The Rise and Fall of Silicon Graphics
or How a Rebellious Youth Briefly Conquered the World
James Henry Clark was born on the 23rd of March in 1944 in Plainview, Texas. Clark’s family was far from wealthy. His father was fond of drinking and couldn’t keep a job. His mother worked at a local doctor’s office making about $225 per month (around $2605 in 2024). Clark’s parents divorced while Clark was still young, and while that salary may seem fine if low adjusted for inflation, Clark’s mother would only have received $175.50 ($2032) after income tax and social security tax, and it was the sole income for a woman and her three children. For himself, Clark was a bit rowdy. His high school highlights include setting off a smoke bomb on the band bus, smuggling a skunk into a school dance, telling his English teacher to go to Hell, drinking, and drag racing. Given the era, I imagine that the drinking was accompanied by chain smoking.
That times were different is… inadequate verbiage. For all the unruly behavior, Clark was only suspended from school twice. On his second suspension, young Clark decided he’d not be returning to school. He chose to join the US Navy and convinced his mother to sign the permission forms. Of course, this is Jim Clark, and the initial days of his naval career didn’t exactly go well. Clark had never taken a multiple choice test. He thought that for many questions more than one of the answers were at least partially true and therefore selected them. The officers in charge of test administration thought that Clark was attempting to fool the computer that checked the answers, and he was immediately sent out to sea with other delinquent recruits where he was given poor treatment, and rough and disgusting chores. The experience of Naval life lit a fire in Clark, and he chose to advance his station in life. He began learning about electronics, taking some general educational courses, and offering loans to other sailors at up to forty percent interest.
His first step was to get his General Education Diploma, which he did. He then enrolled at Tulane. Clark did well at Tulane but transferred to the University of New Orleans from which he received his BS and MA in Physics. He then attended the University of Utah where he earned his Ph.D. in computer science in 1974. From 1974 through 1978, Clark was employed as an assistant professor at UC Santa Cruz, but he left to become an associate professor at Stanford in 1979.
Early in his time at Stanford, Clark worked on a project with Xerox PARC with support from ARPA to develop three dimensional graphics. This led to the creation of the Geometry Engine. In “The Geometry Engine: A VLSI Geometry System for Graphics,” Clark also makes specific reference to Marc Hannah and Lynn Conway as being valuable contributors to the effort. What was the Geometry Engine? It was a special purpose microprocessor that handled matrix math along with point mapping. It featured an instruction set suitable both to 2D and 3D graphics, could generate quadratic/cubic curves and conic sections, worked with both vector and raster based systems, and operated in either integer or floating point systems as needed. In fewer words, Jim Clark and his team at Stanford along with the folks of PARC invented the GPU.
Clark founded Silicon Graphics Inc on the 9th of November in 1981, and he left Stanford early in 1982 to pursue building the company full time with just $25000 in funding (around $85000 in 2024) from a friend and the contents of his own accounts. Accompanying Clark in this adventure were Kurt Akeley, Dave Brown, Tom Davis, Mark Grossman, Marc Hannah, Herb Kuta, Rocky Rhodes, and Abbey Silverstone. While SGI knew they would deal in computers outfitted with a powerful GPU, they did not know precisely what else those computers should feature. As a result, Clark asked potential customers what they’d like to see in a workstation. While at least one potential customer was interested in VMS, NASA’s new Advanced Supercomputing division was very interested in UNIX and they were willing to pay. The division’s director at the time spoke with Clark, and (verbally) committed to purchasing at least eighteen workstations in their first order.
As things began to come together around a product plan, Mayfield invested in the young company. As the development and production of workstations is rather expensive, Clark and SGI’s other founders were forced to sell more and more of the company’s ownership to keep operating. The first product to ship was the IRIS 1000, where IRIS meant Integrated Raster Imaging System, in November of 1983. This machine was intended for use as a terminal for a VAX-11 and featured a Motorola 68000 clocked at 8 MHz with 768K RAM, a Geometry Engine clocked at 6 MHz capable of over six million geometric floating point operations per second, and a 10 Mbps ethernet NIC. The cabinet of the IRIS 1000 was ten inches wide, twenty one inches tall, twenty seven inches deep, and when fully assembled weighed in at one hundred pounds with a ten slot backplane. This machine was followed by the IRIS 1200 which was the same machine but with a twenty slot backplane. These were followed by workstation models 1400 and 1500 in April of 1984 which upgraded the CPU to the Motorola 68010 clocked at 10 MHz with 1.5M of RAM. These machines were differentiated from one another in the size of HDD they featured with the 1500 having been larger. The 1400 featured a 72MB winchester disk, while the 1500 featured 474MB of SMD. Both of these ran a UNIX SVR4 variant with BSD enhancements called GL2, and they featured twenty slot backplanes. The main system boards in these four machines were licensed from Andy Bechtolsheim just before he founded Sun Microsystems. The 1000 and 1200 used the PM1 and the 1400 and 1500 used the PM2. These were not cheap systems with the IRIS 1000 having a price of $22500 (around $67200 in 2024) and the 1400 having a price of $35700 in 1984 (around $106600 in 2024). These twenty slot machines were eighteen inches wide, twenty nine inches tall, and twenty seven inches deep, and fully assembled weighed in at two hundred pounds. By the time the first of these machines sold to Carnegie-Mellon University’s Electronic Imaging Lab, the founders of SGI owned very little of their company.
From nearly the first day that SGI’s hardware was on the market, software developers began trying to exploit the machines’ graphics capabilities. A rather prominent example of this was Wavefront Technologies in Santa Barbara led by Bill Kovacs, Larry Barels, and Mark Sylvester. Their first product was called Preview and launched in 1984 on SGI’s hardware. Their customer list included Universal Studios, NBC, NASA, and Electronic Arts. Naturally, this also informs us that these companies were using SGI hardware.
Given the outline of his youth, it isn’t very surprising that Clark was a hands-off kind of manager. He would hire the brightest minds he could, set a general target, and then let people go after it however they saw fit. There are two narratives for what follows. The first and most common that I’ve read was that Mayfield didn’t much care for Clark’s management style and they brought Ed McCracken formerly of HP in as CEO. The second narrative states that Clark didn’t care for running the company and brought McCracken in on his own accord. Whatever the case, McCracken stated of Clark:
Jim's not a day-to-day person. He works in his own time frame. He takes complex things and makes it simple. It might take a month, a day, or a year. He gets in these moods for a while where he's almost unavailable. He's most effective when he's in that mood.
In August of 1985, the company introduced the IRIS 2000 series of workstations. These machines were all based upon the the PM2 system board featuring the Motorola 68010 clocked at 10 MHz with a floating point coprocessor (SKYFPM-M-03). Naturally, these all featured the graphics engine as well. The IRIS 2000 and 2200 were ten slot backplane, shipped without a disk, and were intended for use as terminals. The 2300 and 2400 were twenty slot backplane and shipped with winchester disks. The IRIS 2500 was rackmount and used SMD disks. The 2000 series used a Geometry Engine clocked at 8 MHz. A few months after the initial launch of these upgraded machines, SGI launched the turbo line. This included the 2300T, 2400T, and 2500T which featured the IP2 system board with a Motorola 68020 clocked at 16 MHz, an FP1 floating point unit, and 2MB to 16MB of RAM. The RAM of the turbo units used a newer, faster, local bus. As a result, the RAM between turbo and non-turbo systems could not be mixed. This was an important bit of information as SGI did offer turbo upgrades for non-turbo systems that would then require the purchase of expensive proprietary memory.
In January of 1986, SGI made their initial public offering raising $17.2 million (nearly $49 million in 2024) with trading having started at $3 per share and topping $30 on the day. The following month, the company introduced the IRIS 3000 line. These are very similar to the IRIS 2000 turbo machines but with Enhanced IRIS Graphics. These featured either ten or twelve Graphics Engines clocked at 10 MHz with either eight or thirty two bitplanes depending upon configuration. The 3000 line could be ordered with either winchester disk drives, ESDI drives, or SMD.
Also in 1986, Control Data Corporation and Silicon Graphics signed a deal under which CDC would resell IRIS machines under CDC’s own branding. As far as I know, no complete listing of which models sold under what naming survives today, but it is known that the IRIS 3130 was resold as the CDC Cyber 910. This would make it a machine with twelve GEs at 10 MHz and ESDI drives.
In March of 1987, Silicon Graphics announced a new machine that marked a major transition for the company. The Professional Iris was a RISC machine built around the R2000 from MIPS Computer Systems (another project started at Stanford and spun out as its own company) clocked at 8 MHz. The company’s press release read:
The first member of the Iris line is the 4D/60, a RISC superworkstation with a 32-bit 8 MHz CPU from MIPS Computer Systems. It offer performance three times that of the Silicon Graphics Iris 3100 series. The graphics performance has been enhanced with 38 custom and semicustom graphic chips. It performs 140,000 32 bit three dimensional floating point transformations per second and renders over 4,500 100-pixel polygons per second with smooth shading and hidden surface removal. It offers 24 colour bit-planes for more than 16 million colours; four user-accessible system planes for overlay or underlay, menu and windowing functions; a 24-bit Z-buffer enabling hidden surface removal with greater accuracy and realism; high-level primitives such as splines and surfaces for more accurate renderings; and a multi-mode graphics windowing environment.
Standard configuration includes 4Mb CPU, eight colour bit-planes for 256 colours); four system planes, a Weitek-based floating point accelerator board; a 170Mb ESDI disk and controller; a 19″ 1,280 by 1,024 60Hz non-interlaced colour monitor; keyboard and mouse; and a floor-standing chassis with 12 VME slots and a 1,000-watt power supply.
Software compatible with the previous generation, it runs Unix System V.3 with a base price of $74,000.
The Professional Iris line included the 4D/60 mentioned in the press release followed by the 4D/50, 4D/70, 4D/80, and 4D/85. All of these featured the R2000 CPU with a floating point coprocessor. The 50 and 60 had an R2000 clocked at 8 MHz, while the 70 was at 12.5 MHz, and the 80 and 85 were clocked at 16.7 MHz. For comparison to other architectures, the 4D/50 was capable of seven million instructions per second, the 70 was capable of ten million, and the 80 was capable of thirteen million. The 50 and 60 had memory configurations starting at 4MB and upgradeable to 12MB. The rest of the lineup started at 8MB and could be upgraded to a maximum of 144MB. The first of the 4D/60, 50, and 70 systems to ship utilized the Clover 1 graphics system. Later models shipped with Clover 2 branded as IRIS GT. IRIS GT brought hardware support for lighting, smooth shading, antialiasing, pan/zoom of images, arbitrarily shaped windows, and other rather modern capabilities. Importantly, the bus for this system was a proprietary 64 bit bus. The actual chips powering all of the graphics capabilities were still the Graphics Engines, but these were updated some and they were capable of twenty million floating point operations per second. The Professional Iris series brought an end to the disk anarchy of the previous lineup and all systems utilized SCSI hard disks, and QIC-120 tape drives were also available. These systems were resold by both Control Data Corporation and Prime Computers. The UNIX version mentioned in the press release was SGI’s 4D1 which would later be renamed IRIX.
On the 29th of March in 1988, Control Data Corporation announced that it would be acquiring twenty percent of Silicon Graphics for $68.9 million (nearly $181 million in 2024) and extending its licensing deal for reselling SGI’s machines with an agreement to purchase $150 million (around $393 million in 2024) in hardware over the next three years.
On the 16th of September in 1988, SGI announced that IBM would be purchasing graphics cards and licensing IRIS GL, the software library for SGI’s graphics, for use in the IBM RS/6000 POWERStation. McCracken commented:
We are pleased to establish a relationship with IBM and look forward to working with them. The agreement reinforces our long-time conviction that three-dimensional graphics will become a mainstream technology in the computer industry. As real-time 3D graphics is made more affordable, the rapid growth that the 3D workstation industry is experiencing will continue to escalate.
The card in question was the IrisVision, and while I refer to it as a card, it was really two cards. The primary card held the Graphics Engine and daughter cards held the framebuffer and z-buffer memories totaling 5MB for the framebuffer and 3.75MB for the z-buffer. The primary card connected to the computer via its MCA bus edge connector, and it provided a DE-15 connector for display attachment. Overall, the IrisVision MCA card’s hardware was extremely similar to the graphics system in the SGI Personal Iris series introduced in 1987. It featured SGI’s fifth generation geometry processing pipeline (referred to as GE5, or Graphics Engine five), either an eight or twenty four bit per pixel frame buffer, and twenty four bits per pixel z-buffer. Also, just as the workstations’ hardware did, the IrisVision implemented the entire IrisGL API in hardware. The primary difference in IrisVision was the presence of a VGA (DE-15) passthrough for 2D graphics. In the course of the IrisVision’s development, an IBM PS/2 running OS/2 was used for testing and development. This resulted not only in a minimal OS/2 driver, but also in an ISA version of the IrisVision being developed. Ultimately, the only major customer SGI had managed to obtain was IBM for the MCA card for the RS/6000 UNIX workstations. Their struggle may have been that the card was priced at $4995 (just over $13000 in 2024). The company ultimately spun off the entire project as a separate company, Pellucid, which didn’t fare well. The former SGI employees who started Pellucid still managed to change the world when they founded 3dfx which used similar technology as well as the passthrough for 2D graphics.
SGI held a rather firm grasp on high-end graphics workstations, but hadn’t yet made a push into the entry level market. This changed with the introduction of the Personal Iris lineup. The line started with the 4D/20 which made use of a R2000 CPU from MIPS clocked at 12.5 MHz achieving ten million instructions per second. The other three machines made use of the R3000. In the 4D/25 the R3000 was clocked at 20 MHz achieving sixteen million instructions per second. In the 4D/30, the clock speed was pushed to 30 MHz and the performance was bumped to twenty seven million instructions per second. The highest performance model was the 4D/35 at 36 MHz and thirty three million instructions per second. Maximum memory supported on these systems was 128MB. Personal Iris systems were sold by both SGI and Control Data as expected, but these systems were also offered rebadged by the somewhat newly reconstituted Groupe Bull. From what I can find, Bull’s sales of rebadged SGI machines weren’t great; they had better luck with NEC hardware. For the naming “Personal Iris” and the thought that SGI would be attacking the “low-end” of the workstation market… the pricing wasn’t all that reflective unless one were to compare to “high-end” SGI machines which could reach lofty prices of about $100000 (about $262000 in 2024). The Personal Iris line started at $20000 (roughly $52000 in 2024).
The other, much higher end and far more expensive, SGI lineup introduced at this time was the PowerSeries which were multi-processor systems (up to eight CPUs) and could be deskside or rackmount. These systems could also support higher clocks at up to 40 MHz which in combination with up to eight processors could mean performance over two hundred thirty million instructions per second. The power of these systems was put to use in the movies The Abyss, Terminator 2, and Jurassic Park among many more.
In March of 1991, Compaq acquired thirteen percent of SGI for $135 million (around $307 million in 2024) along with an agreement to invest another $50 million (about $114 million in 2024) in the development of a new workstation that would be priced at around $7500 (roughly $17100 in 2024).
The most famous and beloved SGI systems were introduced from 1991 to 1995. These models were the Indigo, Indigo 2, and the Indy. The corresponding high-end systems were the Crimson, and Challenge series. The first Indigo system released in 1991 featured a MIPS R3000 CPU clocked at 30 MHz. The Indigo (and Crimson) moved SGI’s systems to 64 bit MIPS CPUs starting with the R4000 at 100 MHz and the R4400 at 150 MHz in 1992. The 150 MHz part in an Indigo could achieve one hundred twenty million instructions per second. The Indigo 2 was first introduced in 1993 with the MIPS R4400 CPU and “Extreme” graphics. The Indy was lowest end SKU of the three, and it was introduced in July of 1993 with a 100 MHz R4000PC CPU, 24 bit graphics system, 16MB of RAM, the IRIX operating system, a fifteen inch monitor, and a price of $4995 (about $10700 in 2024).
On the 13th of March in 1992 announced that it was acquiring MIPS Computer Systems via a stock swap worth about $333 million (around $737 million in 2024). This followed MIPS having had financial problems, high employee turnover, and the exit of the company’s president, Charles Boesenberg, one month earlier. For SGI, the acquisition ensured their part supply. MIPS Computer Systems became MIPS Technologies. The combined company had revenues at around $1 billion (about $2.21 billion in 2024). However, the large acquisition did mean that SGI posted a loss on the year of about $118 million (or $261 million in 2024). This move also briefly brought SGI into the ACE alliance that aimed to build a workstation standard on the MIPS CPU and the UNIX operating system as well as the 80386/486 and NT. This group was built of Compaq, MIPS, Microsoft, DEC, SCO, Acer, CDC, Kubota, Olivetti, NKK, Prime Computer, Pyramid Technology, Siemens, Sony, Sumitomo, Tandem, Zenith, and Wang. SGI and Compaq left the alliance rather promptly. This could be due to their own arrangement not long before, but ACE fell apart completely not much later anyway. I suspect that no strong alliance of fierce competitors would last long in a market that was shrinking due to low-cost commodity hardware and software consistently improving year over year in the PC compatible market. Yet, the SGI Indigo 2, Indy, Challenge and a few more were mildly compliant with the ACE ARC (Advanced RISC Computing) standard.
On the 30th of June in 1992, Silicon Graphics released OpenGL. This was a cross-platform API for both 2D and 3D graphics allowing hardware acceleration of rendering via one or more GPUs descended directly from IRIS GL. Unlike its predecessor, OpenGL did not have windowing, and it didn’t offer a mouse or keyboard API. IRIS GL had been developed before X and other graphical environments were available, and therefore had needed those features, but OpenGL had no such requirements. Another major change in the transition to OpenGL regarded feature availability. IRIS GL presupposed the use of SGI’s hardware. OpenGL could not make such an assumption, and as a result it allowed features not supported by a GPU to be rendered in software by the CPU. One customer this would positively affect was Microsoft who’d licensed IRIS GL for inclusion in NT in 1991.
At the end of 1992, Jim Clark met with Nintendo CEO Hiroshi Yamauchi to discuss bringing 3D graphics to Nintendo’s next game console. In many ways, the Nintendo 64 was an SGI workstation in miniature with a MIPS R4300 CPU clocked at 93.75 MHz offering one hundred twenty five million instructions per second, 4MB of Rambus DRAM at 250 MHz (actually 4.5MB but 512K is visible only to the GPU) which could be doubled with a RAM expansion pack, and the Reality coprocessor clocked at 62.5 MHz which offered the SGI GraphicsEngine (though a more modest version). The system supported 16.8 million colors, a maximum resolution of 640x480, and audio sampled at up to 44.1 KHz. Unfortunately, the design of the system meant that the full capabilities would almost never be fully realized. For example, there was no dedicated sound chip, so high sample rates would tax the CPU, and while the R4300 is 64 bit, the Nintendo 64 had a 32 bit data bus. Yet, showing the nature of the hardware packed into the Nintendo 64 is the Nintendo 64DD. This offered 64MB read/write magnetic disks (similar to Zip), a real time clock, internet connectivity via a 28.8 kbps modem, keyboard, mouse, and audio/video capture effectively transforming the Nintendo 64 into a small workstation. The expansion, after significant delays and a one year two and a half month life on the market, was a commercial failure. The Nintendo 64 itself, however, was a huge success following its release in 1996.
Industrial Light and Magic had been using SGI hardware since 1987, and on the 8th of April in 1993, they announced a partnership with SGI to create the Joint Environment for Digital Imaging, or JEDI. This allowed the two companies to gain insight from each other’s work. SGI got access to much of ILM’s software expertise while ILM got access to the latest and greatest hardware at a discount.
In 1994, Jim Clark left SGI, sold his shares in the company, and went on to partner with Marc Andreessen and start Netscape.
In 1995, SGI spent about $500 million (or $1 billion in 2024) acquiring Alias Research, Kroyer Films, and Wavefront technologies. At roughly the same time, SGI worked with DreamWorks SKG to form DreamWorks Digital Studio where these newly acquired companies’ products could be put to good use.
On the 26th of February in 1996, Silicon Graphics acquired Cray Research for $740 million (or $1.47 billion in 2024). This gave SGI control of around forty percent of the high performance computing market at the time. While many industry analysts speculated about SGI’s motives, Cray was struggling to survive and they had multiple installations at NASA. While SGI had been successful in entertainment, that sector accounted only for something around ten percent of SGI’s annual revenues. The bulk of SGI’s customers were governmental, so much so that SGI created the wholly owned subsidiary Silicon Graphics Federal Inc to hold those contracts and provide service and support for governmental organizations. In this way, SGI was essentially making sure they couldn’t lose one of their largest and most valuable customers, NASA, as they’d be the provider of not only workstations but also the support and service of NASA’s supercomputers. The supercomputer relationship benefited SGI all the way to 2008 with Pleiades.
The new SGI workstations of 1996 were the O2 and O2+ series. These systems were very different from both their predecessors and successors in that they utilized a unified memory architecture via the Memory & Rendering ASIC (MRE). The MRE had direct paths to all parts of the O2 such as the CPU, memory, I/O, compression, display, and imaging. Due to this structure, graphics hardware wasn’t optional but rather integral to the system’s design. The O2 could come equipped with an R5000, RM5200SC, RM7000A, R10000, or R12000 CPU. Frequencies ranged from 180 MHz to 400 MHz, all options had on-board floating point support, and could support up to 1GB of unified memory via eight 128MB DIMMs of one hundred thirty nine pin SDRAM.
The high-end deskside and rackmount options made available at this time were the Origin 2000, Origin 200, and Onyx 2 series. These were multiple CPU systems with distributed, shared-memory architecture called S2MP. The Origin 200 was the entry level system, the Onyx 2 was a step up, and the Origin 2000 was the premium SGI branded system and was rackmount. This series also had Cray Origin at the super-premium level with up to one hundred twenty eight R10000 CPUs. The IRIX operating system shipped with these models supported SMP.
On the 14th of May in 1997, SGI announced the acquisition of ParaGraph International Inc. ParaGraph was a vendor of VRML and web graphics software, and after the acquisition the company and its assets were moved to Mountain View with the new name of Cosmo Software. McCracken commented:
One of the most important long-term growth opportunities for Silicon Graphics is to empower the designers, developers, and service providers of the Second Web. With the acquisition of the leading PC 3D Internet company and the formation of Cosmo Software, we are increasing our investment and reinforcing our leadership in the market for the software and services that will bring about this new interactive medium.
Bringing the technologies of Onyx 2 series to the midrange workstation was the Octane, released in January of 1997. This was the a desktop machine instead of deskside, but it supported dual CPUs. This line featured the crossbar switch that debuted in the high-end and server machines of the prior year. The concept was that instead of a traditional shared bus, each subsystem could communicate with any other without interference. The crossbar switch had seven ports: HEART ASIC (CPU and memory), graphics (Impact [first or second generation] or VPro), XIO B, XIO C, XIO D, built-in I/O, PCI bridge. The Octane did have a higher-end version, the Octane 2, which featured more powerful CPUs and GPUs, higher density memory support, and a beefier PSU. CPUs in the Octane ranged from the R10000 at 175 MHz to the R14000A at 600 MHz, and RAM ranged from 64MB to 2GB.
Silicon Graphics didn’t do too well in 1997 overall. For revenues of $3.6 billion (or $7 billion in 2024) the company posted a loss of $78.6 million (roughly $152 million in 2024). On the 29th of October in 1997, Ed McCracken resigned as did the executive vice president of sales and marketing, Gary Lauer. The company then laid off around nine percent of its employees (about seven hundred people). Richard Belluzzo (formerly at HP) took over as CEO and Robert H. Ewald who was already the executive vice president of computer systems (formerly president of Cray) took over Lauer’s job duties. Some sources claim that McCracken was forced out, but this isn’t accurate. At the annual shareholder meeting in Palo Alto, McCracken announced his resignation stating: “after a great deal of thought, the time is right for me and the company to make a change.” He then proceeded to find and to hire his replacement himself.
Around this time, Silicon Graphics filed a lawsuit against a startup called ArtX. ArtX was founded by Dr. Wei Yen and around nineteen other SGI employees who’d worked on the Nintendo 64. The company’s original goal was to develop a PC graphics chip that would rival 3dfx. Then, in May of 1998, the company gained a contract to develop a graphics processor for Nintendo’s next generation game console, the GameCube. At COMDEX in the autumn of 1999, the company unveiled the Aladdin 7 chipset which shipped as integrated GPUs on K6-2 and K6-3 motherboards made by Acer Labs. ArtX was bought by ATI in February of 2000. ArtX’s technology was incorporated into ATI’s GPUs from 2002 until roughly 2005. SGI’s lawsuit against ArtX was quietly dropped in 1998 without any settlement having been reached.
On the 1st of January in 1998, shortly after taking over as CEO, Belluzzo sold two of SGI’s PCB factories and restructured the company from twenty six groups to just five. SGI then setup MIPS Technologies as its own legal entity (though SGI maintained a majority ownership), terminated the Cosmo software business, and proceeded to make customers hesitant to continue investments into the MIPS architecture by announcing SGI’s intent to migrate to Itanium (and collaborating on projects Monterey and Trillian) while simultaneously launching an IA-32 series of machines running NT known as the Visual Workstation. Additionally, the company began outsourcing the manufacturing of their computers, and cut the operating budget by about $200 million (or $381 million in 2024). In Spring of 1998, the company announced a lawsuit against NVIDIA for patent infringement.
None of this helped to change the overall direction of the company. Revenues fell to $3.1 billion and the company posted a loss of $460 million for 1998. On the 20th of July in 1999, without adequate funding to continue the lawsuit against NVIDIA, SGI and NVIDIA agreed to license one another their respective patent portfolios. The company continued to lose money, and Belluzzo left on the 22nd of August in 1999 to lead Microsoft’s MSN division.
As Bob Bishop took the reigns of SGI, things looked dark. AMD announced their 64 bit architecture in October, PC graphics had made massive strides while remaining significantly less expensive than SGI’s offerings, NT was proving to be a solid and less expensive competitor to UNIX, Linux was eating away at traditional UNIX market segments, and Itanium still hadn’t launched. By this point, the company had no fall back as they’d mostly stopped investment into new MIPS CPUs.
On the 2nd of March in 2000, SGI sold Cray to Tera Computer for $22 million (or $40 million in 2024). Tera promptly renamed itself to Cray Inc as it took on an installed base of six hundred supercomputers and two hundred customers across thirty different countries.
SGI’s final MIPS workstations were the Fuel and Tezro lines. Fuel was introduced in 2002 with the R14000 clocked at 500 MHz, up to 4GB of DDR SDRAM, and SGI’s VPro graphics. Models were available with up to an R16000 CPU clocked at 900 MHz. The Tezro was launched in 2003 starting at $20500 (or $34574 in 2024). This model featured only the R16000 and could be configured at clock speeds from 600 MHz to 1 GHz with 512MB to 8GB of DDR SDRAM and SGI’s VPro graphics. Fuel workstations were single CPU, but Tezro was offered with one to four CPUs. While SGI’s IA-32, Itanium, and Xeon workstations and servers sold, they didn’t make much money. On the 10th of July in 2003, SGI vacated and leased their headquarters to Google.
As SGI’s fortunes continued to decline, the company sold Alias Systems (formerly Alias|Wavefront) for $58 million on the 16th of April in 2004 to Accel-KKR (roughly $95 million in 2024). Then, in November of 2005, SGI was delisted from the NYSE due its stock price sinking below the minimum required. In January of 2006, Dennis McKenna was hired as president and CEO, and named chairman of the board. Bishop remained on the board of directors and served as vice chairman. On the 8th of May, the company filed for bankruptcy protections. The campus leased to Google was sold to Google in June of 2006 for $319 million (or $491 million in 2024). It’s prior home in Mountain View had been sold to the Computer History Museum in 2002. The company emerged from bankruptcy and was relisted in October. The official end of SGI’s MIPS and IRIX came on the 29th of December in 2006 with the final orders being fulfilled in March of 2007. Bob Ewald replaced McKenna as CEO on the 9th of April in 2007. In December of 2008, SGI was again delisted. On the 1st of April in 2009, the company filed for bankruptcy, and was subsequently purchased by Rackable Systems for around $42 million on the 11th of May in 2009 (roughly $65 million in 2024). Rackable renamed itself Silicon Graphics International following the acquisition, and it was later bought by Hewlett Packard Enterprise.