NexGen
From The Matrix
NexGen
Atiq Raza was born in Pakistan, in the family of a talented self-tought radio engineer and this had a great effect on its life - on finishing the school he left for London to master electronic engineering at a college. On his graduation in 1972 he came back to his motherland which wasn't though an ideal place to build up a hi-tech career. He found a job in Telephone Industries of Pakistan and worked there for 6 years watching how his country was turning into a scene of action for religious fanatics and armed gangsters (or even both in one). As a result, at the age of 30 Raza with his wife and child migrated to the US in 1978 where he enrolled first in the University of Oregon and then later Stanford University. Once again it was a radio electronics department.
His further career went through different companies - Synergetics, Trilogy, VLSI Technology. At the same time he was involved in public affairs working for a local community as he felt obliged to give something back to this country: it wasn't even a Pakistani community - when English people left India long time ago they divided it into two states - Pakistan and India, and Atiq tried to overcome that split, though from the USA: "We also created TiE (The IndUS Entrepreneurs) organisation with Kanwal Rekhi, Suhas Patil and and Prabkhu Goel taking an active role. I wasn't that active but I emphasized there should be no dividing line between Indians and Pakistanis. Whenever a Pakistani came to me, I told them also we should remain completely united with Indian organisations and that is the way I have operated for the last 22 years in the United States".
No wonder that the "continental" line went through the Raza's career in the USA. And in 1988 his acquaintance, Rajvir Singh, suggested he work for a newly created Indian startup called NexGen founded by Thampy Thomas and financed by Compaq and ASCII (there was about 20-30% of Indians and Pakistanis among the engineers). Thomas had a trivial plan - to make clones of Intel's i836 and PCs around it. That was a real adventure - as Vinod Khosla from Kleiner Perkins Caufield & Byers recalls later - when that investment company joined financing NexGen in 1990 (right at the time when Raza became CEO), they even lacked for money for current weekly payments.
NexGen had no agreements with Intel, that is why, in contrast to pirates from AMD, they had to build up the processor from the very beginning. The work took 7 years, and the first announced in 1994 solution was fairly interesting. It was really fully compatible with the i386 in the instructions, but its microarchitecture outshined even the 486 and resembled more the Pentium which just appeared on the shelves.
Those 7 years were very fruitful - although the engineers didn't keep within the established terms, the Nx586 was the first, after Pentium, superscaler processor, i.e. it was able to process more than one instruction at a clock. Today a couple of integer units, those for floating point operations, SIMD and so on are no wonder, but at that time a second integer unit was a top-range progress. Add one more technological exploit here: the Pentium has an L1 cache divided in two - 8 KB both for instructions and for data, the so-called Harvard architecture. The Nx586 used a similar solution, just the cache was twice as great - 16 KB for each part.
Plus, the dynamic branch prediction, the 64-bit FSB (against 32 bits of the 486), and the RISC86 architecture. The technology of splitting of complex CISC x86 instructions into simpler ones, RISC-like, which is so popular today, was first tried at the time of Pentium and Nx586. Higher-performance execution units, smaller and more compact, easier addition of new ones. Superscalar support and speculative execution are more points they differ in: the processor execution buffer could simultaneously have three such RISC-like microops, one execution unit for loading/storing addresses and two integer ones, one quite complicated, the other very simple. Besides, it had an integrated L2 cache, contrary to the Pentium.
The NexGen didn't have its own factories, that is why the company covered the same way as Transmeta several years later. Ironically, as a partner for chip production they chose the same firm - IBM. At the beginning of March 1994 they released a 0.5micron version of the 70 MHz processor, on March 10 1995 it was a 0.44 micron solution. 3.5M transistors, 4V core voltage, around 160 square millimeters. The FSB frequency was twice lower than the processor's one - i.e. 35 MHz for the 70MHz processor. In November 1995 they came up with the Nx587 coprocessor. Such a small number of transistors was achieved by switching over the coprocessor onto an external chip. At that time NexGen considered (Intel also admitted that, though started strengthening a FPU in the Pentium with an eye into future) that the х86 code use mostly integer calculations, that is why they didn't imbed the processor into the main chip; instead, they used transistors, for example, to reinforce the L1 cache and embedded the L2 cache controller. But a little bit later, together with the Nx587, the company launched a new version of the processor - Nx586fp which combined the processor and coprocessor; it makes me think that at the time of release of the Nx586 the coprocessor wasn't finished yet.
Like most innovations, the Nx586 didn't avoid some compatibility issues - it used its own Socket-463 (it's interesting that later, being AMD's CTO, Raza chose a 462-pin connector for the Athlon 462). That is why it shipped with its own chipsets - NexGen NxVL or NxPCI. That pushed away a possibility to succeed although it could easily fight against the Pentium. On November 13 1995 first official Nx586 started flooding the market; the P120 and P133 had a bit lower frequency than the PR rating, for example, the P120 was clocked at 111 MHz, but anyway, its integer performance was 1.5 times better than that of the Pentium 120. And as you remember, the Windows 95 deals primarily with integer values.
The processor and chipsets were supported by 7 board makers and 4 manufacturers of third PC echelon, and the price set was quite good - $447 for a 133 MHz processor (much cheaper than the Pentium 133), that is why the base created was excellent, and today we could have Intel and NexGen fighting against each other if it were not IBM that failed to provide sufficient production facilities for the Nx586.
No sooner had the processor put on the silicon than the developers started working on the Nx686. The works were started long before and the fruit appeared soon after the launch of the Nx586. In 1995, October 10, at the MicroProcessor Forum, NexGen showed off samples of its next processor fulfilling them under the Windows 95.
Well, there was something to boast of. The same package, socket and square (180 sq. mm. on the 0.44micron technology though the mass production required 0.35 micron) housed 6 M transistors against 3.5 M of the Nx586, including a 48 KB L1 cache (twice higher frequency as compared with the core, 16 KB for the code, 32 KB for data). Plus:
separate power supply, 3.3 V input, 2.5 V on the core; power consumption - 4 W; clock speed - 180 MHz; FSB frequency - 60 MHz (i.e. the multiplier was elevated from 2 to 3); L2 cache's controller supports up to 2 MB of the external cache at the same processor's frequency; by that time Intel launched its Pentium MMX, that is why the engineers had to add a unit controlling instructions working with multimedia data, to follow its policy of keeping pace with Intel. Isn't it a big step as compared to the Nx586? It's an absolutely modern and competitive processor. At the MPF'95 the developers mentioned its performance must have been twice as great in integer calculations as compared to the Pentium Pro running at the same speed and by 33% in floating point operations.
Atiq Raza, CEO NexGen, clearly realized that fighting against Intel meant to lose. There must be someone with a greater reserve, that is why while the engineers were working on the new product, the managers started talks with another company that desperately needed a new processor to set off against the Pentium as their K5 was just a clone of the good old Am486 in spite of being pin-compatible. So, NexGen entered into negotiations with AMD.
At the beginning of 1996 the companies established an agreement: NexGen was bought by AMD at $850 million. Money? Shortly before that event Intel was forced by the court to pay some $1 billion to AMD... Well, an elegant solution to turn Intel's money against them. All the stuff of NexGen poured into AMD. Raza was at the head of adapting the Nx686 for the needs of AMD, after that he was the prime driver of Athlon. Two former Intels' men, who changed it for NexGen, also took leading positions - Dana Krelle, Vice President of NexGen's Market Development, got a similar post in AMD, and former senior Intel Vice President, Vinod Dham, who joined NexGen in Spring 1995 became COO.
It wasn't an AMD's own processor again, it was still several years before the release of the Athlon, but with the NexGen's Nx686 named AMD K6, Andy Grove would have no reasons to call AMD "Milli-Vanelli", as the song would be pouring from their own house in spite of the music & lyrics being from strange hands.
(source: digital-daily.com)
