64-Bit Question

It's unlikely you spend much time pondering the number of bits a processor can handle. However, recent developments could make it an issue affecting everyone. The 64-bit question has implications for all of computing, but will the impact be as great as 32-bit? Of course, 64-bit processors are nothing new - they've been around in RISC (Reduced Instruction Set Computer) servers and workstations for years, and more recently in games consoles.

64-Bit Question
It's unlikely you spend much time pondering the number of bits a processor can handle. However, recent developments could make it an issue affecting everyone. The 64-bit question has implications for all of computing, but will the impact be as great as 32-bit?

Of course, 64-bit processors are nothing new - they've been around in RISC (Reduced Instruction Set Computer) servers and workstations for years, and more recently in games consoles. However, the landscape is changing, chiefly through the involvement of mass-market chipmaker Intel and its rival AMD, both of which announced new 64-bit chips in 2002. If history is anything to go by, this would imply an impending commoditisation of the technology, leading not just to more affordable 64-bit servers, but also 64-bit desktops, notebooks and PDAs over the next few years.

But is that really on the cards? Do we need 64-bit everywhere anyway? There are also software implications to consider - are 64-bit applications any better than 32-bit, do they run any faster, and can you run existing 32-bit programs on the chips now starting to appear? Moreover, is any 64-bit software actually available for these processors?

Questions like these need to be answered to get beyond the hype and commercial forces currently driving 64-bit technology. In this feature, I'll be looking at the latest 64-bit processor developments and putting them into perspective for those of us living in the real world. As such, a good place to start is with a short layman's guide to what 64-bit computing is all about and what sort of applications it's good for. Then I'll look in more depth at what the new chips from Intel and AMD have to offer, along with the software implications, and where the major system vendors look set to take 64-bit over the next few years.

More bang for your bits

Potentially, there are many things 64-bit processors should do better than their 32-bit counterparts. That said, most advances hinge on the ability to work with larger numbers, which, not surprisingly, can be up to 64-bits long. This allows, among other advantages, for a higher level of accuracy when it comes to big floating-point calculations. But this would seem to be of little value outside the realms of scientific apps - such as mapping of the human genome - and unimportant for corporate transaction processing or servicing web and email servers.

Of wider importance is the ability to address a larger amount of memory. The amount varies depending on the implementation and operating system, but Windows' limit for a 32-bit processor is currently 4GB. With a 64-bit processor, however, you can, in theory, directly address up to 8TB - 2,000 times as much.

While 4GB of memory should be more than enough for desktop applications, in a server that amount is soon used up. The extra space possible with 64-bit addressing would enable much larger SQL databases to be loaded and run in memory, and whole websites to be resident in RAM. As RAM is so much faster than disk, the potential benefits are easy to grasp - programs will run faster, more users will be supported on the same hardware, more complex transactions accommodated and so on.

Properly exploited 64-bit addressing may benefit scientific and technical computing as well as the big transaction-processing systems deployed by large enterprises. As a consequence of design, other features of 64-bit processors can fulfil the larger transaction requirements of enterprises right down to those deployed in smaller companies and on the desktop.

The speed at which the processor is able to access all that extra memory is of critical concern. It's no good being able to access memory in 64-bit chunks if the interface between processor and RAM slices it up into smaller units for transfer. Indeed, Intel has had to come up with some major improvements on its original 64-bit Itanium - the new Itanium 2 provides a threefold increase in bus bandwidth (see Itanium 2 - this time it's serious).

Other manufacturers claim to still outperform the Itanium in this area. Rival AMD, for example, boasts much better throughput for its 64-bit Hammer architecture using a bus technology called HyperTransport, which it helped develop. It also claims to support easier and more effective scaling in multiprocessing systems, which we'll discuss in more detail, along with actual processor implementations of Hammer technology, later on.

In the meantime, there's also the matter of compatibility. It may not be a problem for some of the older 64-bit RISC designs, such as Compaq's (now HP's) Alpha processor or Sun's UltraSPARC, which run custom Unix software. But it certainly is for Intel, which, for the first time, has departed from the well-established x86 instruction set of its earlier 16- and 32-bit processors in favour of a totally new approach, dubbed IA-64.

Based on a technology it calls EPIC (Explicitly Parallel Instruction Computing), the new Intel architecture is claimed to be good for both high-end technical computing and commercial applications. That's not just because it's 64-bit, but because the design of the chip enables the Intel processor to execute a lot more instructions at the same time - that is, in parallel. An Itanium 2 processor can process up to six instructions simultaneously, with more resources to carry out those instructions than most 32-bit chips and some existing 64-bit designs.

The Itanium 2 has 328 on-board registers plus six integer and two floating-point execution units that can all be run in parallel. Added to this, the processor has a lot of logic dedicated to deciding in advance which instructions can be executed in parallel along with facilities to execute those instructions and transform data ahead of time. Chip and compiler designers refer to this as predictive or speculative processing.

This is good stuff, and the Itanium 2 with its 221 million transistors is an impressive 64-bit CISC (Complex Instruction Set Computer) processor, although it has an Achilles heel - the inability to run existing x86 applications natively.

IA-64 is completely different from x86, aka IA-32. Although it's possible to provide a degree of compatibility by mapping x86 instructions and registers to a subset of the new resources, that's far from enough. Instead, the developers have had to build in a special x86 emulation mode, which has to be 'booted' much like a standalone x86 processor to initialise items such as memory descriptors, registers and flags.

Switching into this mode is a far from trivial process, and potential performance isn't promising. Tests with the first Itanium have put it on a par with a Pentium II when running 32-bit code. So although it's possible to run Windows or DOS programs unchanged - even whole operating systems - these won't run as fast as if they were ported to the newer architecture. Moreover, the Itanium clock speeds are relatively slow compared with the latest Intel Xeon and Pentium 4 chips (1GHz for the first Itanium 2), which could lead many companies to upgrade to the latest 32-bit silicon rather than jump to the Itanium.

Software compatibility

To get the full benefit of Intel's 64-bit processor, you need both a 64-bit operating system and applications written and compiled for 64-bit. As well as taking time to achieve (see 64-bit software), this is a big upheaval in a market used to getting faster applications simply by running them, unaltered, on newer hardware.

But the Intel argument for switching to its new architecture leans heavily on how it sees the 64-bit processor being used - a vision not universally shared by its rivals.

Study the Itanium roadmap or read any of the associated white papers and you'll find few references to anything other than servers and technical workstations. On such systems, Intel reasons that the performance benefits of running enhanced 64-bit software, tuned specifically for its IA-64 architecture, outweigh the drawbacks of having to re-write code. It also reasons that sticking with the old x86 architecture would limit further development, as the x86 is unable to take advantage of the parallel processing and other features of the new chip.

That's not how AMD sees it. AMD engineers have based the 64-bit Hammer architecture on an x86 core that can run existing 32-bit applications as quickly, if not faster, than 32-bit chips. Like the existing 32-bit Athlon, Hammer chips feature a RISC processor inside a CISC shell, this time with 64-bit extensions. The x86 instructions are converted into smaller, easier-to-execute RISC operations, enabling the AMD processor to provide a high level of performance with fewer resources than the Itanium. It does have to switch mode when going from 32-bit to 64-bit processing, and vice versa. Whereas with the Itanium that's like booting a completely separate processor, it's no harder for Hammer than for a 32-bit x86 chip, such as a 486 or Pentium, switching into 16-bit operation. Also, because the architecture is x86 based, there should be no degradation in performance when running 32-bit code on the 64-bit AMD chips.

It comes as little surprise, therefore, to find that AMD expects compatibility to be of real importance in the cautious enterprise market. Such customers, it says, make changes gradually, and then only where there's a clear competitive advantage to be gained. AMD hopes the ability to run existing applications faster, with the promise of further software gains later, will be more attractive than having to change both hardware and software horses mid-stream.

The big downside to simply extending the old x86 architecture (the new architecture is unimaginatively referred to as x86-64) is that it isn't so easy to match the kind of parallel and predictive processing possible with IA-64. On the plus side, the first Hammer processors do feature a high degree of parallelism, being able to process up to nine RISC operations (ROPS) at a time. But it can take three or more ROPS to make one CISC instruction, and the AMD silicon also has fewer registers and execution units than the Itanium. This could limit the ability of AMD to compete both in the technical workstation and commercial server markets. However, Hammer does have a technological ace up its sleeve when it comes to multiprocessing.

Intel's approach is the same with the Itanium as with earlier 32-bit processors. Each processor connects to a shared system bus, managed by a dedicated MP chipset. This limits the speed at which processors can access both shared memory and the cache in other chips. With Hammer, both the memory controller and the crossbar switch provided by the MP chipset are built into each AMD processor. Each has its own dedicated memory, plus separate high-bandwidth (HyperTransport) links to the cache and memory associated with the other processors.

Dubbed 'glueless multiprocessing', this approach does away with the need for complex multiprocessing chipsets, making life easier for motherboard manufacturers. Also, it makes for better scalability, with the HyperTransport bus of the Hammer architecture providing greater bandwidth and lower latency than is possible using the shared bus of the Itanium. This should enable AMD to at least get a toehold in the SMP market, with plans to introduce four-way and eight-way servers where previously it was limited to just two-way with the Athlon chip.

But, just as importantly, AMD doesn't see 64-bit computing limited solely to high-end servers and workstations. Two implementations of Hammer are due to be released during 2003. Only one - the Opteron - is for multiprocessing servers and workstations, with three HyperTransport buses per processor to enable scaling up to eight-way.

The other is a 64-bit Athlon with just two such buses, which will be used in ordinary desktop and notebook PCs. The backwards compatibility of Hammer could really score here, giving AMD an edge over Intel, which is sticking with 32-bit processors in this market for the foreseeable future.

Infinity and beyond

'That's all very interesting,' you might say, 'but where are these new 64-bit processors taking us and when will we be able to get hold of them?' The answer to this depends on what type of system you're after.

If it's a server platform, you can get 64-bit solutions already. When it comes to Unix-based mid-range servers, the market is mature and you're almost spoilt for choice, although that's set to change. Compaq, for example, announced two years ago that it would be phasing out the 64-bit Alpha chip (inherited from Digital) in favour of Intel's Itanium. Post merger, the new HP will be continuing that programme and, likewise, replacing its own 64-bit PA-RISC chip with the Intel silicon.

The good news for existing customers is that this will be a gradual process. Itanium-based processing modules are being produced that can be plugged in without having to change other components. Nor will it be necessary for major software changes, especially on PA-RISC servers, as HP co-developed the Itanium and made sure native support for its instruction set was built in.

Elsewhere, Sun continues to plough its own hardware furrow, deploying its 64-bit UltraSPARC processor in servers and high-end workstations. And IBM is continuing to develop its 64-bit Power processor for use in servers running AIX, while at the same time also deploying the Itanium in its x-series line. Plans to port AIX to the Intel 64-bit chip have been shelved - IBM now favours Linux and Windows for this platform.

Lower down the scale, the leading vendors of industry-standard (Intel-based) servers have all announced products based on the Itanium 2, the first 'production' version of the IA-64 chip. All will be multiprocessing systems, although initial shipments will only be two-way and four-way, as the development of suitable chipsets has lagged behind the introduction of the processors. Shipping dates will also be affected by software availability (see 64-bit software), particularly .NET Server 2003, and it's likely to take another year or two for the market to start moving.

As for AMD, at the time of writing none of the big-name vendors had announced servers based on the Opteron processor, which is unlikely to ship in volume before mid-2003. There are lots of rumours, however, and the x86 compatibility features of the Opteron could well see a number of smaller vendors adopting the processor over the next few months. Unfortunately, the lack of software support means these will have to compete against 32-bit Xeon MP systems to begin with, as well as the Intel marketing muscle in this market.

The dynamics are quite different in the desktop sector, though, where AMD processors are seen as a viable alternative to Intel chips. You might reasonably expect 64-bit desktop and notebook PCs, based on Athlon chips, well before the end of 2003. However, users of such systems will, again, have to be content with 32-bit applications for some time.

It's much like when the industry switched from 16-bit to 32-bit computing (see Echoes of the past). However, this time round it's likely to take a lot longer before we start putting the word 'legacy' in front of '32-bit applications' and moving universally into the new 64-bit world.

Author: Alan Stevens
The 64-bit question



Local Articles
Computer Hardware
Home