Contents |
Some Good News for HP
In the most far reaching infrastructure announcement of the last 18 years, HP, ARM, and Calxeda introduced the Moonshot project that will probably change the data center as we know it. At heart this is the enablement of ARM as an architecture in data center servers. This throws the final stone to break the Microsoft/Intel duopoly.
History Repeats?
The fundamental business driver of this dramatic announcement is the consumerization of IT. The last time this happened was the introduction of Windows NT by Microsoft in July 1993. Intel and Microsoft had won the PC battle against Apple, Commodore, and the other personal computer manufacturers. The volume of x86 chips was driving down the cost of chips. The NT operating system was well specified. Initially it was deployed in simple areas in the data center such as file/print. It was functionally enriched; it replaced NetWare as the dominant operating system in the data center, and it expanded to address almost all data center workloads.
Now the Intel platform is used for all computing except for the very high-end, where IBM Power7 systems like the 32-way 795 currently dominate. The only challenge to this duopoly was from Intel, itself, when Itanium was introduced in 2001 (it was rescued from itself by AMD introducing 64-bit support on x86 without requiring a code conversion), and the introduction and gradual adoption of Linux. The bottom line is that x86 architectures from Intel and AMD drive over 80% by value of the world's servers.
History looks to be repeating itself nearly two decades later. This time mobile computing is the consumer catalyst. ARM designs outcompeted Motorola, Intel (the ATOM x86 chip), and other companies in producing processors for smartphones that sip power. Apple’s entry into the smartphone and now the tablet market was based on the ARM chipsets. The low-end of the PC base is being eaten by tablets running on Apples iOS and Google Android platform. Already more Web pages are accessed by smartphones than PCs. The ARM ecosystem of developers has embraced a high-volume low-cost model that provides applications at 10% of the cost of traditional PC applications. Momentum is driving the adoption of ARM processing across all mobile computing and upwards into the PC market. Modern development and leading edge application ideas are coming from this ecosystem.
ARM: A Server Architecture in the Making
The ARM processors have a different DNA, and are particularly efficient at power consumption. A 4-core Calxeda ARM 1.1 GHz processor takes about 5-6 Watts of power fully loaded but drops down to almost zero when not active.
Figure 1 shows the processor-only comparison, using a simple Watts/GHz metric. It shows that the Calxeda ARM processors:
- Are 6-10 times more power efficient than x86 or Power 7 architectures for the processor alone;
- Will need four times more processors than Intel alternatives (Calxeda has half the clock speed and half the number of cores);
- have a lower heat profile that allows a higher density of computing that can be provisioned by traditional power and cooling facilities.
The metrix used does not take into account the power taken by DRAM and other volatile memories or the SATA or SSD drives that are configured into the total solution. These are approximately the same for all solutions and will significantly decrease the total system power/cooling impact.
Missing ARM Elements
The current ARM specification lacks virtualization and security extensions, although ARM say these are areas it intends to address in the future. This initial lack of functionality will not be so important to some very large social service providers, but will be required for ISV and enterprise support.
Perhaps the most important piece missing is an adequate non-volatile storage strategy. HP discussed its future Memristor aspirations and did not tackle the need to make NAND storage an architected extension of main storage. The SSD solution is okay for some applications but still carries the SCSI ptotocol and overheads. It is not adequate for large social sites that are buying large amounts of NAND flash from Fusion-io and other vendors.
Workloads that match the ARM technologies
In general, the workloads that will fit this initial instantiation are:
- Those that do not need virtualization and have their own security built into the applications;
- Those that are naturally scale-out (traditional Web applications);
- Those that are components of very large social and search workloads that are generated in house, and have well documented source code.
Futures
Like initial Windows NT and Intel Servers, the initial ARM products are aimed at niches, and will take years to expand into the marketplace. The power difference is important to some customers but is not a sustainable differentiation. Intel and AMD are very likely to match and even exceed ARM power efficiency. The key business driver will be the lower cost of ARM processors (a factor of 2+) as the economies scale become greater with the relentless drive towards mobility and better applications. It is unlikely that Intel will make it in mobile with Atom, and unlikely that Intel will be allowed to buy ARM. Intel, AMD, and Microsoft will remain good contributers to technology and enterprises; the only difference is that the margins that come with market domination will not be available to these companies. The final act will be the porting of Windows to ARM, 5-10 years from now.
Action Item: CIOs and CTOs of all large organizations need to understand where ARM servers fit and start to plan to take advantage of the lower costs of infrastructure and applications. The migration will take a decade, but like the previous migration will be driven relentlessly by the consumers and businesses using ARM processors as the default.
Footnotes:
