From the article: “full 10 Gbps throughput full-duplex while supporting 10,000 sessions”
Doesn’t sound too impressive if you consider that Netflix has been doing 10-40Gbps with 10,000-40,000 sessions on commodity hardware [1] and is already doing 100Gbps as well [2].
This might be useful if you're implementing an app in the FPGA as well; having the NIC, TCP/IP stack, and app in the same FPGA eliminates PCIe latency. But besides HFT it's hard to find uses for this type of thing.
What about cloud providers like google? They need high performance throughput in their nics in combination with security. It sounds like that benefits from custom chip design.
9 comments
[ 2.9 ms ] story [ 29.3 ms ] threadDoesn’t sound too impressive if you consider that Netflix has been doing 10-40Gbps with 10,000-40,000 sessions on commodity hardware [1] and is already doing 100Gbps as well [2].
[1] https://openconnect.netflix.com/publications/asiabsd_tls_imp... [2] https://medium.com/netflix-techblog/serving-100-gbps-from-an...
> This way, a network-attached FPGA on ingress and egress to a CPU
> can accelerate functions such as encryption, compression,
> memcached and many others in addition to running the
> complete network stack.
The architecture’s resource requirements scale linearly with the number of supported sessions to over 115,000 given today’s 20 nm devices.