New and Old - Both blue, one bold...

[J7SC_Orion]

Just when there are new CPUs on the horizon (looking at you 7950X3d , 13900KS ). I finally got around to finish some long-planned for updates on both my current work & play combo, and an older work & play setup started in late 2018. The latter one is now tagged for some machine learning and rendering tasks in addition to some secondary-station gaming so I might actually have a few months when I am not doing any new builds or modding anything...

 

To recap: My current main system in the primary home office is below (3950x w/6900XT;  5950X and RTX 4090) and had some minor visual updates recently...

 

 

The 4090 in particular rocks...Port Royal at > 29,700, Superposition 8K > 17,200 etc. Add in DLSS3, Frame Insertion, NVidia Reflex and all those goodies -  it is just a sheer joy to play with, especially on a 48 inch 4K120 OLED. However, it is not all fun-and-games. In OctaneBenchmark this sample is among the top results for a single GPU (4090 or otherwise), but...

 

 

 

...with growing emphasis on machine learning for edge AI and visuals such as rendering, even a single and very powerful card has to take a bit of a back seat to multi-setups. OctaneBench has some listings with various over-the-top multi-GPU setups (20x plus). 

 

Since the RTX 4090 'displaced' a 3090 Strix that was now available, it was time to integrate the 3090 into the older Threadripper build with its 2x 2080 Ti in NVLink, not that they ever were slow.... The 2x 2080 Ti score > 21,200 in Port Royal (3090: ~ 16,100) and > 12,600 in Unigine's Superpostion 8K (3030: ~ 8800). However, when the 3090 is added to the 2x 2080 Ti, it becomes a genuine workhorse for rendering and ML: 19,200 CUDAs, 872 TMUs, 288 ROPs, 1416 tensor c, 218 ray tracing, 46 GB VRAM GDDR6/+X. This is where even a very fast 4090 meets its match with its 16,384 CUDAs, 512 TMUs, 176 ROPs, 512 Tensor c, 128 ray tracing, 24 GB VRAM GDDR6X - though to be fair, some of these are not directly comparable between generations. But using the same benchmark, the older 3x RTX GPUs still manage...

 

 

 

Of course, the other side of the coin is just how fast the single 4090 is, especially when taking power efficiency into account.

 

In any case, what absolutely amazed me (as I was expecting otherwise) is the lack of any issues in Windows 10 --- both the NVL/SLI 2080 Tis and the 3090 Strix have their own HDMI monitor inputs for a single 55 inch IPS HDR I have connected, and it is just a matter of setting the Windows display options re. which set should be the primary...so I can game with either the 2x 2080 Ti in NVL/SLI or just with the 3090, all smoothly. And for work tasks, all three easily combine in either setting. 

 

The toughest part really was the 'Frankenstein' operation without making it look hideous (IMO, anyway) and also to make sure that the cooling system could handle 1,700 W + without a problem. 3x D5 pumps and a total of 1080 x 62 mm rad space is enough even when everything is max-co'ed, though down the line, I might add a bit more rad space.  It takes 2x 1300 W Platinum PSUs (Antec HPC) which are connected via their proprietary 'OC Link'. I had those PSUs since 2013/14 - good quality, obviously.

 

Also, it gets pretty crowded on the mobo as you can tell as both the TR CPU and the 3x GPUs are water-cooled, never mind 9x PCIe 8 pin power cables (no wonder I've grown to like that 12VHPWR cable setup the 4090 uses in the first pic). I had to drop the 2x 2080 Ti in NVLink two slots lower than before, meaning one of them floats in its own cage and is connected via a PCIe riser cable (quality counts there w/ Lian Li PCIe 4.0, running 3.0). The 3090 Strix has its own PCIe riser via the top PCIe slot to allow for a roughly 45 degree angle for mounting - it too has its own cage it sits in.

 

So there you have it - some rendering and edge AI tools that cost nothing extra (and good for reasonably-sized proprietary data sets we have) which can run on one machine unsupervised once tasked ,while the newer setup is used for other tasks in another room...

 

 

 

 

Edited January 8, 2023 by J7SC_Orion

[J7SC_Orion]

I am still amazed how seamlessly Win 10 is handling the 3x GPUs (2x + 1x) in the 'Frankenstein' build shown above. I am no friend of MS Windows but I had expected bigger problems - or at least the need to constantly reconfigure graphics settings.

 

For games, it automatically picks the 3090, for rendering and ML, the 3090 and the 2x 2080 Ti combined.

 

I had a nice gaming session on this thing tonight (Forza H 5). The 3090 is perfect for the older 55 inch LG IPS HDR 4K60. Even the four-year old TR 2950X doesn't hinder performance at 4K. It has RAM read speeds of about 103 GB/s (four channel DDR4) which helps with games and of course very much with productivity tasks. 

 

This setup is about 100 feet from the other station which includes the 4090 and the 4K/120 OLED. The 4090 not only has close to three times the number of transistors and is about 1 GHz faster than the 3090 (which itself is no slouch at 2260 MHz), but it is nice to enjoy both while also having separate stations - one can run some bigger data sets while the other does its own thing.

 

Of course they can be networked together now but once I upgrade another system with a 7950X3D in a few months, I'll convert the free-up mobo into a file server - got the place rewired already with Cat 7 +8...