that getting one single RTX Card might outperform your dual-card setup.

In general, speeds depend on a few factors - which can be observed by looking at benchmarks of the same card in different systems and settings - sometimes they vary by quite a lot… So the answer to “is it worth it” or “whats better: two small cards or one bigger one” is often: “it depends”…

For example: In multiple GPU Setups - lower resolutions tend to be mostly immune to speedups on cards with a bigger cache… The higher the resoluton and scale factor, the better the speed gain (in generall)…

Things like pcie-lane, hardware-scheduling, etc… start to play a role in certain scenarios…

In the field of “AI processing”, things also can change quite quickly… At the moment, the Tensor Core support outperforms everything else on the market - but with RDNA3, AMD has introduced “tensor like cores” - once support matures for these, we might see a passing of the speed crown once again. Or not - no one knows… Intel also hast these kind of cores, the theoretical speed of an arc770 actually is quite good for the money (and other usecases show the potential) - but the software side has to mature also…

1 Like
9

no problem to use mutliple GPUs.

Some combinations make no sense, some don´t work…

But it is possible, its implemented and I have run multiple tests and benchmarks over the years…

10

depends on resolution and scale factor…

But one thing is sure: While the 1080TI still is quite a capable gaming card - it is neither capable of FP16 computation nor is it equipped with tensor cores…

With TVAI and alike, we definitely must look at other benchmarks for speed figures :slight_smile:

so you will be getting way more than 25% … more like double, tripple… etc…

head over to the benchmark thread, you will find a lot of numbers there.

1 Like
11

Care to share how you have been doing this for years in TVEAI and TVAI ?

12

I have shared some examples over at the Facebook group when this forum wasn´t used for TVAI a lot back then.

The basic principal is really simple:
1
2
3

I could do a long tutorial of the tricks and bells and whistles, but not enough time, so I just screenshot the machine I am sitting in front of… (Attention: This is no optmized setup I am using, its my “hit and run testbench” which gets torn appart every few days and has totally messed up installations … The cards actually only run at 50%… So take the figures shown with a grain of salt - there is room for improvement, I just sho them to illustrate the procedure and show the basic principal)

just select “all gpus” in the preferences.
load your video
do your thing :slight_smile:

of yourse, there are some tricks that make this run better or not.

  • depending on the card it can be helpful to manually force tvai to use a FP16 model instead of FP32 (by editing the JSON Files manualy)
  • lowering the amount of threads seems to help in most situations when using multiuple cards (in TVAI 2,x the “reduce load” thingy thit exactly that.
  • The cards should be of compatible architecture (in this case its GCN 3.). Some stuff can be mixed, RTX plus a GCN2 card is a bad idea, but a VEGA 56 and a VEGA 64 can do fine… Speeds and VRAM should be equal or at least sufficient…
  • one of the cards is used as display/for windows/other apps - having the possibility to chose individual cards would open up the possibility to dedicate two cards for computation and one for display (e.g. an iGPU)…

So the principal functionality is there and AFAIR has worked on and off better or worse from TVAI V 2… I back then did even run it on two GTX 750TI on a Socket 771 Dual Xeon System (Core2Duo Generation) with three 2GB Maxwell Cards :slight_smile:

The topic has been covered here and there but does not get too much attention. Its already usable, with a little polishing it can come in handy for some people in some cases.

Ah, to show what speeds are on one card:

4

On thing I mentioned, but to make sure: This is a 4GB Card (two in the first screenshot used) from GCN 3 Generation - and all I did here is I edited the JSON File to use FP16 Models and lowered the Thread count… This makes FHD usabel on 4GB VRAM and speeds things up a little… If you happen to have a 6GB Card, you might not even need to alter anything and simply check “use all GPUS” and are good to go… Or use smaller resolutions.

4 Likes
13

Thanks for the detailed post outlining everything.

14

Much thanks to EVERYONE for their responses!

I ultimately got an Asus TUF nVidia 4080 as my mainboard is also Asus. I’m opting to stay with the same brand name for the sake of those blingy little features you don’t get by mix’n’matching brands…

Yeah, the performance difference is more than 4x faster, with a 768x576@25 → 1920x1080@30 going from ~9 hours down to ~2. The word “wow” comes to mind.

If converting to 60FPS, then the video took just under 4 hours, but the results were grotesque, so back to 30FPS I go. So ~2 hrs per episode it is. Still, “wow” applies.

What amazed me more is that 4080 GPU load hovered at 80% and stayed under 60C, whereas the old 6700 used 96~7% load and got up to 65C.

CPU usage doubled (from 20 to 40%), but that’s to be expected, considering the massive data bus of the GPU.

The Cooler Master is a large PC tower case I bought a decade ago. Despite its size (24Lx10Wx18H), I had to remove some 3.5" drive bay mounts as the card was too long, by 5mm.

As much as I wanted the 4090, paying 50% more for what appears to be 30% higher performance over the 4080 wasn’t worth it, though one can fathom secondary costs (e.g. electricity usage) that make up for it in time. Also, the Asus brand’s 4070ti was $200 less than the 4080. PNY’s was $400 less, but I’d have to wait 3 weeks. Gigabyte was $70 less than the Asus, but - it’s not too much of a difference compared to the Asus branding, and - dang - the Asus feels robust and well-built.

And, of course, power usage/system wattage - the card will undoubtedly make up that difference after a year or so, depending on usage…

Thanks again for the advice! Still surprised by this thing…

2 Likes
15

You can lower power usage without delaying rendering times. Check this post in the Topaz Facebook group for reference… It’s a Private group so you need to ask to join… Topaz Video AI | Facebook

Screenshot 2023-06-18 192035
Screenshot 2023-06-18 1920351430×906 151 KB

3 Likes
16

Nice, thanks!!!

1 Like
17

Hey guys,

I’m using TVAI on my M1 Pro with only 16GB ram, think the speed is quiet decent for the power usage but it’s really slow compaired to a recent RTX. Thinking about buying a RTX 4070 TI. I went throught the latest benchmarks and think that it will give me the best “bang for a buck”, at least here in Switzerland. Is it possible that someone, hopefully the maker of this video, could share it somewhere without the need of having a FB account? Would be really cool. I want to run the RTX in a external GPU enclosure connected to a small Elitedesk which I don’t use at the moment.

Hope that someone of this private FB group reads this and could give me some hints how to run the RTX the best way.

@dpcole72-12989
sorry to use your topic for that :slight_smile:

Thanks,
Patrick