Die shot of the MIPS R10000 microprocessor, fabricated by NEC as the NEC VR10000 (with a story!)

CPU-posting on main

MTI = MIPS Technologies (company that made MIPS (Microprocessor without Interlocked Pipeline Stages) processors, they make RISC-V processors now lmao)

At the time when the MIPS R10000, known as the "T5" while in development, was being designed, MTI had made a name for themselves as designers of high-performance computer microprocessors along the lines of the then-new philosophy of reduced instruction set computing (RISC). Actually, their R2000 design was the first commercially-available RISC microprocessor. By the time the T5 was being designed, they were no longer alone in the RISC microprocessor market. Several companies, including IBM and Motorola (joined together in the AIM alliance which produced PowerPC), DEC (who designed the Alpha line of RISC microprocessors after MTI owned them in the 80s when their radically simpler chips were performing better than VAXen), and Sun Microsystems (who were making the SPARC line of microprocessors) were now marketing RISC microprocessors. Not just even marketing but beating MTI in the market they had created. After trying and failing to develop their own complete computer systems alongside their chips, they were having financial difficulties until Silicon Graphics acquired MTI to secure availability of MIPS microprocessors for their famous ("it's a Unix system, I know this!") MIPS-based workstations and servers. Although their new (in 1993) R4000 and R4400 designs performed well compared to their contemporaries, they were quickly being made obsolete by MTI's competitor's new offerings and they were left with a problem:

The MIPS R4000 and the R4400, which is essentially an R4000 with bigger on-die caches, were more or less just an architectural evolution from the R2000. The R4000 made its performance in much the same way as the R2000 did, the classic RISC design process mantra: "let's make it simpler" and thus be able to run it faster. In particular, what this means for the R4000, and what is a key difference from its predecessors and its contemporaries, is a technique called superpipelining. In an instruction pipeline, the maximum speed at which your processor can issue instructions is set by the pipeline stage which takes the longest to complete. Superpipelining is one way of addressing this problem: you can subdivide each pipeline stage into 2 simpler pipeline stages that individually complete faster and thus be able to clock your chip faster without problems. However, this has its limits. Eventually, it becomes impossible to further "deepen" the pipeline like this or clock the processor faster in general without other problems. This is why MTI's competitors opted for the analogous superscalar approach: you can duplicate functional units of your processor and have multiple instructions "in flight" at the same time and usually this also involves multiple pipelines. At the time MTI thought this approach would result in more consistently higher performance (not to mention save die space) but were quickly proven wrong when their competitor's superscalar (and often with other architectural tricks) chips were outperforming the R4000 in spite of MTI's fabrication partners constantly improving their process and releasing chips that ran at higher and higher speeds.

Enter the MIPS R8000 (die not pictured here) in 1994, a weird and expensive 6-chip 4-way superscalar design meant for the high-end microprocessor market while the next-generation T5 (which would become the MIPS R10000, as mentioned earlier) was under development. It didn't sell well because of its high price and the fact that its integer performance, important for general-purpose computing applications, was lacking compared to the 200-MHz R4400 that was being sold by then. It did, however, have impressive floating-point performance, which landed many R8000-based systems in the TOP500 supercomputer list for a time. But this design could never be the high-performance and general-purpose processor MTI needed to compete with their competitor's offerings...

Introduced in 1996, the MIPS R10000 (die IS pictured here) was a significant departure from the architecture of the R4000 (which more or less was directly derived from the first research done at Stanford University where MIPS was initially created over a decade earlier). Dropping the superpipeline approach, the R10000 is a 4-way superscalar processor even capable of executing instructions out of order! Another big change is that it has a branch predictor and speculatively executes instructions after a branch as opposed to the R4000, which used the classic MIPS "branch delay slot" technique to schedule one more instruction in the pipeline after a branch and then stall lol (they should have added even more delay slots, caring about binary compatibility is liberalism). It's hard to find benchmarks for something this old but this design performed at least several times faster than an R4400 at about the same clock speed!

If you like my CPU posting and want me to post more in the future let me know

Also ask me any questions if you want too and I'll try to answer