ACADEMIA
IDC's Worldwide Technical Server Taxonomy, 2007
The main differences reside in the following:
- The enterprise server taxonomy looks at all servers worldwide using OEM reported financial data as a constraint to overall market size and performance. Conversely, technical computing is more of a horizontal view with a focus on the overall system purchase, including everything required to make the system work.
- Revenue reported for technical computing systems includes additional items beyond what is counted in enterprise servers (i.e., systems paid for through R&D progress payments), interconnects between nodes, bundle software, and bundled storage.
- Distinct differences exist depending on the deployment environment, including enterprise datacenter, small site/SMB, telco/hosting, and large scale-out implementations, including Web 2.0 and HPC environments. The current market dynamics are very different, with virtualization redefining the infrastructure side of the enterprise datacenter server market and with technical computing currently seeing high growth due to the demand for additional processing power.
- Multicore processors are expected to expand the differences between the two segments due to differences in the ability to make use of the additional cores. Additional SKU proliferation could occur as OEMs work to optimize systems for increasingly variable workloads.
IN THIS INSIGHT This IDC Insight refines the analysis of the differences in how IDC tracks the overall server and technical computing server markets; it adds a comparison of the two market segments using a common definition for revenue and systems. This was accomplished by applying IDC's enterprise server definitions for revenue and systems to the IDC database for technical servers. SITUATION OVERVIEW To elaborate the characteristics of the overall server market and the technical computing market segment, this document will focus on the following areas:
- Top-level server definitions
- Differences in the definitions for revenue, systems, units, and nodes
- Updated comparisons of overall enterprise server and HPC market numbers for revenues, units, and processors
Top-Level Server Definitions IDC categorizes servers into two main groups:
- Servers: This category comprises all server hardware sales for all purposes, applications, and industries, including volume, midrange-enterprise, and high-end enterprise servers (with midrange and high-end servers classified as "enterprise servers"); server blades; and technical servers. Primary server segmentations are based on server class/price band, chip type, and operating system (OS).
- Technical servers: IDC uses the term technical computing and HPC to encompass the entire market for computers used by scientists, engineers, analysts, and other groups using computationally intensive modeling and simulation applications. Technical servers range from small servers costing less than $5,000 to the large-capability machines valued in the hundreds of millions of dollars. In addition to scientific and engineering applications, technical computing includes related markets/applications areas, including economic analysis, financial analysis, animation, server-based gaming, digital content creation and management, and homeland security database applications. These areas are included in the technical computing market based on a combination of historical development, applications type, computational intensity, and associations with traditional technical markets:
- Technical applications tend to expand to fill the available system capabilities based on the requirements to better understand underlying scientific and engineering principles. Characteristics of technical application are high computational capacity, high memory-to-processor data throughput, and fast-streaming I/O. Some examples of the vendors and brands for these systems include, but are not limited to, technical clusters from Cray, Dell, HP, IBM, NEC, SGI, and Sun; SMPs configured for technical environments, such as IBM's p690, HP's Superdome, or Sun's Sun Fire 2500; NUMA architectures such as the SGI Altix series; and vector computers such as the Cray X1 or the NEC SX-8.
- The difference in definitions shows that HPC/technical computing is more of a broad horizontal view and, in many cases, not a direct subsegment of the enterprise server market.
Differences in Approaches to Measuring Revenue At a top level, the overall server market is based on financial data reported by large OEMs in their quarterly financial calls. One important differentiation in accounting methodologies is that the enterprise server team counts nodes based on average or typical configurations while the HPC team counts complete systems. This means that the HPC number includes additional revenue from clustering interconnects, bundled software, node-to-node interconnects, and in-cabinet storage that our enterprise server team doesn't count. That being said, it is not a simple subtraction relationship between enterprise/general server and HPC system numbers. Another difference is large government-built HPC systems where vendors are paid over time on R&D contracts or as professional service fees. Examples include the Earth Simulator in Japan and the U.S. DARPA HPCS computers. These represent individual systems that can cost multiple hundreds of millions of dollars each but do not show up as product revenue in the vendors' financial reports.
- When these sales are not reported as product revenue by vendors, IDC's server group would exclude these sales.
- The technical computing group at IDC counts these as sales once the customer accepts the system, by working with both the vendor and buyer to determine the value of the computer system.
- For these reasons, quarter-to-quarter comparisons of technical computing data with overall server market data can be difficult and misleading. IDC recommends looking at the relationship of the technical computing market segment to the overall server market over an extended period of time to fully understand the market trends.
IDC has determined that the extra revenue included in the technical computing approach represents under 10% of the overall technical server revenue in a given period. Tables 2 and 3 below show this adjustment as HPC adjusted revenue to allow the revenue to be more accurately compared between the two sets of market revenue. Differences in Approaches to Measuring Systems/Units/Nodes Note: IDC plans to change the technical computing definitions in this area to better align with the overall server definitions, using system/nodes to match the server team's system terminology. Both groups use different terms for systems units and nodes. The enterprise server definition for system units is very close to the technical computing group's node count. In technical computing, "system units" refer to the whole system as purchased on a contract. In the case of clusters, this could be one system with hundreds of nodes. The enterprise server group views a system as an individual compute node, so a cluster with 100 nodes would be counted as 100 system units (compared with 1 system unit in the technical computing group).
- Node: This is the size of the physical SMP building blocks in the system. The processor sockets within a node need to share a common main memory to be considered in the same node. Each node normally has its own copy of the OS, although sometimes it is only a kernel of the OS.
Differences in Terms for Processors, Dies, and Cores Technical Computing Definitions for Processor Elements and Cores Note: IDC plans to change the technical computing definitions in this area to better align with the overall server definitions.
- Core: This term is synonymous with central processing unit (CPU). Each core contains the processor and different levels of on-chip cache, depending on the specific vendor implementation.
- Processor element (PE): This is the location on the server system board that the processor dies and/or processor modules are plugged or soldered into to provide access to main memory and other systems resources. (This is similar to the server definition for processor die.)
Enterprise Server Definitions for Cores and Dies Note: The enterprise group's "die" category is very close to the technical computing group's "processor element" category.
- Core: This term is synonymous with central processing unit (CPU). Each core contains the processor and different levels of on-chip cache, depending on the specific vendor implementation.
- Die: This refers to a single piece of silicon containing a single or multiple cores. (This is similar to the technical computing definition for processor element.)
HPC Technical Computing Compared with All Servers Table 1 shows the overall revenue, units/nodes, and dies/processor shipments for all servers, from 2003 to 2006. Table 2 shows the HPC technical server revenue, adjusted revenue, nodes, and processor elements for the same period. Table 3 shows the changing ratio of HPC compared with all servers:
- The overall server market has grown by an average of 4% a year in revenue and 14% a year in processor dies over the last four years. Growth in 2006 slowed some to 2% in revenue and 9% in processor die shipments.
- The technical computing market segment has seen double-digit revenue growth for four consecutive years since 2003 and broke the $10 billion mark in 2006. This growth was primarily driven by the middle and lower ends of the market, with segments with systems valued below $1 million, which is also the space where clusters dominate in sales. Cluster penetration continues to grow at a high rate in technical sites, resulting in major increases in installed processors.
- Note that the growth rates in HPC have greatly moderated in the last year, to single-digit levels for revenue, although processor and node shipments are still growing at a strong rate in HPC.
Table 3 shows the effects of the different rates of change between the overall server market and technical computing.
- On an apples-to-apples revenue basis, technical computing revenue has gone from 11% of the server market in 2003 to 17% in 2006.
- Looking at system node units, technical computing has grown from 8% in 2003 to 19% in 2006.
- Looking at processor dies (processor elements), technical computing has grown from 12% in 2003 to 26% in 2006.
Note that technical computers are growing quickly in node count and processor count due to the strong demand for increasing computational ability combined with the reduction in the efficiency of clusters as compared with traditional HPC solutions (this requires users to buy more processors to do the same level of work). In effect, the technical computing server market segment is Moore's law constrained, and as the cost of compute cycles moves lower, new use cases are found. IDC notes that many Web 2.0 segments, including search and gaming are also driven by this phenomenon.
| TABLE 1 | ||||||
|---|---|---|---|---|---|---|
| Worldwide Server Revenue and Shipments, 2003–2006 | ||||||
| 2003 | 2004 | 2005 | 2006 | 2003–2006 CAGR (%) | 2005–2006 Growth (%) | |
| Revenue ($B) | 46,149 | 49,146 | 51,268 | 52,251 | 4.2 | 1.9 |
| Unit shipments (nodes) | 5,278,222 | 6,307,484 | 7,050,099 | 7,472,649 | 12.3 | 6.0 |
| Processor die shipments | 8,662,823 | 10,134,624 | 11,712,766 | 12,779,159 | 13.8 | 9.1 |
| Source: IDC, 2007 | ||||||
| TABLE 2 | ||||||
|---|---|---|---|---|---|---|
| Worldwide HPC Technical Server Revenue and Shipments, 2003–2006 | ||||||
| 2003 | 2004 | 2005 | 2006 | 2003–2006 CAGR (%) | 2005–2006 Growth (%) | |
| Revenue ($B) | 5,698 | 7,393 | 9,208 | 10,030 | 20.7 | 8.9 |
| Adjusted revenue (to match enterprise revenue definitions) ($B) | 5,128 | 6,654 | 8,287 | 9,027 | 20.7 | 8.9 |
| Unit shipments (nodes) | 411,327 | 734,510 | 1,215,735 | 1,419,221 | 51.1 | 16.7 |
| Processor element shipments | 1,002,905 | 1,657,827 | 2,681,079 | 3,351,843 | 49.5 | 25.0 |
| Source: IDC, 2007 | ||||||
| TABLE 3 | ||||||
|---|---|---|---|---|---|---|
| Worldwide HPC Technical Server Share of Total Server Revenue and Shipments, 2003–2006 (%) | ||||||
| 2003 | 2004 | 2005 | 2006 | |||
| Revenue | 12.3 | 15.0 | 18.0 | 19.2 | ||
| Adjusted revenue (apples to apples) | 11.1 | 13.5 | 16.2 | 17.3 | ||
| Unit shipments (nodes) | 7.8 | 11.6 | 17.2 | 19.0 | ||
| Processor shipments | 11.6 | 16.4 | 22.6 | 26.1 | ||
| Source: IDC, 2007 | ||||||
- Worldwide and Regional Server 2007–2011 Forecast Update: September 2007 (IDC #208812, October 2007)
- Intel Brings Core Architecture to MP Systems with Quad-Core Tigerton Processor and Caneland Server Platform (IDC #208910, October 2007)
- In Pursuit of Petascale Computing: Initiative Around the World (IDC #208623, September 2007)
- Regional Server Reference Book — 2006 (IDC #207910, August 2007)
- IDC's Worldwide Technical Computing Taxonomy, 2007: How It Fits into the Overall Enterprise Server Taxonomy (IDC #207478, July 2007)
- The Impact of the Many-Core Processors on the HPC Market (IDC #lcUS20774307, July 2007)
- Worldwide and U.S. Enterprise Server 2007–2011 Forecast by Vertical Market and Company Size (IDC #207130, June 2007)
- IDC's Worldwide Hardware and Network Infrastructure Taxonomy, 2007 (IDC #206156, June 2007)
- HPC Storage Trends (IDC #207102, June 2007)
- HPC Storage and Data Management: User/Vendor Perspectives and Survey Results (IDC #206967, June 2007)
- HPC User Update: Tokyo Institute of Technology, Global Scientific Information and Computing Center (IDC #205701, May 2007)
- HPC Technical Computing Market Trends and Areas of Growth: Looking at Market Drivers, User Preferences, and Opportunities (IDC #DR2007_SISEJ, February 2007)
- What Will Be the Next Phase Change in HPC, and Will It Require New Ways of Looking at HPC Systems? (IDC #205025, January 2007)
- Worldwide High-Performance Technical Computing System 2005 Vendor Shares: 2005 Market Census Results (IDC #204866, December 2006)
- DARPA Selects IBM for Phase III of HPCS Program (IDC #204606, December 2006)
- DARPA Selects Cray for Phase III of HPCS Program (IDC #204582, December 2006)