<:#1704,9360><:f>Since its first introduction in 1986, PA-RISC has been a technology leader offering high performance, scalability, and support for both technical
<:P13,1,>and commercial environments. Among its forward looking features has been the ability to support 64-bit virtual addressing. To guide customers who are beginning to consider the future evolution and impact of 64-bit computing, Hewlett-Packard Compa
ny summarizes its view of these issues in this paper and relates them to HP's long term plans for evolving the PA-RISC architecture.
<:#284,9360><:f>
<:#2556,9360><:f>HP's progressive implementation of the PA-RISC architecture provides an excellent example of specific 64-bit feature enhancements which have been delivered to users when they can provide real benefits without incurring undue cost. Examples include 64-b
it floating point registers and 64-bit data paths. PA-RISC will be further enhanced to provide 64-bit flat addressing and 64-bit integer registers at mid-decade, well before these features become a broad market requirement. These enhancements will be evol
utionary and will enhance the current capabilities by providing forward binary compatibility for current PA-RISC systems. With this strategy, PA-RISC is able to provide optimal cost/performance for customers today while offering a clear path for meeting fu
ture customer requirements.
<:#284,9360><:f>
<:#1136,9360><:f>The move towards 64-bit computing, however, faces a number of important challenges in the marketplace. HP believes that the inability to meet these challenges will prevent a broad move to 64-bit computing until late in this decade. Most industry obser
vers agree with this assessment. To become a reality, 64-bit computing m<:f>ust overcome the following obstacles:
<:#284,9360><:f>
<:I0,0,720,0><:#284,9360><:f> * <+#>Microprocessors.<-#> 64-bit microprocessors are more complex and costly to manufacture.
<:#284,9360><:f>
<:I0,0,720,0><:#568,9360><:f> <:f>* <+#>Systems software.<-#> Standards for 64-bit computing do not exist for open systems software -- for operating systems, compilers, graphics, and other middleware.
<:#284,9360><:f>
<:I0,0,720,0><:#852,9360><:f> * <+#>Application software.<-#> Without software standards and broad demand for 64-bit features, the emergence of 64-bit application software will be gradual, extending out through the end of the decade.
<:#284,9360><:f>
<:I0,0,720,0><:#568,9360><:f> *<+#> System cost.<-#> Moving to 64-bit data structures can increase the requirements for main memory and disk capacity, increasing system cost.
<:I0,0,720,0><:#284,9360>
<:I0,0,720,0><:#568,9360><:f> * <+#>Performance.<-#> Without larger caches, 64-bit systems may actually run slower than 32-bit systems.
<:#284,9360><:f>
<:I0,0,720,0><:#568,9360><:f> * <+#>Address flexibility<-#>. In the near term, very few applications will require the flexibility of flat 64-bit addressing.
<:#284,9360><:f>
<:#1136,9360><:f>By the end of the decade, cost-effective 64-bit computing will be enabled by continued advances in silicon density, software standards, memory and disk capacities, and by the emergence of 64-bit application software. HP is committed to providing custom
ers with high-performance today and migration paths to the future through PA-RISC.
<:#284,9360>
<:#1704,9360>PA-RISC will provide forward binary compatibility to protect customers' investments in hardware and software. In association with the Precision Risc Organization -- a group of industry-leading companies with $70 billion in electronics-related revenues from
Europe, North America, and the Pacific Rim -- HP is actively working to expand the range of customer choices for purchasing PA-RISC systems. Hewlett-Packard's state-of-the-art PA-RISC will continue to evolve to meet the needs of the customer.
<:#284,9360><:f><:f>
<+B># # #<:p >
<:#284,9360><:f><+!>PA-RISC: LEADERSHIP ARCHITECTURE TO MEET CUSTOMER NEEDS<-!>
<:#284,9360><:f>
<:f>In 1986, Hewlett-Packard Company became the first company to fully commit its computer product lines to the RISC computing architecture, starting the high-performance Precision Architecture-RISC family. Since that time, the scalability of PA-RISC has b
een demonst<:f>rated through three generations and seven implementations of this architecture. PA-RISC has demonstrated leadership capabilities in several areas, including being the first RISC architecture to support 64-bit virtual addressing. To this day
, HP's PA-RISC remains the only RISC architecture to have implemented a chip with full access to 64-bit virtual address space (in a previous implementation).
<:#284,9360>
<:#1420,9360>With the visibility of 64-bit computing in the industry press, many customers are beginning to ask if they should begin moving to 64-bit computing. The question of 64-bit computing is reviewed in this paper with respect to its status, the barriers to its e
mergence, and HP's plans for PA-RISC. PA-RISC will be enhanced to enable 64-bit computing at mid-decade, well before 64-bit computing becomes a broad market requirement.
<:#284,9360><:f>
<:#284,9360><:f>
<:#284,9360><:f><+!>REALITY CHECK: WHERE IS 64-BIT COMPUTING?<-!>
<:#284,9360><:f>
<:#1420,9360><:f>HP's commitment to provide an enhanced 64-bit PA-RISC can be better understood once the facts about 64-bit computing are separated from the marketing hype. The general question of 64-bit computing can be clarified for today's environment by looking at
three elements: a comparison of "architecture" versus "implementation," the issue of systems software, and the availability of application software.
<:#284,9360><:f>
<:#284,9360><:f><+!>"Architecture" versus "Implementation"<-!>
<:#1136,9360><:f>When examining the reality of 64-bit computing, it is important to distinguish between claims of "architecture" and the realities of "implementation." While the architecture may support a feature, the specific chip implementation might not. Full 64-bi
t computing requires a 64-bit implementation of each of the following characteristics:
<:#284,9360><:f>
<:#284,9360><:f> * Virtual memory address space,
<:#284,9360><:f> * Physical address space,
<:#284,9360><:f> <:f> * Floating point register size,
<:#284,9360><:f> * Integer register size, and
<:I0,0,2160,0><:#284,9360><:f> * Width of the data buses (cache, memory, and internal data paths).
<:#284,9360><:f>
<:#1704,9360><:f>The distinction between "architectura<:f>l" capabilities and actual chip "implementations" is compared for different vendors using these characteristics in Figure 1. As the attached figure shows, no vendor currently offers a full 64-bit RISC implementa
tion at this time. Furthermore, even architectures which are not being promoted as "64-bit" have implemented certain 64-bit features which provide benefit to the user. For example, all leading-edge<:f> processors already provide 64-bit registers for float
ing point operations.
<:#284,9360>
<:#284,9360>
<:#284,9360><:f>
<:#284,9360><:f>
<:#3692,9360><:f>HP's progressive implementation of the PA-RISC architecture is an excellent example of an evolution in architectural implementation to meet customers needs. PA-RISC is a 64-bit capable architecture, and certain 64-bit features have already been impleme
nted in PA-RISC products when they can provide real benefit to the user and do not cause discontinuities in their software environment. HP has implemented 64-bit registers for double-precision floating-point operations in all PA-RISC processors since their
introduction in 1986. Likewise, wider data paths have been provided in PA-RISC implementations and are not exclusive to 64-bit computing. PA-RISC has implemented 64-bit data paths/busses in a range of implementations for both floating point and integer d
ata. More significantly, HP has led implementation of 64-bit cache data paths to support superscalar operations, as shown by the HP 9000 Series 840 which was introduced in 1986! It should be obvious from HP's progressive implementations of PA-RISC that th
e question is not whether the 64-bit architectures are technically possible, the question is whether the current implementation of 64-bit features benefit the customer in today's environment.
<:#284,9360><:f>
<:#284,9360>
<:#284,9360><:f><+!>The Systems Software Issue<-!>
<:#1420,9360><:f>The slow emergence of 64-bit applications in the 1990's is primarily due to key barriers in the area of systems software. These barriers include a lack of 64-bit software support for operating systems and "open systems" software standards, including la
n<:f>guages, compilers, libraries, and application programmer interfaces (APIs). Each of these barriers represents a significant issue to 64-bit computing reality.
<:#284,9360><:f><:f>
<:#1141,9360><:f><+!>Barrier: Operating System Support.<-!> The issue of operating system support for 64-bit operation is very significant to the adoption of 64-bit computing. According to Robert B.K. Dewar, professor of computer science at New York University (NYU),
and Matthew Smosna, systems manager at NYU's Courant Institute,<:f>
<:I0,0,720,720><:f> "...operating system limitations (such as 32-bit kernel data structures) mean that many users of 64-bit machines, like MIPS and Alpha, simply will not be able to use the 64-bit capability. For example, both DEC's own VMS and the Microsoft NT system, w
hich DEC regards as an important option for Alpha, support only 32-bit operation."<:F
@Footnote@<:f><:f> "Taking Stock of Power Processing" by Robert B.K. Dewar and Matthew Smosna,
<+#> Open Systems Today<-#>, February 1, 1993, p.49. <[>Hereinafter referred to as "Power
Processing"]
@Footnote@<:f>
>
<:f><:f>
<:#284,9360><:f>
<:#1704,9360><:f>This lack of operating system support for 64-bit operation is a critical barrier which prevents access to 64-bit functionality. Buying "64-bit" hardware to use with the current operating system limitations is analogous to purchasing an airplane to driv
e on the<:f> local streets. The operating system support is not matched to the technology. Almost all of the operating systems in widespread use today or that are planned for use in the near future (Windows, Windows NT, UNIX, VMS...) are still defined as
"32-bit."
<:#284,9360><:f><:f>
<:#2561,9360><:f><+!>Barrier: Open Systems Software Standards.<-!> A second important barrier to the adoption of 64-bit computing is the lack of open systems software standards, which extends beyond the operating system into software such as languages and graphics APIs
. This lack of standards is a critical issue for today's user of heterogeneous system environments. Today's enterprises require that a variety of systems from
<:f>different vendors work together using industry standards to get the job done. The lack of 64-bit standards to accomplish this task is evident when viewing Figure 2. While this list may not be exhaustive, it points out that the common industry standard
s in use today are based solidly on 32-bit technology. More importantly, most of these software standards have yet to define 64-bit extensions, much less to determine a replacement 64-bit standard!
<:#284,9360><:f>
<:#852,9360><:f>A particularly significant case which demonstrates this lack of standards is the popular C and C++ programming languages. According to Dewar and Smosna, due to the lack of standards for 64-bit computing...<:f>
<:I0,0,720,720><:f> "...in C and C++, you are at the mercy of the compiler developer. At least three competing viewpoints exist on what size 'int' and 'long' <[>integer sizes] should be on the new 64-bit machines...The most disturbing thing about this argument is that,
at least for the moment, manufacturers don't agree, and we are afraid that the arrival of 64-bit architectures will be accompanied by a mess of software incompatibilities, confusing compiler options and difficult porting problems."<:F
@Footnote@<:f><:f> See "Power Processing," <+#><-#>p.53.
@Footnote@<:f>
>
<:f><:f>
<:#284,9360><:f>
<:#1420,9360><:f>HP is actively involved in seeking industry-wide standard definitons for 64-bit C programming extensions. HP is represented in two separate efforts: an industry cooperative committee formed in April 1992 and the ANSI C Committee. These joint meetings
have revealed significant differences between parties on definitional issues. A common ground for an open systems standard for 64-bit C programming extensions has not yet been found.
<:#284,9360><:f>
<:#1988,9360><:f>The lack of standard C programming language definitions for 64-bit computing is merely the tip of the iceberg in the larger problem of open systems standards. There is a wider need for 64-bit standards in languages, compilers, libraries and APIs to ena
ble open systems computing. Efforts like the common open software environment process, which joins HP with other UNIX vendors<:f>, is aimed at promoting a consistent set of APIs that will run across all vendors' systems. But all of these standards which e
nable open<:f> systems and portability are based on 32-bit technology. The 64-bit open systems standards are, at best, still under development.
<:#284,9360><:f><:f>
<:#1993,9360><:f><+!>Overall Software Impact.<-!> The missing 64-bit software pieces -- operating system support and open system standards for languages, libraries, and APIs -- will slow the development of true 64-bit application software and inhibit the adoption of th
e 64-bit computing. HP believes that 64-bit software standards and their broad acceptance are some years away. Until true 64-bit standards for open systems are established, customers with heterogeneous systems on their network risk having to support incom
patible, shifting software implementations of 64-bit computing.
<:#284,9360><:f>By far the most serious question of 64-bit computing is: What customer benefit is available?
<:#284,9360><:f>
<:#1420,9360><:f>The customer will benefit from 64-bit computing when 64-bit application software is available and usable in today's open systems environment. Industry analysts David M. Smith and John Morrell of International Data Corporation (IDC) in Framingham, MA, r
eviewed the intent of independent software vendors (ISVs) in their November 1992 report on 64-bit computing. Their interviews show that...<:f>
<:I0,0,720,720><:f> "...the majority of ISVs are not seeing demand among their users for specific 64-bit functionality, nor do they see many inherent capabilities of which their software products can take advantage. The majority of ISVs will port existing 32-bit software
to the 64-bit platforms without taking advantage of the 64-bit features."<:F
@Footnote@<:f><:f> "Sixty-Four Bit Computing: A Bit Ahead of Its Time?" by David M. Smith and John
Morrell, International Data Corporation (IDC # 7175), November 1992, p.7.
<[>Hereinafter referred to as "Sixty-Four Bit Computing"]
@Footnote@<:f>
>
<:f><:f>
<:#284,9360><:f>
<:f>This lack<:f> of true 64-bit application software (without specific 64-bit functionality) points out the general lack of customer benefit available to the current purchaser of this technology. Furthermore, in commercial markets the adoption of true 64-
bit application software is hindered by the need for 64-bit middleware. Smith and Morrell note that "...64-bit platforms would not have value to the mass commercial market until the higher-level software <[>i.e. 'middleware'] utilized 64-bit features."<:F
@Footnote@<:f><:f> See "Sixty-Four Bit Computing," p.7.
@Footnote@<:f>
>
<:f><:f>
<:#284,9360><:f>
<:#2272,9360><:f>The transition to 64-bit computing will be a gradual evolution, as customer requirements increase and as 64-bit application software becomes available. An approximate timeline for this transition is provided by system shipment estimates from the IDC re
port on 64-bit computing, which is shown in Figure 3. The shipment estimates show a slow transition to 64-bit systems, with the vast majority of shipments in this decade being comprised of 32-bit implementations.
<:f> A key element impacting this transition for most users is that 64-bit software environments which can begin to support 64-bit applications are not expected to become widely available until the 1996 timeframe (according to IDC). The majority of 64-bit
software applications will arrive even later.
<:#284,9360>
<:#284,9360><:f>
<:#284,9360>
<:#1136,9360><:f>With this timeline in mind, HP will provide the first enhanced 64-bit PA-RISC systems at mid-decade, well before 64-bit computing becomes a broad market requirement. Furthermore, HP will provide these enhanced implementations when customers can realize
the practical benefits of this technology, without creating discontinuities in the customer's software environment.
<:#1988,9360><:f>HP's careful evaluation of user needs has helped to make PA-RISC a leadership architecture which provides high-performance in a cost-effective manner to the user. An important point is that PA-RISC provides these capabiliti<:f>es in a manner that is us
able by today's software applications. These capabilities are not vague promises of "future benefit," but represent real value to the user in terms of productivity with their software applications. The cost-effectiveness of PA-RISC is apparent when consid
ering three areas: system cost, performance, and addressing flexibility.
<:#284,9360><:f>
<:#284,9360><:f><+!>System Cost: Optimizing Memory and Disk Resources<-!>
<:#1704,9360><:f>PA-RISC provides high-performance, cost-effective technology by focusing on the proper choice of 32-bit or 64-bit data representations for today's user. This cost issue has been missing from most discussions on 64-bit computing and has a very real cons
equence to the user. The choice of 32-bit or 64-bit data representation can have significant implications for memory and disk capacity. Brian Case, writing for Microprocessor Report, puts this issue into perspective for the user:<:f>
<:I0,0,720,720><:f> "A very important concern is how primitive data types affect the size of user-defined, aggregate data types (records and structures). If basic integers and pointers suddenly double in size, structures may double as well. In some cases, this bloating
is intolerable..."<:F
@Footnote@<:f><:f> "Comparing the New 64-Bit RISCs" by Brian Case, <+#>Microprocessor Report<-#>, Vol. 7,
No. 3, March 8, 1993, p.15.
@Footnote@<:f>
>
<:f><:f>
<:#284,9360><:f>
<:#1136,9360><:f>A serious challenge of the larger 64-bit data and structure sizes is the increased requirements for memory and disk capacity. The choice of 32-bit data representation for use with PA-RISC provides the user with high-performance implementations which ar
e usable by today's applications and avoid the problem of "data explosion" with its consequences on system cost.
<:#284,9360><:f>
<:#284,9360><:f><+!>Performance: Impact on Cache<-!>
<:#852,9360><:f>PA-RISC also provides high-performance, cost-effective technology by delivering balanced system performance. Several areas of system<:f> performance should be considered for comparison to 64-bit computing, including cache performance and
<:f>the use of arithmetic operations.
<:#284,9360><:f>
<:#568,9360><:f>Dewar and Smosna, writing for "Open Systems Today," relate the issue of 64-bit data types to cache performance:<:f>
<:I0,0,720,720><:f> "The trouble with the move to 64 bits is that the data which would have previously required 4 bytes in the cache (addresses and integer data) end up taking 8 bytes, effectively reducing the size of the cache by a factor of two and resulting in more cac
he misses...<[>F]or many applications, the performance degradation from this effect will outweigh advantages from speedy 64-bit arithmetic."<:F
@Footnote@<:f><:f> See "Power Processing," p.53.
@Footnote@<:f>
>
<:f><:f>
<:#284,9360><:f>
<:#1420,9360><:f>This performance degradation due to cache misses is a significant cost to the user. An obvious solution to the cache miss performance problem might seem to be the doubling of cache size (with no corresponding performance advantage). The net result is
that a larger cache system is needed when using 64-bit computing technology than is needed when using a 32-bit system to achieve the same performance.
<:#284,9360><:f><:f><+!>Performance: Use of 64-bit Arithmetic<-!>
<:#2840,9360><:f>The user's application software dictates the type of arithmetic operations that are performed on data. The PA-RISC approach to arithmetic data operations, which is characterized by 64-bit floating point data registers and 32-bit integer registers, prov
ides high compute performance which is matched to the abilities of today's software applications. Current applications only rarely use 64-bit arithmetic for integer data. But in the "full" 64-bit implementations, the 64-bit arithmetic must be performed on
addresses, regardless of the type of arithmetic used for data. The 64-bit implementations must perform extra calculations (on 64-bit addresses) with no current benefit to the user in data arithmetic. Contrast the performance overhead in this approach wit
h the PA-RISC approach, which is usable by today's software applications for high-performance arithmetic operations.
<:#284,9360><:f><-#>
<:f>The cost-effectivenes<:f>s of 64-bit computing technology for system performance is a wide area for discussion. While benefits are possible, real inhibitors are also. Smith and Morrell note in their IDC report that "<[>w]hile 64-bit technology can pro
vide some performance boosts, it can also generate additional overhead..."<:F
@Footnote@<:f><:f> See "Sixty-Four Bit Computing," p.3.
<:#1420,9360><:f>PA-RISC provides a high-performance, cost-effective virtual memory system to meet users' address space needs. The question of address space centers around two issues: the amount of virtual memory address space available to the user and the performance
of the<:f> virtual memory system when accessing it. Smith and Morrell investigated the need for larger amounts of virtual memory in their IDC report and concluded that...<:f>
<:I0,0,720,720><:f> "...<[>t]he advantages of 64-bit addressing capabilities are not apparent when looking at the address space requirements of the average program. For the vast majority of these, 4GB of addressing <[>32-bit] will be sufficient for many years."<:F
@Footnote@<:f><:f> See "Sixty-Four Bit Computing," p.3-4.
@Footnote@<:f>
>
<:f><:f>
<:#284,9360><:f>
<:#1136,9360><:f>The PA-RISC architecture provides a virtual memory system which allows software applications to access 64-bits of address space through the use of segmented addressing. If "quantity" is the issue, the users of PA-RISC are not limited in the amount of v
irtual memory space they can address.
<:#284,9360><:f>
<:#1988,9360><:f>For the vast majority of users, the segmentation approach provides a cost-effective means of addressing 64-bits of virtual memory using object sizes up to 4 Gbytes (32-bit). Until RAM prices decrease significantly, the lack of physical memory in a syst
em will cause most users to encounter the system overhead due to memory "paging" long before any performance overhead due to segmentation is experienced. When physical memory (RAM) becomes more affordable in the future, flat addressing approaches to virtua
l memory system performance will become<:f> increasingly attractive.
<:#284,9360><:f>
<:#1420,9360><:f>HP's PA-RISC architecture provides customers with a proven virtual memory system which is capable of accessing large virtual address spaces. The PA-RISC approach, which utilizes 64-bit segmented addressing, will be enhanced by providing 64-bit flat add
ressing. Most importantly, this enhancement to PA-RISC architecture will maintain forward binary compatibility to protect the customer's hardware and software investments.
<:#284,9360><:f>
<:#284,9360><:f><:f><:f><:f><+!>CONCLUSION<-!>
<:#284,9360><:f>
<:#1420,9360><:f>HP's progressive implementations of PA-RISC demonstrate the scalability of the architecture. The question is not whether the 64-bit architectures are technically possible, the question is whether the current implementation of 64-bit features benefit th
e customer in today's environment. The bottom line for the customer was succinctly captured by META Group, Inc. in their 1993 report on 64-bit computing architectures:<:f>
<:#284,9360>
<:I0,0,720,720><:f> "Users should not be swayed by 64-bit hype, and should consider 64-bit systems only when all components (hardware, OS, DBMS, applications) can be assembled and exploited in an integrated environment."<:F
@Footnote@<:f><:f> "64-Bit Computing Architectures" by META GROUP, Inc., Open Computing & Server
Strategies Service, File No. 252, April 13, 1993.
>
<:f><:f>
<:#284,9360><:f>
<:#284,9360>
<:#1420,9360><:f>Hewlett-Packard's state-of-the-art PA-RISC will continue to evolve to meet the needs of the customer. The first enhanced PA-RISC 64-bit systems will appear at mid-decade, well before 64-bit computing becomes a broad market requirement. HP remains comm
itted to providing customers with high performance today and with migration paths to the future through the scalability of PA-RISC.
