I believe eGPU connectivity and support really fits in well with Framework’s sustainability and modularity. Looking at the ASUS ROG Flow x13 that just came out, it allows for eGPU connectivity with less bottleneck and latency (proprietary connector). I would really like to see a framework motherboard be able to accommodate an adapter of this kind with a connecting eGPU case. It would be the ultimate in modularity and flexibility of a laptop.
1 Like
What amount of bottleneck would you get currently with USB4/TB4 currently with even the highest end graphics cards, I wonder? And if it’d be worth it if USB4/TB4 was already good enough.
Definitely would like to see eGPU’s be able to get full bandwidth, eventually.
Unfortunately Asus took a standard royalty free connection type (Oculink 8i) and moved some stuff around to make it a Asus only product with no hope and the standard going anywhere. Even if it was OK by Asus, the bandwidth is double what the thunderbolt controller on the framework can process, so it just wouldn’t work.
The real problem here is everyone was all hopeful that Thunderbolt 3->4 will be a big bandwidth increase but that just didn’t happen so eGPU’s are just becoming a worse and worse tradeoff.
Fortunately a few hours ago, an Intel exec tweeted out a photo (and then quickly deleted it), showing that TB5 will be the 80Gbps everyone hoped TB4 would be. I don’t think eGPU’s will be a good bet if/until then.
1 Like
The PCIE bottleneck with USB4/TB4 is significant. It is limited to ~4 GB/s with PCIE 3x4, while high end GPUs have 8 times the bandwidth with PCIE 4x16 at ~32 GB/s. It also limits in half the bandwidth of PCIE 4x4 SSDs, some of which can reach ~7 GB/s.
As for Thunderbolt 5, it would presumably double the PCIE bandwidth to ~8 GB/s, but by the time it releases GPUs will be on PCIE Gen 5 x16 with ~64 GB/s bandwidth, continuing to be a significant bottleneck.
1 Like
I’ll take what I can get with Thunderbolt eGPUs, it’s just a shame this bandwidth cat-and-mouse between GPUs and Thunderbolt doesn’t seem to show signs of slowing down. If the chip/GPU shortage is anything to go by though, there’s probably going to be healthy markets in previous-gen GPUs for a good while, long enough where a combination of modern TB and an older GPU might escape current eGPU bottlenecks…
1 Like
Even if modern GPUs have theoretical 32GB/s bandwidth with gen4x16, it doesnt mean they use all of it. IIRC most testing has found limited performance increase on high end GPUs even just between gen3x8 and 3x16
1 Like
For gaming even a RTX 3080 on PCIe 1.1 x16 (4GB/s) averages 75% of the performance of PCIe 4.0 x16: NVIDIA GeForce RTX 3080 PCI-Express Scaling - Performance Summary | TechPowerUp
1 Like