История создания суперкомпьютеров Cray: от CDC к собственному бизнесу.

Supercomputer manufactured by Cray Research

From 1968 to 1972, Seymour Cray of Control Data Corporation (CDC) worked on the CDC 8600, the successor to his earlier CDC 6600 and CDC 7600 designs. The 8600 was essentially made up of four 7600s in a box with an additional special mode that allowed them to operate lock step in a SIMD fashion. Jim Thornton, formerly Cray's engineering partner on earlier designs, had started a more radical project known as the CDC STAR 100. Unlike the 8600's brute force approach to performance, the STAR took an entirely different route. The main processor of the STAR had lower performance than the 7600, but added hardware and instructions to speed up particularly common supercomputer tasks. By 1972, the 8600 had reached a dead end; the machine was so incredibly complex that it was impossible to get one working properly. Even a single faulty component would render the machine non operational. Cray went to William Norris, Control Data's CEO, saying that a redesign from scratch was needed. At the time, the company was in serious financial trouble, and with the STAR in the pipeline as well, Norris could not invest the money. As a result, Cray left CDC and started Cray Research very close to the CDC lab. In the back yard of the land he purchased in Chippewa Falls, Cray and a group of former CDC employees started looking for ideas. At first, the concept of building another supercomputer seemed impossible, but after Cray Research's Chief Technology Officer travelled to Wall Street and found a lineup of investors willing to back Cray, all that was needed was a design. For four years Cray Research designed its first computer. In 1975 the 80 MHz Cray 1 was announced. The excitement was so high that a bidding war for the first machine broke out between Lawrence Livermore National Laboratory and Los Alamos National Laboratory, the latter eventually winning and receiving serial number 001 in 1976 for a six month trial. The National Center for Atmospheric Research (NCAR) was the first official customer of Cray Research in 1977, paying US$8.86 million ($7.9 million plus $1 million for the disks) for serial number 3. The NCAR machine was decommissioned in 1989. The company expected to sell perhaps a dozen of the machines, and set the selling price accordingly, but ultimately over 80 Cray 1s of all types were sold, priced from $5M to $8M. The machine made Seymour Cray a celebrity and his company a success, lasting until the supercomputer crash in the early 1990s. Based on a recommendation by William Perry's study, the NSA purchased a Cray 1 for theoretical research in cryptanalysis. According to Budiansky, "Though standard histories of Cray Research would persist for decades in stating that the company's first customer was Los Alamos National Laboratory, in fact it was NSA "

The 160 MFLOPS Cray 1 was succeeded in 1982 by the 800 MFLOPS Cray X MP, the first Cray multi processing computer. In 1985, the very advanced Cray 2, capable of 1.9 GFLOPS peak performance, succeeded the first two models but met a somewhat limited commercial success because of certain problems at producing sustained performance in real world applications. A more conservatively designed evolutionary successor of the Cray 1 and X MP models was therefore made by the name Cray Y MP and launched in 1988. By comparison, the processor in a typical 2013 smart device, such as a Google Nexus 10 or HTC One, performs at roughly 1 GFLOPS, while the A13 processor in a 2019 iPhone 11 performs at 154.9 GFLOPS, a mark supercomputers succeeding the Cray 1 would not reach until 1994.

Typical scientific workloads consist of reading in large data sets, transforming them in some way and then writing them back out again. Normally the transformations being applied are identical across all of the data points in the set. For instance, the program might add 5 to every number in a set of a million numbers. In simple computers the program would loop over all million numbers, adding five, thereby executing a million instructions saying a = add b, c. Internally the computer solves this instruction in several steps. First it reads the instruction from memory and decodes it, then it collects any additional information it needs, in this case the numbers b and c, and then finally runs the operation and stores the results. The end result is that the computer requires tens or hundreds of millions of cycles to carry out these operations.

In the STAR, new instructions essentially wrote the loops for the user. The user told the machine where in memory the list of numbers was stored, then fed in a single instruction a(1 1000000) = addv b(1 1000000), c(1 1000000). At first glance it appears the savings are limited; in this case the machine fetches and decodes only a single instruction instead of 1,000,000, thereby saving 1,000,000 fetches and decodes, perhaps one fourth of the overall time. The real savings are not so obvious. Internally, the CPU of the computer is built up from a number of separate parts dedicated to a single task, for instance, adding a number, or fetching from memory. Normally, as the instruction flows through the machine, only one part is active at any given time. This means that each sequential step of the entire process must complete before a result can be saved. The addition of an instruction pipeline changes this. In such machines the CPU will "look ahead" and begin fetching succeeding instructions while the current instruction is still being processed. In this assembly line fashion any one instruction still requires as long to complete, but as soon as it finishes executing, the next instruction is right behind it, with most of the steps required for its execution already completed. Vector processors use this technique with one additional trick. Because the data layout is in a known format — a set of numbers arranged sequentially in memory — the pipelines can be tuned to improve the performance of fetches. On the receipt of a vector instruction, special hardware sets up the memory access for the arrays and stuffs the data into the processor as fast as possible. CDC's approach in the STAR used what is today known as a memory memory architecture. This referred to the way the machine gathered data. It set up its pipeline to read from and write to memory directly. This allowed the STAR to use vectors of length not limited by the length of registers, making it highly flexible. Unfortunately, the pipeline had to be very long in order to allow it to have enough instructions in flight to make up for the slow memory. That meant the machine incurred a high cost when switching from processing vectors to performing operations on non vector operands. Additionally, the low scalar performance of the machine meant that after the switch had taken place and the machine was running scalar instructions, the performance was quite poor. The result was rather disappointing real world performance, something that could, perhaps, have been forecast by Amdahl's law.

Cray studied the failure of the STAR and learned from it. He decided that in addition to fast vector processing, his design would also require excellent all around scalar performance. That way when the machine switched modes, it would still provide superior performance. Additionally he noticed that the workloads could be dramatically improved in most cases through the use of registers. Just as earlier machines had ignored the fact that most operations were being applied to many data points, the STAR ignored the fact that those same data points would be repeatedly operated on. Whereas the STAR would read and process the same memory five times to apply five vector operations on a set of data, it would be much faster to read the data into the CPU's registers once, and then apply the five operations. However, there were limitations with this approach. Registers were significantly more expensive in terms of circuitry, so only a limited number could be provided. This implied that Cray's design would have less flexibility in terms of vector sizes. Instead of reading any sized vector several times as in the STAR, the Cray 1 would have to read only a portion of the vector at a time, but it could then run several operations on that data prior to writing the results back to memory. Given typical workloads, Cray felt that the small cost incurred by being required to break large sequential memory accesses into segments was a cost well worth paying. Since the typical vector operation would involve loading a small set of data into the vector registers and then running several operations on it, the vector system of the new design had its own separate pipeline. For instance, the multiplication and addition units were implemented as separate hardware, so the results of one could be internally pipelined into the next, the instruction decode having already been handled in the machine's main pipeline. Cray referred to this concept as chaining, as it allowed programmers to "chain together" several instructions and extract higher performance.

In 1978, a team from the Argonne National Laboratory tested a variety of typical workloads on a Cray 1 as part of a proposal to purchase one for their use, replacing their IBM 370/195. They also planned on testing on the CDC STAR 100 and Burroughs Scientific Computer, but such tests, if they were performed, were not published. The tests were run on the Cray 1 at the National Center for Atmospheric Research (NCAR) in Boulder, Colorado. The only other Cray available at the time was the one at Los Alamos, but accessing this machine required Q clearance. The tests were reported in two ways. The first was a minimum conversion needed to get the program running without errors, but making no attempt to take advantage of the Cray's vectorization. The second included a moderate set of updates to the code, often unwinding loops so they could be vectorized. Generally, the minimal conversions ran roughly the same speed as the 370 to about 2 times its performance (mostly due to a larger exponent range on the Cray), but vectorization led to further increases between 2.5 and 10 times. In one example program, which performed an internal fast Fourier transform, performance improved from the IBM's 47 seconds to 3. another slower MECL 10K 5 4 NOR gate used for address fanout, a 16×4 bit high speed (6 ns) static RAM (SRAM) used for registers and a 1,024×1 bit 48 ns SRAM used for the main memory. These integrated circuits were supplied by Fairchild Semiconductor and Motorola. In all, the Cray 1 contained about 200,000 gates. ICs were mounted on large five layer printed circuit boards, with up to 144 ICs per board. Boards were then mounted back to back for cooling (see below) and placed in twenty four racks containing 72 double boards. The typical module (distinct processing unit) required one or two boards. In all the machine contained 1,662 modules in 113 varieties. Each cable between the modules was a twisted pair, cut to a specific length in order to guarantee the signals arrived at precisely the right time and minimize electrical reflection. Each signal produced by the ECL circuitry was a differential pair, so the signals were balanced. This tended to make the demand on the power supply more constant and reduce switching noise. The load on the power supply was so evenly balanced that Cray boasted that the power supply was unregulated. To the power supply, the entire computer system looked like a simple resistor. The high performance ECL circuitry generated considerable heat, and Cray's designers spent as much effort on the design of the refrigeration system as they did on the rest of the mechanical design. In this case, each circuit board was paired with a second, placed back to back with a sheet of copper between them. The copper sheet conducted heat to the edges of the cage, where liquid Freon running in stainless steel pipes drew it away to the cooling unit below the machine. The first Cray 1 was delayed six months due to problems in the cooling system; lubricant that is normally mixed with the Freon to keep the compressor running would leak through the seals and eventually coat the boards with oil until they shorted out. New welding techniques had to be used to properly seal the tubing. In order to bring maximum speed out of the machine, the entire chassis was bent into a large C shape. Speed dependent portions of the system were placed on the "inside edge" of the chassis, where the wire lengths were shorter. This allowed the cycle time to be decreased to 12.5 ns (80 MHz), not as fast as the 8 ns 8600 he had given up on, but fast enough to beat CDC 7600 and the STAR. NCAR estimated that the overall throughput on the system was 4.5 times that of the CDC 7600. The Cray 1 was built as a 64 bit system, a departure from the 7600/6600, which were 60 bit machines (a change was also planned for the 8600). Addressing was 24 bit, with a maximum of 1,048,576 64 bit words (1 megaword) of main memory, where each word also had eight parity bits for a total of 72 bits per word. Memory was spread across 16 interleaved memory banks, each with a 50 ns cycle time, allowing up to four words to be read per cycle. Smaller configurations could have 0.25 or 0.5 megawords of main memory. Maximum aggregate memory bandwidth was 638 Mbit/s. MFLOPS. (The reciprocal approximation unit could also operate in parallel, but did not deliver a true floating point result two additional multiplications were needed to achieve a full division.) Since the machine was designed to operate on large data sets, the design also dedicated considerable circuitry to I/O. Earlier Cray designs at CDC had included separate computers dedicated to this task, but this was no longer needed. Instead the Cray 1 included four six channel controllers, each of which was given access to main memory once every four cycles. The channels were 16 bits wide and included three control bits and four bits for error correction, so the maximum transfer speed was one word per 100 ns, or 500 thousand words per second for the entire machine. The initial model, the Cray 1A, weighed including the Freon refrigeration system. Configured with 1 million words of main memory, the machine and its power supplies consumed about 115 kW of power; Seymour Cray deliberately made design decisions that sacrificed reliability for speed, but improved his later designs after being questioned on this matter. Similarly, the Cray Operating System (COS) was fairly rudimentary, hardly tested and updated weekly or even daily in the early days.

The Cray 1S, announced in 1979, was an improved Cray 1 that supported a larger main memory of 1, 2 or 4 million words. The larger main memory was made possible through the use of 4,096 x 1 bit bipolar RAM ICs with a 25 ns access time. The Data General minicomputers were optionally replaced with an in house 16 bit design running at 80 MIPS. The I/O subsystem was separated from the main machine, connected to the main system via a 6 Mbit/s control channel and a 100 Mbit/s High Speed Data Channel. This separation made the 1S look like two "half Crays" separated by a few feet, which allowed the I/O system to be expanded as needed. Systems could be bought in a variety of configurations from the S/500 with no I/O and 0.5 million words of memory to the S/4400 with four I/O processors and 4 million words of memory.

The Cray 1M, announced in 1982, replaced the Cray 1S. It had a faster 12 ns cycle time and used less expensive MOS RAM in the main memory. The 1M was supplied in only three versions, the M/1200 with 1 million words in 8 banks, or the M/2200 and M/4200 with 2 or 4 million words in 16 banks. All of these machines included two, three or four I/O processors, and the system added an optional second High Speed Data Channel. Users could add a Solid state Storage Device with 8 to 32 million words of MOS RAM.