We compare and contrast two keynote speeches from opening day (Aug 26th) of the 2014 IEEE Hot Interconnects conference, held at Google’s campus in Mt View, CA. The focus is on disruptive innovation giving rise to radically new hardware..or not?
Abstract: The God Box is Dead, by J.R. Rivers of Cumulus Networks
The maturing landscape in both interconnect technology and consumer expectations is leading a time of innovation in network capacity and utility. Complex systems are being realized by loosely coupling available and emerging open components with relevant, consumable technologies enjoying rapid times to deployment. This talk will highlight the historical precedence and discuss these implications on future system architectures.
In his 2014 Hot Interconnects keynote talk J.R. Rivers, co-founder and CEO of Cumulus Networks, said: “The God Box is dead.” That meant not to expect any new revolutionary pieces of hardware (like the IBM 360 maintrame in the 1960s to Apple’s iPhone in 2007) anytime in the near future. “The IT industry is unlikely to create a new piece of hardware that will elevate a market that is mature,” he added.
Author’s Note: The first time I heard about a “God Box” was during the 1998-2000 fiber optic buildout boom. Many start-ups were making “Multi-Service Provisioning Platforms (MSPPs), which they claimed could do any and all networking functions. Hence, MSPPs were referred to as “God Boxes.” After the fiiber and dot com bust in 2001-2002, many of the new age carriers (CLECs) went bankrupt. They (and not the ILECs) planned to run fiber to commercial buildings in support of either n x G/10G Ethernet OR SONET/SDH OC12 or OC48 access. After their demise. most of the MSPP companies disappeared.
Among the reasons Rivers cited to support his position:
1. The decline of research: Government investment in research is declining, while corporate R & D for most companies is much more Design then Research (Google is a likely exception). Venture Capital and corporate research is evolutionary, seeking incremental improvements in existing technologies and products.
2. The (semiconductor?) supply chain is mature. [No clarification provided]
3. Components improve (in price-performance, power, size, etc) faster than they can be re-invented. Microprocessors vs Network Processors was given as an example, where the former has evolved to lower performance gaps with the latter (due to continuation of Moore’s law).
Other observations and advice for researchers and new product developers:
1. “Go places no man has ever gone before. Don’t rehash a design just to make incremental improvements.”
2. Aim at a broad market- especially for components. Niche markets disappear quicker than most think.
3. Tracking a benchmark is better than the current generation of silicon. To clarify this point, J.R. wrote in a post conference email: “The point here is that many companies work on a piece of technology based on the current version of that technology (for example my CPU is way better that today’s i7) without taking into account likely evolutions in the current form as well as planning a roadmap based what the incumbents are likely to build towards. A recent example of this in the interconnect world is the Fulcrum Microelectronics story.”
4. Need a 4 X speed improvement over what’s available today in silicon, e.g. switch silicon (Broadcom), network processors (Cavium), I/O and Bus interconnect (PLX Technology- recently acquired by Avago Technologies). If that can’t be achieved, the newly targeted silicon won’t gain significant market share.
For a systems developer, time to market and flexibility are vital to success. The hardware produced should facilitate rapid provisioning of services and be able to run various types of software/firmware. J
J.R. shared an experience he had as a systems designer at Cisco where he helped develop their Nuova switches: “After releasing some recent ASIC-heavy products, Cisco found that what really mattered to customers was the provisioning system and the software. In the end, that provisioning system could have sat on industry-standard servers and been almost as successful,” Rivers said.
ONF Executive Director Dan Pitt (Alan’s IEEE colleague for 30+ years) asked J.R. to please tell us how his talk applied to Cumulus Networks – a very innovative software company. Mr. Rivers respectively declined to do so, even though the audience (by a show of hands) indicated they were very interested in Cumulus. In fact, it was one of the reasons I attended the first day of the conference! Rivers said he’d talk to audience members off line about Cumulus, but then he disappeared as the morning break began (at least I couldn’t find him then or later).
Prof. David Patterson (UC Berkeley) challenged Rivers by saying that when Moore’s law ends (perhaps in 2020), there’ll be a lot more opportunities for tech innovation.
Abstract: Flash and Chips: A Case for Solid-State Warehouse-Scale Computers WSC) using Custom Silicon, by David Patterson, PhD and Professor UC Berkeley
The 3G-WSC design emphasis will shift from hardware cost-performance and energy-efficiency to easing application engineering. The reliance on flash memory for long-term storage will create a solid-state server that has both much faster and more consistent storage latency and bandwidth. Custom SoCs connected by optical links will enable servers in a 3GWSC to have better network interfaces and be one- to two-orders of magnitude larger than the servers of today, which should simplify both application development and WSC operations. Thus, a WSC of 2020 will be composed of ~400 3G-WSC servers instead of ~100,000 4U servers.
Such a 3GWSC server would certainly be considered a supercomputer, but unlike those for high performance computing, it will be multiprogrammed—for both interactive and batch applications—be fault tolerant so as to be available 24×7, and be tail tolerant to deliver predictable response times.
Prof. Patterson noted that new market opportunities happen when existing engineering technologies change. One recent example is server architecture changes with virtualization and the move to cloud resident data centers.
The Professor observed that Moore’s law (doubling of transisters per same size die every 18 to 24 months) to three years and will slow to five+ years before it becomes defunct for SRAMs and DRAMs in 2020. Flash memories may or may not perpetuate Moore’s law.
[During my interview with him at 2014 Flash Memory Summit, Professor Simon Sze1 said that Moore’s law for SRAM/DRAMs actually stopped around 2000 or 2001, but continued for Flash memories.]
Note 1. History Session @ Flash Memory Summit, Aug 7th, Santa Clara, CA: Interview with Simon Sze, Co-Inventor of the Floating Gate Transistor by Alan J Weissberger
When Moore’s law is dead or slowing to a crawl, there will be new opportunities for innovative custom silicon, according to Prof. Patterson. “When scaling (of transistors) stops, custom chips costs will drop,” he said. There will then be more ASICs developed and that will be supported by new hardware description languages to facilitate new designs.
All that will give rise to the next generation of Warehouse Scale Computers (WSC). It will be imperative to address “tail tolerance” in order to build new computing machines with predictable response times for all operations. Then, we actually may see new God boxes emerge, the UC Berkeley Professor concluded.
In a post conference email, Prof. Patterson wrote: “The God Box is not dead. The iPhone created the smartphone phenomena in 2007 and the iPad led to the tablet in 2010, both of which outsell PCs. The past is prelude, and the recent past suggests more God Boxes are coming in the near future.”
This author agrees as there are likely to be more instances of disruptive tech innovation in the coming years.
In a post conference email, Mr Rivers responded to Prof Patterson’s comment about the iPhone:
“Prof. Patterson’s comments around iPhone and iPad as related to the PC… the point he missed is that in both the smartphone and tablets, Apple’s hardware innovations are not their differentiators and that world is rife with industry standard hardware… there is not a God Box in that industry anymore (I’ll hold up the iPhone as a GodBox for the first 4 years of existence).”
3 other articles on the excellent 2014 Hot Interconnects conference are at: viodi.com