TPX
(Thomas K.)
January 22, 2024, 5:17pm
45
This is shared address space, if you could fill up the memory with the GPU then the thing wouldn’t work at all.
Basically apple is screwing people over.
It used to be e.g. CPU 8 GB & GPU 4 GB.
Both could write to the memory independently.
Now they have to share it
So, 8GB for CPU and GPU, of course the transfer is slower than with separate but you now have 4 GB less.
It would be efficient if the data that is written could also be read by both, but since it is always called “shared” I don’t think this is the case.
The bandwidth also changes with the chip installed, while the neural engine is the same everywhere.
If these large models were made, around 95% of customers would probably disappear because they would no longer be able to use them.
jo.vo
January 22, 2024, 5:27pm
46
Since the RAM is accessible to both CPU and GPU it can in fact be freely distributed.
TVAI reports 48GB VRAM from my 64GB in total - and the bandwidth isn’t really shabby, either at 800GB/s.
If Apple was „screwing“ people with that, everyone with an integrated graphics and shared memory would do so.
In fact Apple Silicon is about the only solution where that shared memory approach does make real sense as this chips have (at least for the Max and Ultra chips) high enough memory bandwidth to not really bottleneck.
But TVAI doesn’t really make good use of that nor of a full second Neural engine or double the CPU/GPU cores (see the comparison Max vs. Ultra with the only little speed gain).
TPX
(Thomas K.)
January 22, 2024, 6:01pm
47
Have you ever tried this?
I would be interested because I usually suffer from a lack of memory on the CPU side, already have 64 GB.
Don’t compare 4000$ (M2 Studio Ultra, 800 GB/s) minimum systems with 1000$ systems.
2000$ M2 Max 32 GB = 400GBs
And anyone who uses integrated GPUs with this software is beyond help.
Just imagine 128 GB with an iGPU.
An Intel Iris Xe 96EU has 2 teraflops fp32, the M2 Ultra is 10x as fast in purely computational terms.
That would be like comparing a PCI H100 with 96 GB at 1.6 terabytes per second for 10,000$ with the M2 Ultra, that makes no sense.
The H100 PCI-E would have 248 teraflops fp16, that’s three RTX 4090s at 700 watts.
Can be used in the cloud, even in packs of 4 or 8.
Building a computer for a friend. Thought I’d try this out even though I only spent one thousand dollars on it, and it’s not meant to be able to run TVAI. Just games and not even the newest of them.
Topaz Video AI v4.1.0
System Information
OS: Windows v11.23
CPU: 13th Gen Intel(R) Core(TM) i5-13400 31.68 GB
GPU: NVIDIA GeForce RTX 2080 Ti 10.782 GB
Processing Settings
device: 0 vram: 1 instances: 1
Input Resolution: 1920x1080
Benchmark Results
Artemis 1X: 15.88 fps 2X: 11.54 fps 4X: 04.07 fps
Iris 1X: 15.90 fps 2X: 09.01 fps 4X: 02.97 fps
Proteus 1X: 15.64 fps 2X: 11.44 fps 4X: 04.04 fps
Gaia 1X: 04.99 fps 2X: 03.38 fps 4X: 02.33 fps
Nyx 1X: 06.30 fps 2X: 05.39 fps
Nyx Fast 1X: 13.26 fps
4X Slowmo Apollo: 07.20 fps APFast: 62.79 fps Chronos: 10.88 fps CHFast: 18.67 fps
1 Like
TPX
(Thomas K.)
January 23, 2024, 7:33am
49
The 2080ti is a strong card still.
Faster than my Quadro RTX 5000.
And this year, it becomes 6 yeas old.
Nice build.
I was surprised by how well it did. I got it refurbished for about $380. He’s getting a killer deal.
1 Like
TPX
(Thomas K.)
January 23, 2024, 3:03pm
51
He is able to play Cyberpunk with DLSS and Raytracing activated in 1080p, thats nice.
1 Like
The end came for my GTX 1070 when Nvidia disabled B-Frames in FFMPEG encoding for them.
It lasted me 7 years!
Now I have 4090
2 Likes
TPX
(Thomas K.)
January 23, 2024, 7:26pm
54
yea, a bit useless, i know.
We dont know which model was used, but its a nice overview.
3 Likes
TPX
(Thomas K.)
January 23, 2024, 7:30pm
55
You’re not making my decision very easy.
I see all the people here in the forum with their 4090 causing problems.
I’m also interested to see if nvidia isn’t plagued by another GPU bug.
I’m stuck somewhere between 7900 XTX, 4070tiS, 4090, W7800 and RTX ADA 4500.
Yes, I know, all different cards.
I really appreciate the reliability of the Pro cards now, but everything good costs money.
On the other hand, I would like to try gaming cards again, the Radeons also run with the Pro drivers.
TPX
(Thomas K.)
January 23, 2024, 7:58pm
57
Oh man, reminds me of the time with the GTX 670 OC from Gigabyte, the RMA took forever and I had to call in another service provider who confirmed that the GPU causes a black screen after a certain time.
I did get an second one and after the old card did get back from RMA i did sold the new one as used on ebay.
Things you can see on the screen are often related to the Vram.
TPX
(Thomas K.)
January 23, 2024, 8:23pm
59
Yes, even the FE.
Just been back on reddit and this reminds me so hard of the gaming GPU days, hours and hours of searching for a bug.
I think the 4090 is built on edge because of the power target.
They are trying to force it into a corset and it would need more power.
One manufacturer left the market, they already knew why.
My first 4090 died after 9 months. but until that I had zero issues with it. RMA process was no problem (Zotac).
I have had zero issues with the new one.
I have a 1200w Platinum PSU and am using a 3rd party power cable as the squid was too inflexible and I couldn’t get my side panel on with it.
Card: Zotac RTX 4090 Trinity OC
Power cable: https://www.amazon.ca/gp/product/B0BJZCCMY1/ref=ppx_yo_dt_b_asin_title_o02_s00?ie=UTF8&th=1
PSU: https://www.newegg.ca/super-flower-leadex-platinum-se-sf-1200f14mp-v2-1200w/p/1HU-024C-00037?Item=9SIAU16JBM3310
I guess what I am saying is I would buy it again, even though I had a card failure. Zotac warranty is 3 years.
Pretty good replacement deal! Free upgrade!
Maybe if my 4090 fails again, I will get a free 5090 upgrade
1 Like
TPX
(Thomas K.)
January 23, 2024, 9:37pm
66
So this cable makes 100x more sense than the rough thing from nvidia or the aibs.
Why don’t the ixxxxs just sell something like this, it would be so much easier.
I wouldn’t trust the PSU, but that’s a matter of taste.
Sounds to me like a load change that fails.
You might have to increase the voltage for the memory (minimally), but that only workes for special cards.
1 Like
Sapirus
January 24, 2024, 12:06am
68
Topaz Video AI v4.1.0
System Information
OS: Windows v11.23
CPU: AMD Ryzen 9 5950X 16-Core Processor 63.923 GB
GPU: AMD Radeon RX 7900 XTX 23.938 GB
Processing Settings
device: 0 vram: 1 instances: 1
Input Resolution: 1920x1080
Benchmark Results
Artemis 1X: 22.52 fps 2X: 10.14 fps 4X: 02.78 fps
Iris 1X: 25.71 fps 2X: 12.69 fps 4X: 03.58 fps
Proteus 1X: 25.18 fps 2X: 13.10 fps 4X: 03.84 fps
Gaia 1X: 10.78 fps 2X: 07.56 fps 4X: 03.09 fps
Nyx 1X: 10.80 fps 2X: 08.80 fps
Nyx Fast 1X: 18.97 fps
4X Slowmo Apollo: 26.36 fps APFast: 45.76 fps Chronos: 16.10 fps CHFast: 24.21 fps
2 Likes
The PSU brand actually has an excellent track record and reputation - EVGA uses their components.
1 Like
Windows 11 debloated
No Antivirus or cleaner programs
ASUS-PRIME Z790-A WIFI (Bios 1604)
i9-13900K + Kraken AIO
4090 drivers + 546.33 studio drivers
128 Gig DDR5-4800 UDIMM 1.1V CL40
Motherboard bios all auto no OC
Corsair HX1200 PS
4 onboard Nvme 4th gen drives
Boot drive 2 TB
Cache drive 2 TB
Media drive 4 TB
Mixdown drive 2 TB
Topaz Video AI v4.1.0
System Information
OS: Windows v11.22
CPU: 13th Gen Intel(R) Core(TM) i9-13900K 127.75 GB
GPU: NVIDIA GeForce RTX 4090 23.59 GB
Processing Settings
device: 0 vram: 1 instances: 1
Input Resolution: 720x480
Benchmark Results
Artemis 1X: 128.76 fps 2X: 79.57 fps 4X: 22.45 fps
Iris 1X: 166.56 fps 2X: 105.03 fps 4X: 30.35 fps
Proteus 1X: 227.95 fps 2X: 121.95 fps 4X: 30.27 fps
Gaia 1X: 83.75 fps 2X: 62.16 fps 4X: 26.76 fps
Nyx 1X: 54.16 fps 2X: 49.39 fps
Nyx Fast 1X: 95.76 fps
4X Slowmo Apollo: 230.46 fps APFast: 324.34 fps Chronos: 169.83 fps CHFast: 203.36 fps
Topaz Video AI v4.1.0
System Information
OS: Windows v11.22
CPU: 13th Gen Intel(R) Core(TM) i9-13900K 127.75 GB
GPU: NVIDIA GeForce RTX 4090 23.59 GB
Processing Settings
device: 0 vram: 1 instances: 1
Input Resolution: 1920x1080
Benchmark Results
Artemis 1X: 41.75 fps 2X: 19.68 fps 4X: 04.63 fps
Iris 1X: 38.52 fps 2X: 19.61 fps 4X: 04.83 fps
Proteus 1X: 42.96 fps 2X: 18.88 fps 4X: 05.05 fps
Gaia 1X: 15.79 fps 2X: 10.96 fps 4X: 04.37 fps
Nyx 1X: 17.67 fps 2X: 14.71 fps
Nyx Fast 1X: 30.83 fps
4X Slowmo Apollo: 44.47 fps APFast: 72.53 fps Chronos: 33.51 fps CHFast: 34.85 fps
Topaz Video AI v4.1.0
System Information
OS: Windows v11.22
CPU: 13th Gen Intel(R) Core(TM) i9-13900K 127.75 GB
GPU: NVIDIA GeForce RTX 4090 23.59 GB
Processing Settings
device: 0 vram: 1 instances: 1
Input Resolution: 3840x2160
Benchmark Results
Artemis 1X: 08.66 fps 2X: 04.24 fps 4X: 01.15 fps
Iris 1X: 08.52 fps 2X: 04.54 fps 4X: 01.16 fps
Proteus 1X: 08.76 fps 2X: 04.58 fps 4X: 01.18 fps
Gaia 1X: 03.36 fps 2X: 02.30 fps 4X: 01.03 fps
Nyx 1X: 02.94 fps 2X: 03.60 fps
Nyx Fast 1X: 05.20 fps
4X Slowmo Apollo: 16.27 fps APFast: 22.21 fps Chronos: 07.21 fps CHFast: 12.65 fps
1 Like