What is the best system config for ue5 performance?

I have an old system and by the time I buy a new system, probably UE 5.10 will be released.

So what kind of system will work best?

Does UE require more threads or more VRAM for a performance boost?

Current System-

quad core intel 750
24 GB ddr3 RAM
GPU= AMD saphire r9 390 - 8GB VRAM.

The issue is Metahuman LOD for the groom is not more than 1 and shaders take hours to compile.

I have to create Cinematic animations with UE5. Edit: The aim is to make Hollywood style 3D movies and advertisements, which will run on TV, cinemas, Film Festivals or Netflix. The idea is that, 4k frames can be rendered faster than traditional render farms, using real-time rendering in UE, without bothering about potential performance bottlenecks. Thus more VRAM for handling bigger scenes and textures.

So what should I invest in - Processor with as many cores as possible, probably AMD which has 32 cores (I believe more cores than Intel.)

How much RAM will be better? People say I am crazy when I tell them I have 24 GB RAM. Is this much enough for a new system or what is better? say 64 GB.

It will be a totally new system, so I will probably need another GPU, so how much VRAM should I go for? GTX 3090ti with 24 GB VRAM or less?

Obviously, they are not available right now, with all the GPU getting picked up by crypto miners and are being sold at insane rates.

Hopefully, by the time I am ready to buy them, even better GPUs will come out and the prices might go down.

I am running my current system for 11 years, so I would like to invest in a new system that can handle UE5 for the next 7-10 years.

Please suggest configs.

Thanks

I’m on a 8gb RAM / 4gb VRAM PC, sometimes on a 8gb RAM / 2gb VRAM system and you’re talking about 64gb systems with 24gb VRAM?
I laughed so loud. :joy:

1 Like

Lol Man Jokes aside. I just got a Texture Streaming Pool error. Had to put it at 6000MB.

That means it can use about 6GB of my 8GB VRAM. (I hope I am right)

My aim is not to make games but movies and that also fast using the real time rendering, I am thinking for a long term system.

Atleast I will not have to wait to render 1 frame for an hour, which normal renderers use for cinematic quality.

So suggest configs, even if they are low or high. Something where I will not have to bother about bottlenecks in performance rendering and still get hollywood kind of quality.

Well, if money is no object, For extremely high quality video making I’d suggest ((at-least)) something like a thread ripper, 32 GB of ram, and your v-card should focus on memory space rather than just total speed. Until you convert it into a video file, your scenes are going to be hogging tons of resources if you’re going for movie quality.

So I have 27 Gb of VRAM. I still have that message sometimes when using Quixels assets.

Though as I see there is 11 Gb of internal memory and it is quite busy. The rest is external and it is always almost empty.

https://answers.unrealengine.com/questions/343646/texture-streaming-pool-over.html

27 GB of VRAM? Which graphic card do you have?

Yes man I got it sorted out, but I have decreased the LOD bias of each object to 2048 and below, otherwise, it was going over 6 GB of VRAM. That is why I need the best system, though I am not sure, when I can buy it, but right now I am practicing UE5 with my current system and trying to find an optimized way out.

Exactly my concern, with megascans assets, I have just created a garden lawn with grass and plants and the texture pool is going over 6GB. I had to reduce the LOD bias for texture down to 2048, but for movies, that might not work.

To get high details, I will have to get a GPU with max memory space possible.

So this is what I have figured out the best rig for Unreal - I have to plan and earn enough so that I can buy it by 2022 - 2023.

Ryzen Threadripper Pro 5000WX - 2 of these can be inserted in one motherboard, (2-way symmetric multiprocessing (SMP) support.) so a total of 128 cores - meaning 256 threads. (I will probably be able to afford only one at a time, not sure I will even find it at my place or not, but not buying anything below 64 cores)

Motherboard - probably gigabyte or Asus server boards that can support Threadripper Pro 5000WX.

RAM - if I am spending so much I would go with at least 128 GB RAM. Server motherboards can support up to 2 TB.

GPU -GEFORCE RTX 3090 with 24 GB VRAM.

Any suggestions here? I am talking about the best config you can buy.

I’m wondering how much a system like that costs…
10k? 20k? No idea, just guessing.
It would be great to play games at 2.000.000 fps at 16k resolution.

Around $8000. Again not building this for games but for 4k+ movie rendering, meanwhile, I will be using my current rig to make animations and earn for it.

That’s too little in my experience. A good rule of thumb is to set aside system RAM for the VRAM of your graphics card (because the driver is likely to do this!) and then set aside 2 GB per TB of disk storage (for disk cache) and then make sure you have at least 32 GB left. If you also run virtual machines for whatever reason (work, research, isolation, server development) then you need to remove that RAM, too. When I upgraded to a >20 GB VRAM graphics card, I had to upgrade to 128 GB of system RAM. (I also use Hyper-V VMs.)

Similarly, when compiling shaders, the more threads, the better. When playing games, you actually want higher CPU MHz, but for development, more threads is often actually the better trade-off. If you have less than 16 real cores, you can absolutely upgrade. And, finally, make sure both your OS, and your apps, and your project, are all on NVMe drives (m.2 form factor, using PCI-express, not SATA or USB.)

If you want the best money can buy TODAY, then it’s actually a hard call between a 3975WX (for the clock speed across all cores) or something like a dual-socket EPYC 7800 series. Intel Core / Xeon can almost compete, but not quite, right now (likely to change soon.) And, in a few months, the 5975WX will show up to throw more confusion into the mix – but that’s not something that can be bought today.

Also, the top-end Quadros are better than the RTX-es for development, because the drivers really are better tuned for tools like Maya, and you get twice the VRAM, and NVIDIA is slightly more likely to answer support questions compared to RTX devices. It’s a small difference in practice, though.

Yes I was planning to buy 128 GB RAM, maybe even more, depending on what is the cost at that time, especially if buying a one TB RAM will not cost me a lot.

I will keep this formula in mind.

I have 4 cores right now- 8 threads. I will buy a new system and keep this system for graphic works and video editing purposes.

NVMe drives are not easily available at this time at my place, but hopefully, by the time I buy the system, they might be readily available. I have literally no experience with these.

It will take me some time to arrange the budget for the system, so by then hopefully the Threadripper Pro 5000WX will be out.

When I bought this system, I went for a Quadro 4000 card ( only one fan), available at that time. Had a horrible experience with it, had so much heat problem, then I bought this AMD R9 390, with 3 Fans. Cooling is not an issue anymore though the drivers are â– â– â– â– .

Quadro cards are way expensive than RTX and I am not sure how the cooling works nowadays. Obviously better than before but I would stay away from them due to my bad experience with Quadro cards.

Thanks a lot for your input, it really helps.

There are companies that sell custom waterblocks for a variety of Quadro cards. If you’re going with a custom water loop for your CPU anyway, it might make sense to put that in there, too. That would probably reduce noise, too, which is nice, and typically it also leaves more space in the case, which essentially means you can use a smaller case.

I am not sure I will find custom water loops for my CPU or cards at my place.

I had bought a gigabyte server motherboard for my current system because it had a custom cool loop. Never found anyone in the city who could provide the coolant at my geographical location. But If I am spending so much money I might look for branded options if available at that time.

A Quadro 8000 (NVIDIA Quadro RTX 8000) or 3090 ti, I will choose when I am ready to buy the system. Quadro does have double VRAM.

Thanks

Well, I was just going through Quadro 8000 review and I found out that it has no backplate and has a blower kind of fan system instead of Titan RTX, which uses a dual-fan cooling system. Even my current R9 390 has 3 fans cooling system.

For everyone who wishes to buy a Quadro with a blower kind of fan, especially in a place which has a lot of dust, I would recommend that not buy a blower kind.

What happens with blower is that in long term, a lot of dust gets trapped inside the GPU.
I had to get a special 7 star-shaped driver to open my Quadro 4000 and clean it every year because even dust increases the heat problem.

With 2 fan or 3 fan system, the cooling pipes are big and the whole surface is very open. For my R9 390 I took it out of case after 5 years and used a simple air blower to blow away all the dust, without really opening the GPU part. That is it, it got cleaned and my temp went down from 50 degrees to 35 degrees at idle. I would always go for a 3 or 2 fan system GPU if I cannot find a cooling system for it.

So I would go for a Titan instead of a Quadro. Now my question is if I add 2 x 24 GB Titan in my new system, will Unreal detect both the cards and combine them as 48GB or it can only use 24 GB / 1 card at a time?

I think this is one instance there the developers need to add their knowledge. Could someone who knows how to ping them, please assist me in this.
When a thread hit a bit of code that needs to have data given by another thread there will be a delay. To avoid this the processors have a prediction algorithm to keep the scheduling at optimum. But most effort on the AMD side is going towards optimize for the single processor scheduling. At least when it comes to Ryzen and Threadripper models. The EPYC processors are a different matter as the multiprocessor setup is more common. This is the hardware part of the answer.
But the Unreal Engine 5 (EA) might have some extra intelligence built in the code of the engine and thus sneak past the scheduling dilemma. For example (I am a noob on this) the Lumen part of the code could work more independent from the Nanite part and therefore UE5 developers have added some code to detect if there is a multiprocessor setup, and thus make some excellent use of this possibility.
On a side note, wonder if the Unreal Engine 5 will work well on the EPYC?

Ryzen Threadripper Pro 5000WX - 2 of these can be inserted in one motherboard, (2-way symmetric multiprocessing (SMP) support. So AMD is going for a multiprocessor with 5000.

So yes will Unreal be able to use this feature in the future? I mean it does not make sense that I get that expensive system and Unreal cannot use the Multiprocessor part.

1 Like

That’s not how it actually works. It sounds a little bit like you’re confounding hyper-threading (where one thread will execute while the other is stalled on missing the cache,) with (micro-)tasking code structure. They’re not really related.

In general, at the code design level, we build tasks that are big enough that a core stall is much smaller than the size of computation for each (micro-)task. This has been done in multi-core systems for a very long time (heck, I did it for BeOS in 1998) and is nothing new.

What is new-ish, is that game engines now must use tasking systems; a single thread is not enough to feed modern hungry graphics cards, much less the physical and particle and global illumination and audio simulations that go with them. Unreal has had a worker-thread / task-scheduling system for a long time, and one of the reasons to use Chaos instead of PhysX is that it’s allegedly better at tasking across more cores, so there’s some benefit to be had there.

For content creation, the longest tasks that benefit the most from multi-core, is lighting processing (“build” button) and shader compilation. Both of these have been multi-core all the way back to original UE4, so there’s nothing new needed there.

2 Likes