Correct. The question you have to answer for yourself is “will I get by with x8/x8 PCIe Gen4 or Gen5 slots?” - or do I need to buy into a much more expensive Threadripper/Xeon platform that offers more PCIe expansion capabilities.
You should consider that you can buy multiple desktop-quality machines for the price of even 1 workstation-class.
For a maximum of flexibility I am looking for motherboards with x8/x8 PCIe Gen5 slots + additional x4 PCIe Gen4/5 slots that have bifurcation enabled.
Hot contenders for me are
Only x8/x8 on PCIe Gen 4, but at least it doesn’t break the bank
Yes, I have checked on both these MB’s. Right now I will keep the Threadripper PRo/Asus WRX motherboard combo I have right now. And I will be patient, wait and see if they come out with a smaller TR Pro motherboard in a smaller format. Might be wishful thinking.
As a side note I do find it curious noone has made an mATX x8/x8 motherboard though, the lanes and tech had been there for even freaking budget $50 A520 ATX motherboards. I get not everyone appreciates such a setup and many prefer x16 / x4, but seriously…?
Yep, but there is no technical reason why it has to be that way, I am pretty confident most n00bs don’t really care if it’s x16 + x4 or x16 + x0 → x8 + x8, and the tech already exists for ATX so why not mATX?
I guess it is just because ITX is the new small size king. Incidentally, it is possible to do an x8+x8 riser using mITX, and as long as you are watercooling one of the cards the Sliger SM580 case would allow you to do just that. Oh well.
IF you take a solid X16 to a X8 + X8. How exactly does this cut the performance of the CPU. I would think it cuts the performance of the GPU in half. So technically, if I have 2 GPU only going X8 each. Then a single X16 out perform the X8 + X8?
This actually depends on the card in question. Fun fact: The RX 6600 XT is an 8 lane GPU. So running dual x8 with those cards, you essentially lose 0 performance.
It also depends on the PCIe generation. A gen 3.0 x16 is the same bandwidth as a gen 4.0 x8 is the same bandwidth as a gen 5.0 x4.
While yes you will always be able to find use cases where two x16 PCIe 5.0 is required for maximum performance, most GPUs are happy with x8 PCIe 5.0 today. Going any faster is pointless - unless you are pushing state of the art AI networks or something, then you might not even need full speed PCIe 5.0.
Incidentally no GPUs support 5.0 as of now. It’s all 4.0.
Right now, I use a single A6000, but the updated version of the 6000 is coming out soon, so the A6000 on the used market will get highly inexpensive. So I would like to get a second A6000, but even at getting one at a sweet price. I still want to ensure I get the full performance out when needed.
Even encode/decode of 4k 60fps raw video is less than 2GB/second, one of the most transfer-intensive workloads, which PCIe 3.0 x8 has plenty of headroom for.
Much of the benefit of newer PCIe generation support is to allow at least as good performance with fewer lanes.
There’s tools to measure the I/O bandwidth, so you can find out in advance if it’d make any difference.
They exist, e.g. the ASRock Rack X570D4U series does PCIe4 x8/x8, and the X470D4U before it can do PCIe3 x8/x8.
For common desktop boards, there doesn’t appear to be much use case for this layout: two mainstream high-performance GPUs wouldn’t fit in a µATX chassis, and most other common PCIe cards are below x8.
