Re: [tsvwg] a new method of congestion control

[[e-mail]]

To me, it appears that Nvidia is great on Marketing;

Alone for RoCE they re-invented at least three difference "better than anything before" CC algorithms (I believe the current one is a variation of a delay-based congestion controller, with avoiding some of the signal noise by depending on a very hardware-centric implementation, and auto-magic tuning optimizations when end-hosts communicate via compliant switches).

AI "workloads" should be differentiated between training and inference; the former is usually signified by very high bandwidth demands (e.g. uncompressed, or low-complexity compressed video) at significiant higher-than-realtime data ingestion, but in every case a very low latency response requirement.

The latter has much lower bandwidth rate requirements (e.g. sensors in real-time) but still low latency requirements.

I wouldn't hold my breath, and would only be interested in learning about the scalability limits and real-world issues of these improvements created behind closed doors...

Richard



Am 05.06.2023 um 02:53 schrieb Dave Taht:
> announced by nvidia here:
> https://www.zdnet.com/article/nvidia-unveils-new-kind-of-ethernet-for-ai-grace-hopper-superchip-in-full-production/
>
> I have no idea what an AI workload looks like, familiarity with a ton
> of DC l2 protocols, and there are hints in this post about telemetry.
>
> Anyone have clue here?
>