April 24, 2000
"May you live in interesting times" was an ancient Chinese curse. By interesting times, they meant war, revolution, earthquake, famine, or similar class of events. While the year 2000 is not a disaster, for the computer industry it sure is interesting. And one thing that will be destroyed is a lot of preconceptions.
If you are a power user, almost everything you bought in 1999 will be technically obsolete sometime in 2000. While most of us are pretty well used to the rapid pace of advance in the computer industry, this year is the most dynamic I have seen in 35 years of working in this industry. Many of the hardware advances happening this year are related by the underlying technology, and some are coincidental, such as the wide range of new processors from several sources. I'll make some recommendations for power users and others at the end of part one.
The traditional expectation (Moore's law) in computing is that processor power doubles ever 18 months at the same price point. This year, it will double in less than 12 months. Memory speeds during the 1990s increased only incrementally while processors went from 25 MHz to 500 MHz. 100-MHz memory in 1988 was the first significant increase in that decade, but 1999 delivered 133 MHz, and this year we will see those numbers double.
In addition to the hardware revolution, there is a significant software evolution that we will see this year. Advances in operating systems, office applications, Web enabling software and the underlying tools will bring a wider range of software to the whole range of systems from mainframes to the not so lowly personal computer. I'll be treating the software evolution in Part Two of this series.
Hardware Revolution I use the term revolution in hardware because there are enough quantitative improvements in hardware this year to make a qualitative difference. For example, it will be easy by the end of this year to buy a system that can handle a high-speed Internet connection (DSL, cable) and display the full stream dynamically in 3-D without any special (expensive) hardware. Today, that takes a high-end configuration or a workstation with top-line components. By year end, it will be off the shelf at near today's prices for a standard system.
The year 2000 is unusual in that many new processors, memory systems, and peripherals will be arriving in a 12-month time frame. By the end of this year, new systems will represent a real quantum jump over systems available at the end of 1999. In a single year, processor frequencies will double or more, memory bandwidth will more than double, more cache will move onto the processor chip, and new and extended instruction sets, especially floating point and multimedia, will multiply these hardware performance increases.
The current turmoil we are seeing in the processor arena is the result of the semiconductor industry making several technology transitions at the same time. Semiconductor process technology is moving most CMOS processors to 0.18 micron geometry, chip design is moving to five and six layers of metal, some with copper interconnects rather than the more resistive aluminum. Low-capacitance insulation also enables faster clock speeds, and better cooling handles higher power, which also drives speed.
The memory bandwidth problem, which limits current generation chips, is being alleviated by 100 and 133-MHz SDRAM. These speeds will jump to 200 and 266 MHz this year, with plans for 400 MHz and higher speed memory in the pipeline. Level 1 cache internal to the CPU has reached 128K in the AMD Athalon, twice as much memory as the original IBM PC. Level 2 cache, originally chips on the motherboard, moved into the CPU package and are now being integrated onto the processor chip directly, running at full processor speed. Motherboard cache is being used as a third level of cache, now up to 2 Mbytes and larger.
Processors for servers and high-end systems are pushing the limits of technology to reach well into the gigaflop (billion floating point ops/sec) range per chip, and to multiple teraflop with a thousand or more processors connected as a single system. This level of performance places extreme demands on power supplies, cooling, circuit design, and physical construction. What is learned from this level will gradually appear in smaller systems in the next generations of chips.
Gigaflops for your PC? We are already there with 1-GHz Athalons with 3DNow and Pentium III SSE instruction sets, with PowerPC (Mac) systems G4+ soon to follow. As software support for these special instruction sets broadens, things like voice recognition and 3-D display will become standard features in all systems. The Gigaherz barrier, once a formidable limit, was broken in April by AMD's Athalon, and two days later, by Intel's Pentium III.
That barrier will be old history by the end of this year. Intel demonstrated a 1.5-GHz Willamette slated for the fourth quarter, a successor to Pentium III; and Samsung announced a 1.6-GHz Alpha for year's end (See EE Times for more on the 1.6-GHz Alpha and Ace's Hardware for the 1.5-GHz Willamette.)
Sixty-four bit processors such as Sun Ultrasparc, IBM Power, and Compaq Alpha, have been available for servers and high-end workstations for years, at a high price. Soon they will be available for power users as Intel's Itanium and Pentium IV (Willamette), and AMD's Thunderbird plus the 64 bit x86 Sledgehammer come closer to delivery. As high-end PC and Mac prices drop below $2000, fast Alpha systems are coming into range just below $3000.
Check out the chip columns on Byte (Mark Hachman or Alex Wolfe) for an in-depth look at some of these chips. Also look at these articles:
Server systems are the business world's workhorses, and the demands for more processing capacity are accelerating. Eight and 16-processor SMP systems are where the hottest action is in most business servers, but high-end systems from IBM and Sun have 32 and 64 processor systems available now. Sun's new Ultrasparc III is designed for systems with up to 1,024 processors. Intel's 64-bit Itanium and McKinley are designed to compete in this space.
A detailed discussion of server alternatives will be complex and lengthy, more than enough for a later column. The real megachips are from Hewlett-Packard and IBM. Not satisfied with a mere 20 million transistors on a tiny chip, these giants are putting more than 100 million on the HP 8700 and 170 million on the IBM Power4. The Power4 is designed with Chip MultiProcessing (CMP), which is more than one full processor on a chip.
The Power4 has two CPUs and interconnects for three other chips built in. The Power4 can build an eight-processor system in a hand-sized module. Also near the megachip range is Fujitsu's Sparc-V, a 65 million circuit design that is for their HAL (honest) range of very large systems. In addition, Compaq's Alpha will reach this level of complexity with the next generation EV7 chip, labeled 21364.
All of these chips use a lot of power, up to 125 watts for each Power4 chip. This power level makes power supply and cooling critical issues in the physical construction of the processor module. These chips represent what is currently possible with semiconductor technology if you push the limits. Don't expect one for your home system anytime soon.
Memory has been a performance bottleneck for years in the server market. A comment from the Alpha 21364 design team points this out. "Sites related in a database study using the TPC-C benchmark, the CPU (21164/400) was stalled waiting for main memory on three of every four cycles." (Microprocessor Report (MPR) 8/5/96 page 18). IBM also is attacking the bandwidth issue in the Power4 with a 10 Gbyte/sec main memory and 45 Gbyte/sec interprocessor bandwidth. (MPR 10/6/99) HP addressed the bandwidth problem starting with the PA 7200 design.
While those numbers are impressive, even the PC is moving up rapidly. From the 533 Mbyte/sec of the 66-MHz bus Pentiums, nothing changed until PC100 memory became available, boosting bandwidth to 800 Mbyte/sec. Now 133-MHz memory makes 1.066 Gbyte/sec bandwidth available, with 1.6 Gbyte/sec (200-MHz DDR) and 2.1 Gbyte/sec (266-MHz DDR) memory later this year. DDR is Double Data Rate memory, sending data on the leading and trailing edges of the memory clock.
Intel is planning a Quad Data Rate memory to reach 3.2 Gbyte/sec in 2001. RAMBUS, a different memory architecture, has the potential to be faster than SDRAM, but suffers from a more complex design and lower volumes, making it very expensive, between 6x and 11x current SDRAM costs. Higher-performance SDRAM designs may close this market before it gets established.
While bandwidth stagnated from 1995 until 1999, in one year we have more than doubled it, with another doubling to come next year. What drives this sudden jump in bandwidth? Processor performance needs. Processor performance, measured by a wide range of benchmarks, has become more and more limited by the inability of the memory system to keep the processor supplied with data and instructions.
Increasing the cache size helped for a while, but the limiting factor was bus speed since the L2 cache is limited by the front-side bus (FSB) in PC systems. It was relatively easy to boost FSB speeds to 100 and 133 MHz, but that is getting more difficult where motherboards must be designed for high volume, low-cost manufacturing. Double and Quad rate transfer will serve the coming generation of processors, but beyond that, more will be needed. It's a classic example of the Red Queen's race, where you have to run as hard as you can just to stay in place.
As processor and memory speeds increase, and peripherals as well, the current structure of PCI and AGP slots is running out of capacity. When AGP moved the video demands off of the PCI channels, it extended the PCI lifetime a little. While current PCs don't usually stress the PCI subsystem, the new chips and peripherals coming this year will.
Two competing proposals, Next Generation I/O (NGIO) and Future I/O (FIO), are now working on a common specification to solve this problem. This effort is now being led by an industry group named InfiniBand. PCI-X, an extension of the PCI specification originated by Intel, has reached version 1.0, with products announced at Comdex 1999. This is a mid-term extension of the PCI bus that will double the total PCI bus throughput from 532 Mbyte/sec to 1 Gbyte/sec. Most of the PCI-X bus and cards will be seen only in server systems where total throughput is critical as server systems get larger.
Disk drives and video cards have long been the primary load on the PCI bus. In servers, the disk drives were the only significant load until Gigabit Ethernet came along. Two Gigabit Ethernet channels running full duplex (125 Mbyte/sec in each direction) are enough to saturate a complete PCI bus. Large servers have implemented multiple PCI bus systems interfaced to fast memory, but this approach also has limits. PCI-X is a bridge solution until the combined FIO and NGIO teams come up with a longer term solution.
The InfiniBand specification is expected to be done by the fourth quarter 2000, with products 12 months later. New I/O capacity will be available none too soon. New disk drives from IBM and Seagate are already providing dramatic increases in speed and capacity. One of the key factors in disk-drive speed is how fast data is transferred from the disk to the drive's cache. This is not the channel speed, but an internal transfer rate that sets a limiting throughput.
The new IBM glass platters will transfer at 55 Mbyte/sec, the buffer at 66Mbyte/sec on an ATA/100 channel. Segate has concentrated on faster access with its X15 Cheetah, which has a 3.9 ms average access time, a huge jump over current 7+ ms drives. Both of these drives will strain the capacity of PCI when arrays of them are connected to a server. One IBM drive will saturate a PC IDE channel.
First consider your needs in view of the changes happening this year. If you can get along with what you have for six months or more, the best option is to wait for the new technology to reach retail shelves while you research the alternatives. If a current system is already too slow, consider one of the inexpensive upgrades listed below, or if an upgrade can't be done, look to the new AMD Athlon systems for the best buy and availability.
If you are a power user, my advice is to make inexpensive upgrades as discussed below until complete new systems (or motherboards) can be bought with DDR memory. Either the AMD Thunderbird, expected in Q2, or the Pentium IV in Q4 would be the top performing chip. Both will run faster than 1 GHz, so DDR or RAMBUS memory would be needed for best performance. Your current x86 choices are limited to buying a RAMBUS based system with a fast Pentium III or a 1GHz Athalon with 133-MHz SDRAM memory.
If your work is heavy numeric, especially floating point (FP), take a hard look at the Alpha systems because the Alpha 21264 chip has the best FP performance today, and that advantage will continue even when Intel's Itanium is released. The Alpha FP performance is two to six times that of x86 chips at the same clock rate. Both Linux and FreeBSD are available for the Alpha, as well as TrueUnix and NT 4.0. Be aware that Windows 2000 is not being ported to the Alpha.
The bad news is that until DDR memory is available with motherboards for it, that memory bottleneck will be present on every x86 system bought before then, except for expensive RAMBUS systems. I face the same dilemma as most: all my systems are Super 7 based, with PC100 memory at best. I will be making inexpensive upgrades to extend my systems until I can make a jump to Slot A with DDR memory.
In the shorter run, Socket A with Spitfire chips from AMD will be available soon (May) and less expensive than the Slot A/Thunderbird combination.
Most regular PC users will not be limited by memory bandwidth until a later generation, so they need not wait for DDR. People with the older socket 7 and Super 7 motherboards can either make inexpensive upgrades of processor or processor and motherboard while retaining current peripherals, or choose Socket 370 or Socket A for a new motherboard or system. Be aware that new Celeron and Pentium III chips will not run in older motherboards, even if the socket or slot seems correct.
If your motherboard is a year or more old, it probably will not support the new chips. Check with the supplier for BIOS upgrades, but you may have to replace it. I haven't given any detail on video cards because that only affects display performance, important to users but not an architecture issue. The current high end cards come with 32 or 64 Mbytes of very fast memory, and in some cases, two processor chips. These high end cards (see link) evolve much faster than the PC itself. Unless you need the very latest for gaming or serious graphics work, cards from a generation or two back, about a year old, are plenty fast enough.
Here are some examples drawn from my motley collection of systems. My current main workstation, a K6-2/450 runs mostly well enough with an older Matrox G100 card, three or more generations old. I will replace that VGA card with a near current PCI or AGP card for a doubling of video performance. Unless 3-D games are your main use, AGP cards have little performance advantage, but they do free up a PCI slot. Another inexpensive upgrade is adding memory, up to 128 Mbytes for most systems.
For my older systems like the K6-200, a new motherboard such as the FIC VA-503+ is available under $90, with K6-2/500 processors around $50 from a number of vendors. This MB/processor combination can use either SIMM or DIMM memory and controllers from your current system, making it a cheap step up. About 128 Mbytes of PC100 RAM is currently under $100, so that big jump to 500 MHz will cost around $250 with new memory.
Vendors I have used are listed below. The fastest processor for current Super 7 motherboards is the AMD K6-2/550, and there is not likely to be a faster successor. A note on overclocking, which is becoming popular among gamers. Overclocking is possible, but it stresses the CPU and generates a lot more heat. Processors with cheap fans and power supplies will not work well, if at all, and there is a real risk of burning out the CPU or damaging the motherboard. If your system is used for work of any sort, this is a bad idea. It's cheap just to upgrade and much safer. If it is just a fun machine, check out the overclock sites and experiment. Be prepared to replace equipment.
Every year, one or two technologies make big-capacity or performance jumps. This year, every major component of computers will see big jumps in capabilities. Suddenly the current architecture is at its limits, where last year it looked like there was plenty of room to grow. One example that will impact a lot of people is the Super 7 motherboard system. It has been extended well beyond the original design for Pentiums, and has reached the limits for new processors without a major redesign.
New PC systems based on Socket 370 and Slot 1 from Intel have been around for a few years, and AMD has the Slot A Athlon, with Socket A coming in May. The slot approaches only scale up to two processors without a lot of additional supporting circuitry. Memory access and cache coherence for SMP servers demand much higher performance, which raises motherboard prices quite a bit, and extends engineering times.
Today, Slot A and Socket A have more room for growth than the older Intel designs. One particular advantage is independence of bus and memory speeds in this design. Current Slot A systems run the bus at 200 MHz, with either 100 or 133-MHz memories. Later this year, a 266-MHz Slot A will accept DDR memory as well. Scaling up the Slot A speed should not be a difficult technical problem, but integrating it into the flow of motherboards and processor production lines will be tricky.
The fastest current systems use RAMBUS memory, at a significant cost premium. Later this year, DDR memory will tilt the advantage the other way, and costs should be much closer to current SDRAM levels. Longer term, it is impossible to predict which approach will provide better overall performance, but the cost issue is much clearer. The DDR and QDR SDRAM approaches will be less expensive. We do live in interesting times, but for once, it's all to our advantage.
All content on this site is Copyright 2001 by Bill Nicholls