I've been searching this forum and the various vMix FB groups, for peoples experiences running a Threadripper 3960x with vMix.

I'm really struggling to find new i9 10980XE chips or X299 motherboards, but I can still find new 3960x's and TRX40's at discounted prices.
As I also do a lot of video editing I'm also considering the new Threadripper 7960 & TRX50.

Thanks in advance.

Generally speaking, I've had very good luck with the Threadrippers of the world. When it comes to the heavy duty moving of data around the system, PCI-E lanes matter. Most consumer CPUs nowadaysonly offer 20-24 PCI-E lanes to the CPU; the Threadripper 39xx series offers 64. Intel's i9-xxxXE series (basically, Xeon chips that didn't pass stringent checks for thermal/whatever there) offers 40-48 lanes depending on the generation.

If you are serious about mixing with many inputs/large datasources, you need more lanes to the CPU to push data through. If you look at the block diagram of a non-Threadripper Ryzen motherboard, for example, you will see that most of the PCIE slots are wired to the chipset, and then the chipset itself only has so many lanes to the CPU; on consumer CPUs it's typically 4, but can be more on prosumer boards. So if you are moving data from, say, a BMD Decklink card to the GPU for encoding, you end up bottlenecking on the northbridge when pushing data around. Same thing for the motherboard's Ethernet ports for example, especially a 10G port - that gets wired to the chipset and is sharing bandwidth from there to the CPU.

Here's a block diagram from the Asus ROG Zenith Extreme Alpha sTRX4 motherboard (used with a 3960/3970/etc)


So based on this, you can see where you might want to plug devices in (or not), depending on the purpose. You have 4 PCI-E slots that are directly wired to the CPU, two at 16x (16 lanes), one at 8x, and one that might be 8x or might be 4x, depending on if you are using NVMe slot M.2_2. Based on this, and depending on how the slots are actually laid out on the physical motherboard, you might actually want to end up connecting your GPU to PCI-E #3, because most "consumer" GPUs are now taking more than two slots, so occluding the maybe-4x/maybe8x slot hurts you less than occluding both the 8x slot and the 16x slot if you were to place it in the top slot. This is something you wouldn't be able to do on a consumer board, because only the top slot is typically wired straight to the CPU on those boards, and the GPU is absolutely something you need to push a lot of data to/from.

In fact, based on this board layout, I'd probably connect my boot NVMe drive to either M.2_3 (especially if I'm not using the SATA6G ports), or on the DIMM.2 NVMe riser, because the OS drive isn't necessarily where you are going to need the full speed of directly wired lanes. However, if you plan on doing MultiCorder recordings, especially, say, NDI straight to disk, no reencode, then you'd want the drives you are using for that plugged into M.2_1 and M.2_2 (along with, of course, the appropriate NVMe media that can handle the write speeds you are looking for; not all NVMe is created equal, and my personal favorite would be the Samsung 970 Pro, as it was the last MLC drive from Samsung; as of the 980 series both Evo and Pro line are using TLC).

You can sometimes find block diagrams for chipsets, but how those chipsets end up being implemented on a given motherboard can change. Sometimes you can find them in the manuals, sometimes you need to go looking. You can generally find something though. For the Threadripper Pro series, I haven't done a build there yet, but I've been eyeballing the ASUS Pro WS WRX90E-SAGE for a future build. The Threadripper Pro's are even more badass because they are running with 128 PCIE lanes, giving you *all the IO* to do *whatever you want to*. You can find the manual for the WRX90E-SAGE here. Page A-1 of the manual contains a block diagram for the motherboard.

This motherboard features 7 x16 physical slots, 4 that are wired at 16x to the CPU, and 3 more (all 16x physical, but one is only 8x electrically) possibly combined with some of the USB3 ports (not sure on the "with REDRIVER" annotation). Still need to look more into that. Also has a total of 4!! NVMe slots that are directly wired to the CPU, without going through the chipset bottleneck.

But generally, when searching for a motherboard, start with the processor you want to use, find a board that has the physical geometry you are looking for (how many and what types of PCI-E lanes, etc), and then look for a block diagram for that motherboard and see how everything is wired up.

Generally speaking, a GPU *needs* to be wired to the CPU. Decklink cards *should* be wired to the CPU if you don't want to run into bandwidth issues (ie, if you want more than, say, 2 SDI links in total). If you are doing large amounts of NDI, you *may* need to wire a 10G NIC directly to the CPU, depending on if you are doing full-frame/high-bandwidth NDI or NDI|HX. If you are trying to do multicording, you probably want to do that to NVMe drives that have high write speeds and are directly wired to the CPU, or to SSDs that are connected to a HBA (host-bus adapter; sometimes called a drive controller) which is wired to the CPU. If you look at... basically every motherboard block diagram, the onboard SATA ports are wired to the chipset, so your bottleneck will be at the link from the CPU to the chipset.

The biggest thing to keep in mind is that people think about obvious bottlenecks: CPU utilization. RAM utilization. NIC utilization. IOPS on disks. But there are other, hidden bottlenecks present in the system, based on where your IO is happening, and how it's plumbed into the CPU. Don't stick dozens of things on slots/ports connected to the chipset and then wonder where your performance went.

If I had one bone to pick with Intel, it was how they ended up throttling their consumer CPUs around the 6th or 7th generation Intel, whenever they introduced the i9 line. Prior to then, you could get -XE processors (although they weren't yet branded as such) in the i7 line. First they moved them to the i9's, and then they brought regular consumer CPUs into the i9 lineup that only had (I think) the 20 PCI-E lanes (plus the DMI to the chipset), and then the (formerly i7 then formerly i9) CPUs became the i9-nnnXE's

Honestly, unless the use case was very small, I wouldn't build a switcher on anything less than an i9-NNNXE, a Xeon processor (or a DP Xeon, depending on how much IO), or a Threadripper (or an EPYC, I suppose, AMD's server CPU). And my personal preference would be a Threadripper/Pro. However, I must acknowledge a tendency to overbuild, not just for today but tomorrow. But I can understand how compelling a Xeon-DP could be, based on availability of used systems/etc. A single Xeon or i9-NNNXE again is more... limited (40/48 lanes depending on the generation), but if you aren't doing a ton, could see them working fairly nicely.

From the Gigabyte TRX50 Aero manual, on page 7.



The statement is a little contradictory, wherein they state that there are two X16 slots, but that a GPU should be placed in PCIEX16_1. I think the caution here, however, is to avoid placing it in PCIEX16_3, which is only a gen4 bus... not that it makes much of a difference because the graphics card you linked only uses PCIE gen 4, so the "benefits" of a gen5 slot aren't used. But it could also be cautionary that PCIEX16_2 is shared with something else.

The AMD TRX50 platform allows for up to 48 lanes of gen5 PCIE and a total of 92 PCIE lanes (88 of which are usable, because 4 lanes are reserved for the chipset). It's unclear if the chipset is using gen5 lanes or gen4 lanes based on what I've looked at so far (a quick skim). According to the block diagram you shared, 2x 16 and 3x nvme (which are 4 lanes reserved) is a total of 44 lanes (16 + 16 + 4 + 4 + 4), so it's possible that it's all straight wired.

But here's the thing: PCIEX16_2 doesn't have to be 16x electrically for you. Both of your BMD cards are only 8 lane cards (and PCIE gen 3 as well, so it doesn't matter which gen of slot you plug it into). I think it's very unlikely that if PCIEX16_2 is shared that it's shared down to only 4 lanes, as opposed to 8.

I can't speak to fit of the GPU and the motherboard without taking a closer look at both, but it does appear that PCIEX16_2 is spaced 3 slots down from PCIEX16_1 (on an old board which had slots in every... slot... PCIEX16_2 looks like it would be in slot 4) as M2B_CPU and M2C_CPU are occupying slots 3 and 2 respectively, which means that anything less than a 3 slot high GPU *should* fit. Of course, this is based on me looking at the diagram on page 4 of the manual, without ever having seen the board before, your mileage may vary, past performance is no indication of future success, etc, etc.

Also, with only 2 NDI|HX cameras you'll hardly be touching the bandwidth on the 10G NIC, so it shouldn't matter where that's plumbed in (that's well under 1G link speeds). Just keep in mind that NDI|HX will be leaning on the CPU for decodes, as nothing in life is free (you are already using these cameras, probably, so you already know this, but... talking about potential bottlenecks and stuff). 2 cameras shouldn't be that high of a load, but it is something to keep in mind.