back to article Amazon gobbles chips firm Annapurna to speed up cloudy data centers

Amazon has agreed to buy a secretive Israeli chipmaker in an apparent bid to build new custom tech for its massive data centers. Israeli financial newspaper Calcalist was the first to report that Bezos & Co had snapped up Yoqneam, Israel–based Annapurna Labs, and Amazon later confirmed the purchase to the Wall Street Journal …

  1. Voland's right hand Silver badge

    There are other reasons for the networking costs

    Well, while I have considerable respect for James, I have to disagree with him somewhat.

    Networking equipment itself is also generally following the Moore's law. Networking equipment cost != networking costs and this is where he is being a bit disingenuous.

    Most cloud providers tried (inclusive of Amazon) to do DIY networking done by software engineer as a replacement for proper resilience which is using algorithms and protocols for which there is a mathematical proof of convergence.

    For Amazon this cheapskating experiment ended up in the majestic clusterf***s of 2013 when they had several zone-wide outages due to unhandled network failures.

    From that moment onwards Amazon has been buying proper equipment to augment the DIY. This expense however is a natural result of a failed networking experiment (for which James should do some Mea Culpa) combined with a failed design (again, Mea Culpa needed) and the cost of fixing it (again - Mea Culpa James). It is not your normal purchasing and cost curve to which Moore law should apply so he should stop complaining.

    1. Anonymous Coward
      Anonymous Coward

      Re: There are other reasons for the networking costs

      "Networking equipment itself is also generally following the Moore's law."

      Is it? Really?

      Can you back that up?

      1. Voland's right hand Silver badge

        Re: There are other reasons for the networking costs

        Can you back that up?

        Cost per GBit forwarded is in line with it. Have a look at the 1G and 10G port cost curves for example. When 1GBit came out 15 years ago it was in the range of several k per router card (the original GSR trident and its Juniper equivalent) and ~ 2K per NIC. It is now 10£ with the curve complying with Moore's law all along. The same goes for mid-range routing/switching processor costs normalized by PPS or GBit. 10GB ports and NICs are also going down the same way.

        The problem is that the requirement for Gbit forwarded has grown more or less with the same curve (or faster) resulting in the end-cost of networking equipment required per rack, per datacenter and per company being constant or growing. That is expected - it "serves" a Moore's law governed population of compute.

        Asking for networking kit cost in a datacenter to drop in absolute numbers is a bit disingenious as this in fact implies a faster than Moore's law drop.

  2. well meaning but ultimately self defeating

    Pretty scary

    What I see as the most interesting aspect of this acquisition is our friends at AWS have bought in a tonne of semiconductor designers. Whilst they are focussed on networking at the moment, do you honestly think that Dr Vogels has not structurally disassembled his workload and identified functional use-case where an ARM based processor will have a price/perf advantage. If I was Intel right now, I would be a little nervous as if the economics of this work our correctly from a unit cost perspective you can be sure that the chocolate factory will be replicating.

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Biting the hand that feeds IT © 1998–2020