Peter Christy

Intel's AppUp: A Hybrid Cloud for SMB (a clever example of the power of virtualization!)

Like EMC’s VCube introduction last month (VCubes — a really clever idea) Intel’s introduction of AppUp -- an SMB offering leveraging virtualization -- is at once pretty obvious and quite profound. To understand this duality you first need to have a realistic picture of the discouraging nature of SMB IT today. For example, in an auto dealership you’ll probably find many different Windows applications each running on a separate PC or small server. If one of these servers fails or needs an upgrade it means ordering and configuring a new and matching server, carefully reinstalling the application and then grabbing the old data. The power of the server is rarely an issue. With today’s multi-core systems, a single small server typically has more than enough power to roll up and run all these applications on a single, shared machine. The question has always been how to into a more attractive and modern Cloud-like vision with as little disruption as possible.

Intel's AppUp is new and innovative model for delivery IT services to an SMB, developed and operated by Intel, and delivered to SMB customers by Intel business partners (including SP’s, MSP’s and ISV’s). Technically, the customer gets a small, virtualized server and runs the applications they choose (selected from an Intel “App Store”) as virtual machines on that server.  Software and updates are delivered over the network by Intel.  The user’s data resides locally on the server (providing robustness in the case of communications failures) and is backed up into the Cloud.  The overall system is maintained and supported by the business partner.

The AppUp business model is unquestionably modern. There are no up-front costs for the customer (the server is owned by the business partner). The customer pays for the apps on an ongoing, as used basis, based on the number of enabled users.

So on the one hand, AppUp is a straightforward usage of existing ideas, but on the other hand, it’s revolutionary.

• From the point of view of the customer, application selection and installation is dramatically simplified: click and go; and hardware maintenance is greatly simplified as well since the applications share a server, and since any dependence of applications on specific server hardware is eliminated.
• From the point of view of the SP, MSP or ISV who provides support to the customer, the task is greatly simplified. The hardware is generic, based on an Intel definition, and then implemented and sold by multiple hardware OEM’s.
• Usage is moved to an on-demand, pay-for-what you use model. All costs are fees for use rather than up-front capital outlays. There is no need to buy hardware or software up-front. In Intel’s business model, the business partner owns the hardware and lends it to the customer.
• From the point of view of the ISV it’s a way of leveraging many of the benefits associated with the Cloud (no setup, on-demand access, pay for what you use) without changing the software architecture.
• Since the applications execute on the customer’s premises, it avoids entirely the issues of having sensitive data in the Cloud as well as making the solution robust with respect to communications failures.

AppUp is another good example of how a relatively simple use of virtualization enables quite dramatic changes in IT.

Intel's 22nm Tick

Intel detailed their 22nm process and declared it ready for prime time (volume production). There were a lot of interesting technical details to geek over (in this case a new vertical "fin" structure for the transistors). As always, a new process shrinks the size of a given microprocessor and that means more can be produced from a single wafer and that brings the cost down (even taking into account the 5-10% increased cost of the new process). More impressive perhaps is the fact that the 22nm process yields a 50% reduction in the dynamic power consumed (for the same processor design and same clock rate). Although we might like to think of this as the relentless result of "Moore's Law" in fact it represented ten years of hard work. Intel can see continuing progress at least for the next couple of processes (each a 3 year cycle) but one can assume only as the result of decades of hard work. There is nothing simple about keeping Moore's Law going.

Tick/Tock the Relentless Intel Clock

It's very easy to understand Intel's general roadmap (which is not to suggest that it's easy to execute!): Every three years Intel introduces a new process characterized by the smallest feature size. At the moment a 32 nanometer process is superseding 45 nm, and 22nm is well on the way (complete microprocessors in fab now, memory parts made a year ago). As the feature size gets smaller you can make more complex chips (use more transistors), reduce power, increase processing power, or reduce cost, but not necessarily all at once. As time goes on fab cleanliness and process control improve as well, making it possible to fab larger chips. When a new process is introduced, in the first "Tick" phase Intel moves an existing design onto the new architecture (a "shrink"). Because the chip gets smaller it gets cheaper so Intel's margin increases or they can choose to lower the prices, as the market requires. As part of the shrink, the team performs other optimizations (improves timing or yield for example). In the second —"Tock" — phase, Intel introduces a new microcomputer architecture onto what is now a very stable and high volume process. At IDF Intel introduced the next architecture, code name "Sandy Bridge," as the Tock of 32 ns. Sandy Bridge integrates still more of the whole computer onto the CPU chip. Specifically, Sandy Bridge has the most elaborate integration of the PC graphics GPU by far. Many low-end modern PC's use an Intel GPU that is integrated with the "chip set." With Sandy Bridge the GPU moves on the CPU chip and is built with the most advanced process (in the past the chip set might use an older process) thereby increasing the performance both due to process and due to tighter CPU integration. Next year Sandy Bridge will be moved onto the new 22 nm process and get smaller, faster and more power efficient and Intel will be well down the path of the next process. For most of Intel's competitors, Tick/Tock must be like Chinese Water Torture.

Higher Virtual Machine Density

Just like clockwork (a joke for Intel followers) Intel introduced the Westmere server chips. These chips are fab'ed on Intel's 32ns process whereas the preceding generation of Nehalem was fab'ed on 45 nm. Before you dismiss that as semiconductor trickery, remember that a smaller process makes a design smaller, faster and more power efficient, which means for a multi-core CPU Intel can pack in a lot more processing power at more or less the same fab cost (die size) and within the same heat and power footprint (something a server really cares about). Some of the power comes from the intrinsically faster circuits, some comes from the intrinsically lower power circuits so you can run the CPU faster at the same power consumption, and some comes from the ability to pack in more cores (six) at the high end. Intel drives relentlessly on a "tick/tock" process (the joke) which introduces a new process every two years (they argue they could make even more money if they could introduce a new process every 18 months). Now that we're well into the multi-core age, and with applications like virtualization that make very good use of more cores (to run more VM's on the same hardware platform) the rate of increase in performance is pretty surprising. Cisco's UCS posted the best numbers on VMware's VMmark benchmark (the best available test for virtualization throughput on a given platform), and recorded a 42% improvement over the previous Nehalem generation, only a year ago (the fine print: the UCS score was for a 12 core configuration; there are a few higher scores for 32 core systems and above). For a virtualization use that means fewer servers, more VM's or both. Cisco's best-of-breed score certainly validates that UCS is a competent high-complexity server system (for those who might have doubted Cisco building servers for that reason), and it may also show that Cisco "secret sauce" that lets them include more DIMM parts than is possible with a conventional Nehalem server is starting to bear real fruit due to the rapid increase in chip processing power (which it stands to reason needs to be balanced with a larger memory).

Processors Still Matter

As an old microprocessor geek I have to admit that I'm looking forward to the Intel Analyst Summit this week. It's hard to exaggerate the impact that today's virtualized, multi-core CPU's are having on how we do computing and even how we think about it. It was also fascinating to note that the Apple iPad incorporates an Apple processor from Apple's purchase of Palo Alto semiconductor. PA Semi is the creation of Dan Dobberpuhl. I met Dan back in my DEC semiconductor days and hold him in the highest regard as a true semiconductor wizard. After my DEC semi days, Dan and Rich Witek did a series of innovative chips including a very low power ARM processor that Intel acquired when it bought most of DEC's semiconductor assets. The rumor was that PA Semi was formed (back in 2003) to build very low power versions of IBM's Power architecture aiming precisely at Apple and the Mac, but then Apple embraced the X86 putting all that into question. I would have to assume that Apple is using PA Semi to create extraordinarily power efficient CPU's which when combined with software with very careful power management can translate directly into products with battery weight and higher battery life, something we're all interested in. Apple has already shown an ability to exploit advances in Intel CPU's quickly so the question is how much they can kick that up with their own CPU's. Engineering and releasing your own CPU is too expensive for most device companies, but then Apple ain't like most. — pchristy@irg-intl.com

Hafnium Rocks Your World!

The latest Intel Developer Forum demonstrated clearly that progress in the world of microprocessors hasn't slowed down. The keynotes are available on the Web, including a great discussion with Gordon Moore who has to be one of the nicest and most thoughtful technology barons the world has ever seen. Over the last couple of years Intel has retrenched back to a company that really focuses on the core computer offerings and that roadmap has lots of life in it that includes continuing architectural improvements to support bigger picture topics like virtualization, the continuing evolution to multi-core architectures as the means of increasing computer power while saving on electric power, the pending integration of high-performance graphics onto the CPU as well as the integration of a lot of standard radio technology. Moore's Law progress won't cease in the next five years for sure. BTW, Hafnium [Hf] is an element that is at the core of Intel's approach to continue to make transistors smaller and smaller.

As The Chip Turns

Steve Jobs might be strange but he sure isn't stupid.  When Apple moved from IBM to Intel for CPU's for the Mac a lot of people wondered whether this was in the best interests of the Mac cultists since Intel iconoclasm was always part of that. Apple understood that since Intel now had the power minimization religion the Intel processors would overtake IBM in power efficiency. It turns out there are other powerful benefits. When Intel brings out a new CPU now it tends to have architectural visibility (e.g. more cores, virtualization capability, …). The pipeline between Intel and the customer on the Windows side is long and hard often requiring some support in Windows and then by the ODM and computer vendor.  In contrast, here in Silicon Valley Apple and Intel have a small team of bright engineers and if it is to Apple's benefit they can easily be the first to market with a MAC that leverages a new Intel CPU. Intel loves it because the not-to-price-sensitive MAC purchaser is a great way to take advantage of Intel's newest. So expect to see some share growth for Apple in areas like laptops especially given the fact that you can run Windows virtualized if you want.  And speaking of competition, we would guess that AMD is in for a pretty hard time now that the CPU competition has moved to mine-has-more-cores-than-yours (naa, naa, naa!). AMD is better than Intel at many things but getting new processes into volume production is not and is not likely to be one of them. Smaller features and more transistors is the way you get more cores, and Intel seems destined to win those battles for the foreseeable future.