Hacker News Reader

Comment by archi42 | Hacker News Reader

The hobbyist in me is disappointed: Even the ExaNIC X10 seems to come at 350US$ used, and forget about the four figure UltraScale boards :( Of course, the FPGA can do a lot of stuff in addition, so for HPC this might be really nice to offload application-specific stuff really early (because sustained 2*10G traffic is bound to go somewhere, e.g. CPU[s]).

View on HN

$350 is cheap for such a powerful NIC. You could pay $10k for a Napatech, or $2k for a Solarflare, and not get this level of programmability.

libexanic is a remarkably clean user-space kernel-bypass library that allows you to do processing on early fragments of the packet while the rest are still being received.

Comparing used prices to new prices is dishonest. It costs $2,961.00 new.

[0] https://www.shi.com/Products/ProductDetail.aspx?SHISystemID=...

I think we get them for much less.

But Netronomes are remarkably cheap, and you can "offload" eBPC to run on the card's custom cores.

Netronome cards are a gigantic PITA to program, and they are not capable of the precise timing that we need.

I don't know where the heck that price came from. The cards are more like $1200 new.

https://www.cdw.com/product/exablaze-exanic-x10-network-adap...

https://m.aliexpress.com/item/4000523308171.html

Ultrascale. Significantly cheaper. Could it be made to work out is it subject to your comments concerning kintex, pcie straddling?

Aside from that did you know you were going to do this when you did verilog-ethernet etc?

Hence the hobbyist. As a professional engineer I mostly care if the investment into expensive hardware (and developing/adapting a software stack) is worth the gains. And I can totally see that adding a $2k card to a bunch of $5k servers can be cheaper than throwing more servers at a problem (especially after saving on power, rack space and cooling). But in that context I'd only consider a used $350 card for dev/eval work. I don't want to tell the customer that his infrastructure was down for a few days because I convinced him to cheap out on a NIC.

> But in that context I'd only consider a used $350 card for dev/eval work. I don't want to tell the customer that his infrastructure was down for a few days because I convinced him to cheap out on a NIC.

I would argue that it strongly depends on the price differential and how resilient your system is. If your system is properly failure tolerant, and you can buy twice as much hardware for the same price by accepting a 20% failure rate (say), then it would be strongly advantageous to buy all used hardware.

Reply on news.ycombinator.com