Any Carbon users who were using UAD Apollo before?

[LukeHoward]

Great comments – just to respond to a few:

Thunderbolt is closely related to PCIe - a relatively direct communication with the CPU, whereas Ethernet is a 'bolt-on' for network communications and requires translation to 'ip'.

AVB operates at Layer 2 so does not go through the IP stack. The default latency (presentation time offset) specified by the AVB spec is 2ms for Class A traffic but hardware implementations such as RME’s can tune this down as low as 0.3ms. This does require cooperating devices for low latency in both directions as the offset is set at the output (”talker” in AVB-speak). It would be difficult for a non-realtime software implementation such as the CoreAudio one to achieve these latencies though and you are correct that PCIe and even USB interfaces are able to offer lower latencies.

Thunderbolt is fast (40Gb/s) and Ethernet is slow (10Gb/s if you're lucky, more usually 1Gb/s).

Audio bandwidth is relatively tiny, 1Gb/s is enough for many hundreds of audio channels (particularly in the case of AVB where the switches cooperate to reserve bandwidth).

Thunderbolt is a point-to-point communication medium, intended for the CPU to talk directly to peripherals whereas Ethernet is a many-to-many communication protocol with lots of things built in to deal with collisions and delays in devices speaking to each other. These overheads create inefficiencies.

Not really an issue with modern Ethernet switches because each port is its own collision domain.

[dominicperry]

Originally Posted by LukeHoward

Great comments – just to respond to a few:



AVB operates at Layer 2 so does not go through the IP stack. The default latency (presentation time offset) specified by the AVB spec is 2ms for Class A traffic but hardware implementations such as RME’s can tune this down as low as 0.3ms. This does require cooperating devices for low latency in both directions as the offset is set at the output (”talker” in AVB-speak). It would be difficult for a non-realtime software implementation such as the CoreAudio one to achieve these latencies though and you are correct that PCIe and even USB interfaces are able to offer lower latencies.



Audio bandwidth is relatively tiny, 1Gb/s is enough for many hundreds of audio channels (particularly in the case of AVB where the switches cooperate to reserve bandwidth).



Not really an issue with modern Ethernet switches because each port is its own collision domain.

I won't argue with any of your points Luke, as my technical network knowledge hasn't been enhanced much since about 2005. If you want to know about DECnet or SNA gateways, then I can probably help.....

So, given that AVB bypasses some or all of the reasons that I identified as being the cause of Ethernet slowness, can you explain why Carbon still has more latency than a direct-attached TB interface? Is it most likely drivers, or Avid's AVB implementation being poorer than RME's, or perhaps something else. Genuinely curious to know.....

Dominic

[LukeHoward]

Originally Posted by dominicperry

I won't argue with any of your points Luke, as my technical network knowledge hasn't been enhanced much since about 2005. If you want to know about DECnet or SNA gateways, then I can probably help.....

Hah! I'm sure there's someone still using DECnet somewhere!

So, given that AVB bypasses some or all of the reasons that I identified as being the cause of Ethernet slowness, can you explain why Carbon still has more latency than a direct-attached TB interface? Is it most likely drivers, or Avid's AVB implementation being poorer than RME's, or perhaps something else. Genuinely curious to know.....

I'm going to punt that they are using the default presentation time offset of 2ms, because they can't change CoreAudio and, indeed, even if they could, it might be harder to get under that in a non-realtime software implementation (for pretty much the reasons you mention).

I should add that latency is pretty subjective and the Carbon may well be fine for VI use, I have not tried it. AVB in CoreAudio circa 2015 was not great with small buffer sizes but, things have probably improved since then.

[chrisdee]

What's got lower latency recording VI's and recording vocals between Carbon and MTRX Studio?

Would be nice to know the numbers for both HDX and Native PCIe.