Benchmarks Show RTX 4090, 4070 Ti Take A Big Performance Hit With External GPU Boxes
![4090 external](https://images.hothardware.com/contentimages/newsitem/63756/content/4090-external.jpg)
A user on Weibo recently demonstrated this use case with a GeForce RTX 4070 Ti Super and a GeForce RTX 4090 connected externally via OCuLink. OCuLink is an external PCIe connection; while Thunderbolt 4 tops out at a paltry 40 Gbps, OCuLink's PCIe 4.0 x4 connection offers ~63 Gbps. This increase is notable, but there are significant drawbacks to be aware of. According to these performance results, users can still expect a significant drop in performance for GPUs connected in this manner.
![4090 score](https://images.hothardware.com/contentimages/newsitem/63756/content/4090-score.jpg)
The NVIDIA GeForce RTX 4090 is notorious for needing the fastest CPU possible, especially at lower resolutions. When gaming, for example, a less-capable CPU will bottleneck this GPU at any resolution below 4K, and even at 4K in some games. The test machine's Core Ultra 5 125H is certainly pretty fast, but a 28W mobile part doesn't hold a candle to something like a Core i9-14900K. With that said, given that we're comparing Graphics scores here, that's less of a factor.
![4070ti](https://images.hothardware.com/contentimages/newsitem/63756/content/4070ti.jpg)
When the GeForce RTX 4090 was tested using Time Spy Extreme, which runs at 4K resolution, the bottleneck appeared to be significantly less than the standard 1440p Time Spy test. Once again, the CPU is less important at higher resolutions, where the bottleneck typically moves over to the GPU. This will free up the GeForce RTX 4090 to flex more of its brute-force muscle, even when connected via OCuLink.
Even with this drop in performance, these OCuLink connected GPUs can still be an attractive option for users who frequently use a docking station but absolutely need to be mobile. When connected to a less graphically-capable device, such as a laptop, they can still offer a much higher level of graphics performance. The reduced PCIe bandwidth also has a smaller effect on compute and AI workloads that can reside directly on the GPU.