NexGen: Gone But Not Forgotten
If you want to buy a new CPU for your desktop PC these days, you’re pretty much limited to AMD or Intel processors. But four decades ago, the CPU market was quite different and far more diverse, with numerous products from multiple companies on offer.
One such example was NexGen, which had backing from big PC names, plenty of capital, and a team of great engineers. NexGen also had some brilliant ideas to set them apart from the crowd. But then, why did they only last for ten years?
Imitation was the norm 40 years ago
To begin our story, we must head back to the mid-1980s, a time when central microprocessors were still relatively new inventions. Intel’s first CPU, the 4004, had been released a decade before, but by the time the 8086 was launched in 1978, the market was awash with competing models, many of which were essentially clones.
Some of these derivatives were produced under license, some others were reversed-engineered, and in a few cases, companies poached design team engineers to create new chips. For example, the Zilog Z80 was created by some of the Intel 4040 team, and it proved to be faster and cheaper than the 8080, on which it was based.
It was no different for the Intel i386 (also known as the 80386), a processor that’s of particular interest to this tale. Launched in 1985, Intel’s first 32-bit processor was powerful and expensive, and would go on to be subject to various cloning experiments by AMD, Cyrix, IBM, and others.
In that same year, Thampy Thomas, a doctorate of electronic engineering and semiconductor entrepreneur, launched a new fabless processor design company called NexGen. He co-founded it with Nick Tredennick, another well-regarded engineer, NexGen had grand ambitions to beat Intel and the rest — without recourse to design imitation.
With funding from the likes of Compaq (in the US) and Olivetti S.p.A. (Italy), software publisher ASCII Corp (Japan), and venture capitalist firm Kleiner Perkins, Tredennick used his connections in the industry and academia to accrue some of the best engineers around.
Compaq was more than just a giver of funds, though — the PC giant was concerned that Intel had plans to move away from the CISC-based x86 architecture used in the i386, the CPU of choice in Compaq’s computers, and just stick to its RISC designs instead.
Thomas had become enamored with RISC (Reduced Instruction Set Computers) during his time at Stanford and felt that with the limits of the transistor budget available in processor design at that time, this approach was the future.
RISC vs CISC: Two ways to crunch the numbers
RISC essentially involves having the hardware carry out all of the instructions in as fast a time as possible — usually just a single clock cycle. To achieve this, the parts that carry out the tasks only work with the simplest of fixed-length commands. Multiplying two values, for example, requires the processor to be issued with at least four instructions to load the two pieces of data, multiply them, and then store the result somewhere.
Complex Instruction Set Computer (CISC) designs go the other way — the processor can be issued with variable-length commands that the chip itself breaks down into all of the various instructions required. The above multiplying example can be done with just one command. The downside to this is that a fair amount of the transistor budget in the processor design has to be spent on hardware that will handle the translation.
RISC processors are ultimately less complex, smaller, and cheaper to fabricate than CISC ones, but they need appropriate software support — operating systems and compilers need to manage the instruction processing because a RISC-based CPU can’t really do it.
But at the time of i386, the mainstream computer industry was using software developed primarily for CISC architectures, the majority being 16-bit and 32-bit x86 processors. If Intel really was going to change direction, Compaq needed to be ready to stay ahead of the competition.
So with its backers keen for a design to be able to switch to a new paradigm, IBM booked to fabricate the chips, and NexGen equipped to design it, everything was ready to go.
Nx586: A hesitant start yields a promising first model
What really sold Compaq on NexGen was the idea it had for its CPUs — eventually marketed as RISC86, the architecture would consist of a fast and efficient RISC core but with hardware that would convert CISC instructions, making them fully compatible with x86-based software. In other words, the engineers were aiming to have the best of both worlds.
But getting something to market wasn’t plain sailing. Considerable time had been spent on the so-called F86 project, with the first revision comprising 8 separate chips, making it financially non-viable.
The work didn’t bear fruits until 1994, almost 8 years after the company was founded, and during that time Intel settled its mind about RISC, and committed to the x86 architecture with the release of the i486 (1986) and the first Pentium, codenamed the P5, in 1993.
NexGen’s first commercially available processor, the Nx586, was markedly different from the i486, Pentium, and the raft of clones. The chip wasn’t pin-compatible with any motherboard on the market, so NexGen had to design and produce its own chipset (NxVL) — operating systems initially struggled to correctly recognize the Nx586 and the motherboard, too.
Fortunately, the upstart CPU had enough to set it apart from the competition. Manufactured by IBM using its 500 nm process node, clock speeds were just as good as Intel’s, ranging from 70 to 110 MHz. The cache system was especially strong, with 16 kB of L1 instruction and data cache respectively, and unlike the P5, the L2 cache controller was built into the CPU, reducing all important latencies.
Like the Pentium, the Nx586 was superscalar in nature (three-way execution), processing two integer math and one address instructions at the same time; it also had a decent branch predictor unit and 64-bit wide data buses all around.
With AMD coming late to the market with its P5 competitor (the AMD K5 didn’t launch until 1996) and Cyrix’s 6×86 lacking in performance, NexGen had made an impressive entrance into the CPU fray, with a product that was cheaper than Intel and performed just as well.
The battle gets harder but the momentum gathers
While the NexGen Nx586 was strong in integer calculations, it couldn’t compete against the Pentium when it came to floating point math, as it had no physical support for such numbers. This was okay for the DOS-world of business, but PC enthusiasts (especially those who enjoyed gaming) would have to wait until the end of 1995 before NexGen could something about this.
Delays prevented the Nx587 floating point co-processor from being ready at the same time as the Nx586, and eventually, the two would be merged into a single package (two dies on the same CPU board).
As was common at the time, NexGen used the P-rating system to label its products — for example, the Nx586-P80 ran at 75MHz, but had a performance equivalent to Intel’s 80MHz Pentium. Models with the floating point co-processor were given a PF moniker (e.g. Nx586-PF100).
Unfortunately, despite Compaq’s financial commitment to the project, the company didn’t widely use the new chips in its PCs. The need for bespoke motherboards and higher production costs for the multi-die PF processors took the shine off NexGen’s achievements.
Nor did it help that Intel was making rapid progress with its Pentium models. By the start of 1996, the P5 was hitting 166MHz, and there were versions of the architecture that offered backward compatibility with the older i486-era chipsets.
NexGen wasn’t finished though, having already demonstrated its second generation RISC86 processor in October 1995, the Nx686. The company promised to fix everything that was holding back the first one.
To begin with, the new chip would be a single die, smaller in size and with much higher clocks. Deep inside the layout, the L1 data cache size and the number of integer pipelines and x86 instruction decoders would all be doubled.
The Nx686 would even offer new multimedia instruction extensions and be usable in motherboards with the NxVL chipset. In short, the perfect update for current Nx586 users and a far better product to sway customers away from Intel.
The latter was especially important as just a month after NexGen’s press launch, Intel released its brand new P6 architecture, under the name of Pentium Pro. Just like the Nx586, it too dynamically decoded x86 commands into much smaller instructions, making it easier to improve the efficiency of the whole processor.
Sadly, the Nx686 never reached the market, at least not as a NexGen product.
The AMD acquisition: Lofty ambitions meet a surprising announcement
For over a decade, AMD had two major processor design teams: one working on x86 clones and the other developing RISC chips for the embedded market. Around the same time as NexGen was announcing its second generation of CPUs, AMD launched its Am5x86 processor.
This would be AMD’s last Intel-derived chip as the company’s management decided to utilize the strengths of its RISC designs in the desktop and workstation market. Hence why the successor to the Am5x86, called the K5, would take the same approach as NexGen — a RISC-based core, with x86 instruction decoders.
On paper, the K5 was a Pentium-beating design, but the reality of manufacturing such a chip proved to be more challenging and time-consuming than AMD hoped for. The first K5 models weren’t ready until March 1996 and unusually for a new architecture release, they weren’t all-singing-and-dancing versions of the design.
Labeled with the codename SSA/5, these chips had the branch prediction unit completely disabled (due to a wonky branch target buffer), had longer processing wait states for certain operations, and had a high heat output due to being made on the older 500 nm process.
AMD would eventually fix those issues within 8 months, but the damage had already been done, and the K5’s reputation was firmly in the bargain-bin sector of the processor world. Keen to avoid making the same mistakes with their next architecture, the company made a rather surprising decision.
AMD bought NexGen for $857 million in stock in October 1995.
It soon transpired that this was money very well spent.
The purchase of small semiconductor firms by larger corporations wasn’t unusual, even back then, but with this costly acquisition, AMD was openly admitting it didn’t have the capability to directly compete against Intel.
It soon transpired that this was money very well spent. AMD took NexGen’s Nx686 design and rejigged a few elements — for example, the pin layout and associated wiring were adjusted to be compatible with Intel’s Socket 7 format, with the dedicated 64-but bus for the L2 cache being dropped as a result.
The modified design was released as the widely successful AMD K6 series of CPUs in April 1997 and the new chip performed just as well as the updated Pentiums, except in applications heavy with floating point calculation. Naturally, it cost less to buy than an equivalent Pentium, as this was the standard marketing approach for Intel’s competitors.
AMD would go on to further revise the Nx686-based architecture, expanding the multimedia support (marketed as 3DNow!), and use the CISC front end, RISC execution concept for every x86 CPU released after the K6 (as would Intel).
An everlasting end to an innovative company
And that was the end of NexGen. Any remaining Nx586 chips still on shelves or in working computers would rapidly disappear, replaced by AMD versions, and the name itself would quickly just become a footnote in the history of the microprocessor.
But what a footnote.
All modern x86 CPUs, Intel’s and AMD’s, owe tribute to NexGen’s approach to architecture design. The fundamental execution stages in Core and Ryzen processors are RISC in nature and sport front ends packed with cleverly designed instruction decoders to convert CISC-type commands.
So while the company itself is long gone, its legacy is still going strong and very much not forgotten.
VFAB’s Gone But Not Forgotten Series
The story of key hardware and electronics companies that at one point were leaders and pioneers in the tech industry, but are now defunct. We cover the most prominent part of their history, innovations, successes and controversies.
Masthead credit: Dependency Injection