News & Analysis
Bechtolsheim brainstorms on next networking wave
Rick Merritt
10/17/2012 9:46 AM EDT
Looking for a few good chip makers
Click on image to enlarge.
Fewer and fewer companies will make bigger and bigger volumes of merchant chips, Bechtolsheim said. Arista Networks, the switch maker he co-founded, uses only merchant chips from the likes of Fulcrum (now Intel) and Broadcom.
“Even large companies don’t have the volume to drive these things effectively,” he said, a backhand reference to his former employer Cisco Systems, still one of the world’s largest ASIC designers.
Next: Top 10 list
Click on image to enlarge.
Fewer and fewer companies will make bigger and bigger volumes of merchant chips, Bechtolsheim said. Arista Networks, the switch maker he co-founded, uses only merchant chips from the likes of Fulcrum (now Intel) and Broadcom.
“Even large companies don’t have the volume to drive these things effectively,” he said, a backhand reference to his former employer Cisco Systems, still one of the world’s largest ASIC designers.
Next: Top 10 list
|
|
|
|
|
Navigate to related information
rick.merritt
10/18/2012 2:39 PM EDT
Does Andy have it right? wrong? What do you think?
Sign in to Reply
elPresidente
10/19/2012 4:40 PM EDT
That "100,000 chips or it ain't worth it" has been the mantra for decades - nothing new.
100Gb/s FPGAs have been around for a couple of years now. Again, nothing new.
No mention of 160Gb/s, or 400Gb/s, which are upcoming. Sounds to me like he's trying to solve procurement channel problems in his investment portfolio, vs being a visionary driver of industry.
Sign in to Reply
Philip Papadopoulos
10/26/2012 5:43 PM EDT
There is a fundamental utility limit of the network entering/exiting a server -- that's memory bandwidth. 400Gb/s is 50GB/sec, which is on par with achievable total memory bandwidth of common servers. When your network becomes faster than memory speed, you have to wonder about practical utility of that extra unusable capacity. I would argue that the practical limit is closer to single memory channel speed, not aggregate system memory bandwidth. Today, a 2100MHz DDR3 Channel is 17GB/sec ~ 136Gbps.
Sign in to Reply
krh
10/20/2012 12:33 AM EDT
$$ per bit...Andy has it right for the next 5 years...
Sign in to Reply
rick.merritt
10/22/2012 4:39 PM EDT
Two great and opposing views here.
I'd love to hear some more thoughtful opinions on whether Andy has the future of networking right and why or why not.
Sign in to Reply
Philip Papadopoulos
10/26/2012 5:36 PM EDT
1) Andy's talking about the server space. I think he's bang on. The adoption curves look to be very similar to past networking tech uptakes.
2) He's not talking about the desktop space and/or wireless space. GigE is now pressed down on el-cheapo DIY motherboards, but you don't really need greater than 1GbE in your house for the next 1/2 decade) Hence, you cannot necessarily make the same volume argument for bringing down the price of 10GbE at the same rate.
That means that the absolute price of 10G won't drop as quickly as it did for 1G, but it will get there. Of that, I have no doubt. Afterall, as people rely on network (buzzword cloud) services, the number of servers continues to rise. And Ideally you keep constant BW/core as you deploy big systems. As core count rises, so do your network bandwidth needs.
Sign in to Reply