Re: [Coin] The Future of P4, Revisited

[[e-mail]]

Jeff, 

 

You did not read the Octeon 10 slides I sent out. This asic supports the SERDES bus, so the DPU can be used in a switch. Second, this asic support 1 Tbps which a good rate to test many applications.

 

Hemant

 

From: Coin <[e-mail]> On Behalf Of Jeff Tantsura
Sent: Monday, May 15, 2023 5:45 PM
To: hemant=[e-mail]
Cc: Hesham ElBakoury <[e-mail]>; Toerless Eckert <[e-mail]>; Marie-Jose Montpetit <[e-mail]>; coin <[e-mail]>; [e-mail]
Subject: Re: [Coin] The Future of P4, Revisited

 

Hermann,

 

You are comparing apples to oranges (or actually apples to blackberries :))

Trident4 (with 256x 100G SerDes) is a 25T device.

Jericho (2C+) is 14.4T device.

Octeon is a smartNIC/DPU NPU - (400G now, 800G in 2 years?) that should not be compared to the above.

 

There’s a number of DPUs on the market that expose P4 APIs (partial spec + private extensions), non of them can replace (economically) a switch/router ASIC, I believe there a Pensando based Aruba switch (public data) - look at the price and power per bit.

 

P4 - there’s quite some stuff in production that uses P4, Google is also pushing PINS in SONiC/SAI, while in the beginning it was a genius strategy to (up)sell BF, now the scope is somewhat beyond that.

P4 is indeed quite limited as compared to “native” SDKs, with an argument that is provides normalized/common platform/layer to program HW, one might however argue that SAI will eventually be the only winner in this place.

Lack of wide ecosystem and development of compliers towards different HW could eventually be the end of P4 (and potentially push to develop something new).

 

My 0.2

Jeff





May 15, 2023, at 1:19 PM, hemant=[e-mail] <mailto:hemant=[e-mail]>  wrote:

 

Broadcom NPL is proprietary, and Broadcom develops customer code in NPL for switch. Further Trident/Jericho 2 are 10-12 Tbps asics while Octeon 10 is a 1 Tbps asic. My company’s P4 compiler makes Octeon P4 programmable.

 

Hemant