PCIe connections located farther from the central processing unit
PCIe connections located farther from the central processing unit
Hi Everyone, It’s tough to locate solid research or studies on this topic. I might have overlooked something quite basic already in these discussions. My motherboard has six PCIe ports – four x16, one 4-pin and one 1-pin. For the sake of discussion, let’s assume all four x16 slots operate at full speed. Any performance limits or “it doesn’t really matter” remarks aside, my question is: Does placing the GPU in a slot farther from the motherboard affect its speed or storage performance? I’m imagining moving it to the first slot for testing, running a benchmark, then relocating it to the fourth slot and repeating. In this scenario, both slots are physically x16 with 8x speed enabled. I think SSD tests could be swapped for GPU tests since any noticeable difference would likely be negligible. Has anyone ever checked this before? I can picture the results being almost insignificant, but it’s interesting to explore. Trying to search online or on these forums is full of “Wow, did you know your x16 slot isn’t really X16?” reactions! I considered testing it myself, but with my setup it’s not that simple – too many add-in cards to move around.
julian_PVP
06-01-2022, 05:22 PM #2In theory, yes—but given how fast light travels, calculating such tiny differences becomes extremely challenging.
dirty_donut
06-02-2022, 09:41 PM #4In theory, yes. Although longer traces shouldn't slow down data transfer directly, they would add delay, potentially affecting performance if your app is latency-sensitive. Most current GPUs maintain good speed even at PCIe 2.0 rates, so any difference would likely be negligible. For SSDs, the extra delay would probably not matter much. Latency becomes more critical when many small transfers occur, as the start time of each transfer can become significant compared to its completion. If transfers are lengthy, the time before they finish matters more than the speed itself. I believe most transfers to a GPU or from an SSD tend to be faster than numerous tiny requests. The main risk would be signal quality loss with longer traces, possibly increasing errors and thus impacting performance. This should generally be acceptable unless you're dealing with very high volumes of small data. The biggest concern is usually the overall speed versus delay trade-off, rather than absolute latency values. Signal integrity could be affected, but with quality components it should remain manageable. This explains why PCIe versions are often limited to shorter traces—balancing speed and reliability.
Major_Limitz
06-08-2022, 07:30 AM #5Thanks for the update. I really appreciate your attention to detail. I was wondering if someone actually put this through testing. It’s fine—I’ll just have to disassemble and try it out eventually. As originally mentioned, I’m mainly interested in understanding the precise difference between the numbers, whether it’s a tiny percentage or zero. Either way, it would be interesting.
No noticeable variation found between PCI-E lanes from the same device. The signal operates at 100 MHz, encoding bits onto the bus with each frequency tick, organized into packets where minor distance changes don’t matter. The electrical speed through copper falls within 80-90% of light speed—roughly 250,000 meters per second or 250 meters per nanosecond. If you pull dedicated PCI-E lanes from a chipset, performance shifts because additional signals must be multiplexed with other data (like SATA or USB), introducing upstream delays via another PCI-E connection.
The expected variation around 0.00014% is not something you need to worry about because it's beyond the accuracy limits of any standard test (margin of error). In today’s multi-threaded systems, countless external factors cause test outcomes to shift by just a few milliseconds, making a nanosecond difference impossible to detect. Specialized tools would be necessary, which you’d likely not afford.
It seems like these ideas rely on assumptions about manufacturers being fully transparent. In reality, there aren't always real-world tests for every scenario. While it's reasonable to expect high accuracy—about 99%—it's important to question where the motivation comes from. After all, isn't it interesting how performance can sometimes fall short of expectations despite claims?