Gigabit Ethernet (RJ-45) vs SFP+ / SFP28 (Fiber) gaming latency?

[pixelwave]

Hi,

I am currently looking into switching from Gigabit Ethernet (RJ-45) to SFP+/SFP28 (Fiber).

I have one general question and I got different answers.

I plan to hook up Mellanox ConnectX-4 Lx EN (Dual SFP28) cards with optical transceivers (10/25GbE) to a 10GbE SFP+ Switch.

Compared to hooking up the onboard Gigabit Ethernet of a PC to that switch (RJ-45 1GbE port) or using a the above stated connection with the Mellanox card - which connection has the better latency aka is better for things like online gaming:

• 1 GbE Ethernet (RJ-45) ?
• 10 GbE SFP+ (Fiber) ?

The specs of the card state sub-microsecond latency, Google says "SFP+ uses simplified electronics without encoding, and its typical latency is around 300 nanoseconds per link“.

For Gigabit Ethernet I found values of "ranges from 50 to 125 microseconds“. On FS.com website they state 0.1 microseconds (SFP+ Fiber) compared to 2.6 (they only compare to 10GbE in that overview but I guess 1GbE would be similar?).

So I would assume the Mellanox ConnectX-4 Lx SFP+ (Fiber) combination would be better than the Gigabit Ethernet (RJ-45) when it comes to latency / online gaming?

But then I also found this statement pointing out Gigabit Ethernet might be better:

„In general, SFP+ has a higher latency than Gigabit Ethernet due to its higher data rate and the more complex signaling involved. The latency of SFP+ can vary depending on several factors, including the distance between the communicating devices, the quality of the optical fiber, and the network equipment used.

In some cases, the higher data rate of SFP+ may offset the higher latency, resulting in overall faster transmission times compared to Gigabit Ethernet. However, for applications that require low latency and real-time interaction, such as online gaming or video conferencing, Gigabit Ethernet may still be a better option due to its lower latency."

I guess in general it might be negligible since we are talking sub-microsecond / nanoseconds and there are other factors introducing way higher latencies but just out of curiosity I wonder which method would be faster ...

 

[i386]

I...

in your local network the latencies will be probably in microseconds. Traffic over the internet will be milliseconds. -> it doesn't matter for online gaming if you use fiber or copper at home.

In your lan the sfp+/10gbe copper will be faster (more bandwidth and lower latency) than 1gbe copper

 

[LodeRunner]

Whatever latency may be added with optics or DACs will be effectively immeasurable at the OS level. Apart from faulty hardware, there's no gamer on this planet that would be able to tell you if they were on copper or fiber, gig or 10 gig, without looking at the NIC and relevant information in the OS.

Most modern games have some form of latency correction anyway that will be far more infuriating than the micro/nanoseconds added by a transceiver.

 

[Siman]

You're not going to notice it, but for connecting anything within about 10 meters of your devices DAC/copper offers less latency than fiber. The conversion to light and back to electrical signals take time. The rest honestly depends on payload size, buffers, ect... but again we are talking in so minute differences... You would be better off getting a new monitor, GPU, ect...

 

[pixelwave]

You're not going to notice it, but for connecting anything within about 10 meters of your devices DAC/copper offers less latency than fiber. The conversion to light and back to electrical signals take time.

Thanks for the feedback! Although that contradicts the fs.com statements:

 

[Siman]

You're forgetting the distance factor. After a certain length the benefits of not converting lessen. That is why generally its best for in rack routing to use DACs. If you're going to the distribution or core switch it would be best to run fiber. 10meters is where the convergence happens on average. But yes the encoding takes time on ethernet, should have specified only DACs.

 

[pixelwave]

But they specified DACs as well or do I read the table wrong? My understanding is DACs are mostly more cost efficient for shorter cable runs and this table compares SFP+ Fiber vs SFP+ DAC ... with Fiber having a third the latency of DAC for a complete link run? Not specifying length I assume this is in general valid for the whole length spectrum which is limited for DAC anyways ...

 

[mach3.2]

Arista found DACs to have slightly lower latency though...
I'm more inclined to trust Arista over FS since Arista's paper discussed their methodology vs FS's unexplained numbers.

https://www.arista.com/assets/data/pdf/Copper-Faster-Than-Fiber-Brief.pdf

But as i386 mentioned, you aren't going to perceive any meaningful difference between 10GBASE-T, SFP+ fibre and SFP+ DAC for gaming.

 

[Siman]

I think you're getting the table confused with what information I'm trying to relay. Ethernet 10baseT SFPs have to encode to transmit data, that adds latency (as you can see with the table). DAC and Fiber modules don't need to. What you're looking at with the table is the computational overhead of the different devices, not the transfer medium itself. The table would be best represented as the time delay from data to leave the port, not the actual distant end receive latency. This would also depend highly on many factors: SFP quality, switch vender, SFP vender, ect... IMHO it's not the best way to say something "has less latency".

 

[pixelwave]

But I am just comparing Fiber and DAC .. do not use 10baseT.

 

[Siman]

Dac still has less overall latency vs fiber, over short distances. It's also very well documented, IDK where, what , and how FS.com is getting its numbers, like mach3.2 said doesn't sound right even for what they posted. Overall device to device dac has lower latency, my personal network experience and what arista posted is the same.

 

[nexox]

That FS table has confused many people, not only is it just kind of poorly done (no units, doesn't really explain what it measures, and the rows just... do multiplication for the reader?) but it also conflicts with basically everything else published or generally known about DAC vs Fiber.

In the end, you're looking at 2.5 uS additional latency using cat6, while the best round trip time you can hope for between two hosts on a 10G network is more like 100uS, so make your decision on price, power consumption, distance, or whichever comes in a color you prefer, not latency.

 

[ano]

fs.com is good guys in general, but they also do weird stuff

DAC is faster, DAC is also horrible, and I hate it, and I also use it.

for short 0.5 its good, for power usage its good. example, we are building storage/ceph racks, your getting 20 nodes x2 per rack, you save 12-15w PER server when using DAC... it adds up. its also better for latency, marginally in realworld, more in theory