> This is a bad take. Here's my take: (...) 5. They get to continue to purchase the cheaper IP switches and retain their existing hardware without retooling eve

> This is a bad take. Here's my take: (...) 5. They get to continue to purchase the cheaper IP switches and retain their existing hardware without retooling everything. OP's TCP offload engines do not require any retooling at all, if you consider stuff like buying IP switches not retooling. All you need to do is basically buy a network card. Also, if you read the article you'll eventually stumble upon the bar chart where they claim that the one-way write latency of TTPoE is only 0.7E-6 seconds faster than existing tech like infiniband, and it's also the only somewhat hard data they show. Does that justify the investment of developing their whole software+hardware stack? I'm sure the project was fun and should look great on a CV, but overall it doesn't look like it passes the smell test.