As my experence, if your GPU is too powerful like 4090, the bottleneck will be the CPU. You should choose the CPU with more memory bandwidth.
1 Like
Yeah, I tried a variety of virtual machines via Paperspace and found that better GPUs weren’t making much difference past the mid-range (I think around RTX4000), probably because I couldn’t choose the other specs.
Even my older gaming laptop (i7-9750H, 2.6 GHz, 6 cores, 12 logical processors, 16 GB RAM, RTX 2080 8GB GDDR6) seemed at times slower than my Surface Pro 8 (i7-1185G7, 3 GHz, 4 cores, 8 logical processors, 16 GB RAM, Iris Xe G7). The 2080 appears to score 5-6 times higher than the Iris on benchmark tests. So I’m currently planning to prioritize the CPU in my next purchase.
I was trying out the benchmark feature and got unrealistically good speeds, which I believe you discussed on the benchmark result threads. Every machine I’ve used for TVAI has always started out with a relatively high number of FPS, which quickly declines to a fraction of the initial speed throughout the first few minutes of upscaling. On the other hand, previews and very short clips can run pretty fast.
Based on your explanation it sounds like this is when the CPU cache is filling up. Does your choice of exporting to PNGs provide the side benefit of maintaining initial speed, because the individual outputs are so small? That sounds even more appealing…although I’m not sure how slow and novice-unfriendly the process is to convert PNGs to H.265. (Thanks once again for your expertise!)
I used batch files when I started out.
Here’s a sample bat file:
"[Full path to non-TVAI ffmpeg.exe]" -hide_banner -stats_period 2.0 -nostdin -framerate 23.976 -y -i "[full path to input folder]\%%6d.png" -c:v libx265 -crf 22 -pix_fmt' yuv420p -preset slow -x265-params aq-mode=3 -sws_flags spline+accurate_rnd+full_chroma_int -vf "colorspace=bt709:iall=bt601-6-525:fast=1" -color_range 1 -colorspace 1 -color_primaries 1 -color_trc 1 "[Full output path].mkv"
Change all the paths. Double check the -framerate and -crf values. I just put the most common values I use.
The colorspace stuff… This works for DVD to FHD. Might not work for other resolutions.
I don’t think it’s that. CPU cache is for repetitive tasks and is usually in the megabytes range of size. I do think there is something that fills up. My first guess would be SSD cache, but several users have shown that their SSDs can handle the intensive writing that TVAI can do. So that might not be it.
I have noticed that dark and empty scenes process faster. Like end credits.
1 Like
Well I feel like an idiot. Now that I have a higher-end system and am taking this hobby seriously, I finally tried using ProRes HQ for a big file instead of VP9 Best. With ProRes, the speeds do not slow down at all. I’m getting a sustained 40+ fps for Proteus 480->1080, as opposed to a sustained 2.5-3.5 fps using VP9 Best. And compressing with Hybrid really doesn’t take long – though it’s slower at around 20 fps. Anyway, I’m very glad I followed your advice to compress later!
3 Likes
Drive throughput isn’t important as TVAI reads the source file at a pretty low rate (usually well below 50MB/s) so even a single hard drive can keep up. It writes in larger chunks, but as long as the drive has a normal amount of cache, that’s not an issue either.
Personally, I’ve never seen the drive be a bottleneck. Maybe some REALLY old SATA 2 drives with little or no cache would be a problem. I can’t imagine any SSD being a bottleneck. Even if you were reading a 60Mbits/second source that’s still less than 8MB/sec. and well under the average speed of SSD drives.
I bought two cheap 2TB SSDs and configured them into a stripped 4TB drive. While processing in FHD PNG files, after about two months I started to see speed decreases because everything had to wait for the SSDs. I do multiple passes in TVAI.
Because of that, I replaced them with 4 used 1TB WD Velociraptors.
SSDs do lose speed after much writing. PNG is a little extreme, but I recommend that everyone should process to some lossless format. The size differences are not going to be much compared to PNG. (Okay uncompressed tiff would be more extreme.) My little anecdote doesn’t even include UHD size files.
So my conclusion is this: If you can find SSDs that are made for heavy writing go for it, but it will eventually slow things down. (Could be months or years depending on so many things. They last longer and write faster depending on what cells they’re made with. From longest lasting to shortest: SLC, MLC, TLC [3D NAND], QLC and I’ve heard they’re making a quintuple layer cell now.)
HDDs are slower, but putting several in a stripped configuration will get the speed needed. They will have the same speed until they die.
You did TRIM the SSD regularly, right? Without regular TRIMing, writing to the SSD slows down because the memory controller has to look for no longer occupied cells. When TRIMed, all unoccupied cells are marked as available, saving the controller a lot of work at write time.
I use a Crucial 500GB SATA3 drive as my write drive for TVAI. I have 4 PCIe4 NVME drives available (2 of which are Samsung 990 Pros) but since I write to the drive and then move the files to my NAS I don’t want to wear out the more expensive NVME drives. When I started I using this SSD I had previously written about 1 TB to the drive. I have now written and erased over 41 TB and have used 4% of the drive’s lifetime. It still can write at over 300 MB/s - unless there was a video with over a 2Gbit/sec bit rate it would still be able to keep up.
YMMV
To be fair, TRIM is usually enabled automatically in Windows 10. I won’t even check (let alone do it manually) unless my Crucial SSDs drop below ~340MB/s on VERY large file copies (my main use case). That’s happened only once in the last 5 years, I googled how to re-enable it, and now I just let W10 do its thing.
1 Like
Would that do anything? (And I agree that DaveL is correct about Windows 10 and 11 being built with optimizing for SSDs in mind.) I’m writing ~2 TBs to the drive over the space of ~2 hours, then deleting it once it all gets converted back into movies. Then I do it all again for the next batch of movies. Not the typical use of an SSD. I doubt there’s a way to tell the SSD to not bother emptying the SLC or DRAM cache into the NAND because it’s going to be deleted in a few hours.
Anyway, the point is some SSDs just can’t handle the load as well or as long as others. So we cannot say that any SSD will increase speeds in TVAI.
If a drive that has been properly maintained slows down so much that TVAI can’t read or write as fast as it wants to, then there is something seriously wrong with the drive. TVAI simply doesn’t read or write that fast. Even a TVAI process that reads a 30mbit/second original at 3X real-time will struggle to reach 100Mbits/sec.
While Trim is SUPPOSED to be automatic (issued after a delete), I have seen many Windows installations (including my own Win 11 install) where write to SSDs slows dramatically and is restored after an explicit TRIM. Automatic trim in Windows has been flaky since before Windows 11 was released.
Finally, it is a fact that you can’t write to static memory as fast as you can read from it. Most drive specs only measure write processes for volumes less than the drive cache. If you over-run the cache, write speeds drop dramatically - often as much as 70% slower, including the latest and fastest NVME drives.
It depends on your export format.
Harddisk speed is more important if you output as 16bit TIFF files.
Sure…but not what we are discussing - i.e. the virtues of SSDs versus HDs. The suggestion was made that striping SSDs was required to support TVAI. My point is that any SSD should be able to keep up with a “standard” TVAI use case without issue.
And for the good of this thread, this is my current hardware config:
Intel i9-12900KF AIO cooled with Corsair H150i Elite Capellix
Gigabyte Aourus Master X690
32GB DDR5-6400 Corsair Dominator RAM
EVGA RTX-3080 Ti Hybrid (AIO water Cooled)
I saw a significant increase in processing speed by upgrading to 6400 RAM from 4400. I picked up about 6 fps in almost every model.
I remember some user has talked about the ssd being a bottleneck in the past.
Here is the quote from old post.
Yes, those Tiffs are pretty demanding.
This is when i run them in Resolve for some color grading (3000x1700p 16bit)
2 Likes
I am hoping that search for more-for-less continues. We can’t keep chasing more speed only with new GPUs. Eventually the GPU outruns the CPU, bus and RAM and we get bottlenecking. And we can’t keep buying new high performance computer just to keep the that high performance GPU happy. Not without a mortgage on the home.
I have been looking for similar “down to brass tacks” discussions about the non-obvious tweaks and gottcas lurks in getting more performance from a computer. Bottlenecking is one that takes a deeper understanding of how TVAI and the hardware interact.
Certainly an eye opener about different types of SSDs:
| Type | Bits per Cell | Voltage States | Endurance (P/E Cycles) | Use Cases |
|---|---|---|---|---|
| SLC | 1 | 2 | 100,000 | High-performance applications |
| MLC | 2 | 4 | 10,000 | General consumer use |
| TLC | 3 | 8 | 3,000 | Budget-friendly consumer products |
| QLC | 4 | 16 | 1,000 | Read-heavy workloads, archival use |
| PLC | 5 | 32 | TBD | Future applications, high-capacity needs |
And there is a price to be paid for endurance:
| SSD Type | Price per TB (Approx.) | Performance | Endurance (P/E Cycles) |
|---|---|---|---|
| SLC | $45 - $100 | Highest | 50,000 - 100,000 |
| MLC | $20 - $50 | Moderate | 3,000 - 10,000 |
| TLC | $15 - $30 | Lower | 1,000 - 3,000 |
| QLC | $10 - $20 | Lowest | 100 - 1,000 |
Not sure the dollar figures are accurate, but the relative pricing tells the story.
Come on WD VelociRaptor WD1000DHTZ! 
