Systems built for speed

Two years ago, only a few vendors produced Alpha systems, and consumers had a limited choice of packaging, features, and price. Now, finding an Alpha system is easy. Many clone manufacturers, integrators, and system developers build Alpha machines--­from longtime Alpha advocates such as Aspen Systems and Carrera Computers to market newcomers such as CompuSys and Advanced Business Technology. Alpha chips are available from two foundries, and two more Alpha foundries might begin producing chips soon. The market's number of resellers is growing, and Alpha systems' prices range from $99 motherboards to $100,000 multiprocessor behemoths. Alpha CPUs span three generations, and available Alpha systems range in age from recently announced to more than 5 years old. Because of the variety of Alphas now available, consumers frequently wonder how to determine a particular Alpha machine's capabilities in relation to other computers. A look at the past and future of Alpha chips will help you understand what is inside different systems and what to expect for the price you pay.

Start at the Bottom
You can find entry-level Alpha systems complete with RAM, a hard disk, graphics, sound, and network cards for less than $500 in clearance stores. I've even seen new motherboards for less than $100 on the Internet. Digital Equipment calls the chips in most systems at this end of the cost spectrum low-cost Alpha (LCA) processors. These processors have the designations DECchip 21066, 21068, and 21066a. Digital designed them primarily for embedded applications for realtime devices, controllers, and laptops. LCA chips include a cache controller, PCI interface, and rudimentary graphics controller. To achieve this level of integration, LCA processors leave off some components that are standard on most chips. The chips limit the external data bus to 64 bits, use only off-chip cache, offer a maximum of 1MB of cache, and have only 34 external address lines. These limitations reduce LCA chips' complexity, power consumption, cost, and performance. These processors are not the barn burner number crunchers most people expect when they think about Alpha machines.

Digital no longer manufactures LCA chips, and Windows 2000 (Win2K--­formerly Windows NT 5.0) doesn't support them. LCA chips came in speeds of 66MHz (the 21068, or LCA4S), 166MHz (the 21066, or LCA4), and 233MHz (the 21066a, or LCA45). The 21068 chips are rare. They serve mostly as embedded controllers for realtime applications. The 21066 and 21066a processors received wider distribution. Digital shipped them as part of two motherboard packages: the AXPpci33 motherboard (which became known as the Noname board because Digital never gave it a code name), and the Multia Universal Desktop Box (also known as the Multia or the UDB). When they first came out in 1994, LCA systems were more expensive than most PCs, primarily because of the large quantity of software tools that came with the systems. Now that their prices are low, these units have become popular among hobbyists.

The Original Consumer Alpha
Digital introduced the 21064 microprocessor in a series of systems that don't run NT. DEC 3000, 4000, and 7000 AXP servers and workstations, which Digital based on the TURBOchannel or Futurebus for I/O support, can't run NT. The later AlphaStation and AlphaServer systems, which Digital based on the PCI bus, were the first Alpha systems Digital designed with NT in mind.

The 21064 has a dual-issue architecture with an external bus as wide as 128 bits. The 21064's data and instruction Level 1 caches are 8KB each. The 21064a has as much as 16KB each of Level 1 data and instruction cache. The 21064 came in frequencies ranging from 100MHz to 300MHz, and it was the first chip in The Guinness Book of World Records. In 1992, the chip set a record as the fastest microprocessor available. The 21064 is no longer the fastest CPU, but it still performs respectably; its floating-point performance rivals that of some current x86 platforms.

The 21064 was available on a variety of motherboards and in a variety of systems. Digital's 21064 systems include the Alpha XL minitower PC; the AlphaStation 200, 250, 255, and 400 workstations; and the AlphaServer 1000, 2000, and 2100 servers. All these machines run NT and other OSs, including OpenVMS, Linux, and Digital UNIX. Digital manufactured the EB64 and PC64 motherboards for OEMs to use to build 21064 systems. OEMs used the EB64 primarily for design planning and the PC64 for production of Alpha system clones. PC64 motherboards support processors ranging in speed from 233MHz to 300MHz, cache ranging in size from 512KB to 2MB, and as much as 512MB of 128-bit RAM. The PC64 uses a baby AT form factor and provides three dedicated PCI slots, two ISA slots, and one shared PCI/ISA slot for I/O.

The oddest NT-capable 21064 system is the DECpc AXP150 (also known as the DEC 2000 Model 300 or 500 and by its code name, Jensen). When Digital began shipping this 100-percent EISA box in 1993, it was the first Alpha system to run NT, and it turned out to be the only non-PCI Alpha system to run NT. The DECpc AXP150 CPU runs at 150MHz. Digital had to slow the processor from its original design specification of 151.5MHz to meet supercomputer export restrictions of the time. The system was as fast as modern 90MHz-to-100MHz Pentium machines at a time when Intel was just announcing the 60MHz and 66MHz Pentium chips. The DECpc AXP150 has eight SIMM slots for as much as 128MB of RAM interleaved in two banks of four SIMMs each. The system offers few expansion options because it lacks a PCI bus and the processor is not upgradable. The DECpc AXP150's NT hardware abstraction layer (HAL) is too complex to warrant updating this Pentium-class performer to Win2K.

Today's Models
In 1994, Digital introduced the Alpha 21164 processor. The first 21164 ran at 250MHz; the chip now runs as fast as 700MHz and can reach 767MHz when it runs in conjunction with KryoTech's phase-change processor-cooling system. All systems that use this chip run NT. The 21164 is a quad-issue, superscalar processor with 8KB each of Level 1 data and instruction cache and 96KB of Level 2 cache, all on chip. The 21164 has a whopping 10 million transistors. Digital has referred to this chip as the BIPS chip because it was the first CPU to break the billion instructions per second (BIPS) threshold.

21164 systems vary widely in their amounts of Level 3 cache. The chips support as much as 64MB of cache, but most systems use between 1MB and 4MB. Most high-end systems use 8MB. These cache sizes might seem like overkill from a CISC perspective, but they offer RISC processors a performance improvement of up to 25 percent.

The 21164 processor received a midlife power boost in its EV56 implementation, which supports processor speeds above 400MHz. (For information about how Alphas got their EV names, see the sidebar "Electro What?") The EV56 was the first chip to change the Alpha instruction set. RISC processors need 64-bit data alignment for optimal performance. Early Alpha chips waste clock cycles trying to compensate for misaligned data from x86 programs. The EV56 adds byte- and word-alignment instructions to speed up translations of x86 operation codes. This architectural change showcases the Alpha's ability to include new functions without rendering previous implementations obsolete. Digital produced a simple software upgrade that lets older Alpha processors use the byte-alignment instructions through optimized code that emulates the functionality of the new instruction set. Similar software lets the Alpha architecture correct chip, board, or system design errors. Design flaws such as the infamous Pentium floating point divide (FDIV) bug are not serious problems for the Alpha, because you can solve them through firmware or HAL upgrades instead of chip swaps or OS modifications.

The 21164PC is a low-end, low-cost variant of the 21164. The 21164PC does not include the 21164's on-chip Level 2 cache. The 21164PC has a 16KB Level 1 instruction cache and an 8KB Level 1 data cache. Its off-chip Level 2 cache ranges from 512KB to 4MB. This chip appears in many low-cost Alpha systems with speeds that range from 533MHz to 600MHz. You can also find it in 400MHz and 466MHz systems, but those implementations are rare. The 21164PC's lack of on-chip Level 2 cache makes its systems about two-thirds as fast as 21164 systems. The reduced-price chip has only 3.4 million transistors, which reduces its die size and increases the number of chips per wafer. The 21164 performs integer operations at speeds that rival high-end Intel chips and performs floating-point operations at about twice contemporary Intel systems' speeds.

The 21164PC introduced an enhancement to the Alpha instruction set: motion video instructions (MVI), which make the chip exceptionally versatile at multimedia applications. The MVI instruction set accelerates the chip's performance of standards-based multimedia, particularly Digital Versatile Disc (DVD), Moving Pictures Expert Group (MPEG), and conferencing data streams. The instructions fall into three categories: pixel error, pack/unpack, and data clamping. Unlike Intel's MMX, which increases cache size and further divides an already-small register to improve general application performance, MVI delivers realtime MPEG software encoding, which saves users substantially on video-capture hardware. Microsoft's NetShow Services use MVI to improve MPEG compression of live video streams.

Motherboards for the 21164 include Compaq's EB164, PC164, PC164LX, PC164SX, and AlphaPC 164RX. The first four motherboards offer two 64-bit PCI slots, two 32-bit PCI slots, and two ISA slots. The PC164 lets you configure eight banks of memory in one 256-bit memory path with CPUs as fast as 600MHz. The PC164LX supports processors as fast as 800MHz. The PC164SX holds 21164PC processors that range in speed from 400MHz to 533MHz. The AlphaPC 164RX holds 600MHz 21164PC processors and uses a 75MHz bus, which improves performance over the PC164SX by more than 50 percent. Samsung manufactures the MB164LX motherboard, which is similar to Compaq's PC164LX, and the MB164UX motherboard. Samsung's 21164 motherboards support clock speeds as fast as 800MHz, as much as 3GB of RAM, 4MB of Level 3 cache, six PCI slots (including one 64-bit PCI slot), and built-in Ultra Wide SCSI and Ethernet capabilities. Samsung also produces an MB164BX motherboard, which is similar to the MB164UX except that it doesn't include Ultra Wide SCSI and Ethernet interfaces. Other vendors, including DeskStation Technology and DCG Computers, also offer motherboards for 21164 processors.

The 21164 microprocessor comes in a variety of systems. Compaq's AlphaServer series (including AlphaServer 800 and the very large memory--­VLM--­capable 4000 and 8000 families), AlphaStation series, and Personal Workstation family come with 21164 processors. (For more information about the 21164 processor, see Joel Sloss' sidebars "The Alpha 21164," March 1996, and "New Alpha CPU Raises the Bar," October 1996.)

Now Arriving
In February 1998, Digital and Samsung announced the first implementations of the 21264 microprocessor, which Compaq is now shipping. (For a detailed description of the 21264's characteristics and functionality, see Tom R. Halfhill, "The Alpha 21264," December 1998.)

Compaq plans to produce a system called the Compaq 1000 XP (its code name is Monet), which offers a 500MHz 21264 that users can upgrade to faster 21264 chips, in the first quarter of 1999. I expect the system to have four PCI slots and one PCI/ISA slot, 4MB of external Level 2 cache, and as much as 4GB of Error Correcting Code (ECC) Synchronous DRAM (SDRAM). The workstation's I/O facilities include built-in Universal Serial Bus (USB), 10/100Base-T Ethernet, audio, and Ultra Wide SCSI components. Compaq is also working on Alpha-based ProLiant servers, although no details are available at press time.

System configurations for the 21264 will eventually include one-, two-, and four-processor systems. The support logic on the chip's motherboards supports a wide range of memory, cache, and I/O options. Compaq recently announced the AlphaPC 264DP, a high-end motherboard with room for two 500 MHz 21264 processors (each with up to 4MB of 128-bit Level 2 Cache) and 4GB of RAM. Compaq's 21264 servers will include upgrades of existing systems such as the 4000 and 8000 servers and new technologies such as the Adaptive Partitioned Multi-Processor (APMP) series that has the code name Wildfire. Compaq has slated Wildfire's release for 1999. The system has a modular, building-block design, and it will run as many as sixty-four 21264 processors.

Looking Ahead
Now that Intel and Samsung manufacture Alpha processors, the timeline for production of 21264 systems has accelerated. The 21264 will appear in three distinct technologies: 0.35, 0.25, and 0.18 microns. Performance increases and cost reductions will accompany each new implementation.

Alpha's next generations, 21364 and 21464 chips, are already on Compaq's drawing boards. Plans beyond the 21464 are as open as the Alpha architecture. Future Alpha chips might include copper technologies, fully depleted Silicon On Insulator (SOI) multiple CPUs per chip, and a technology similar to Explicitly Parallel Instruction Computing (EPIC). Compaq and Samsung plan to keep Alpha machines at least twice as fast as and less expensive than their IA64 counterparts. (For a performance comparison of several generations of Alpha and Intel processors, see the sidebar "The Performance Curve.") Alpha processors will be viable into the next century. They will offer benefits for users who need top performance or want to migrate to a mature 64-bit platform today.