when Ritsuko times out

K5 (1996) – Major flop
K6 (1997) – Minor success
K7 (1999) – Success
K8 (2003) – Major success
K10 (2007) – Flop

AMD K5 (1996)
The AMD K5 was basically a 29000 (a rather successful RISC CPU, used in the Boeing 777, just to cite one example) with a x86 front end tacked onto it. The 29K logic was unmodified. It ran hot (spawning the ‘AMD overheats’ rumours that you can still find nowadays, even though they’re not true anymore), had issues getting up to speed (hence the PR rating), came out a year late, and was just slow. It was faster than the Cyrix 6×86 in floating point, but that really isn’t saying much. This CPU earned AMD its bad reputation. Luckily for AMD, all the FUD being spread about AMD nowadays is plain false, as they cleaned up their act with the K6 and beyond.

AMD K6 (1997)
AMD got the Nx686 when they bought Nexgen, and because they were going nowhere with their own inhouse K6 and they needed a solution to the horrible beating the K5 was taking (especially with the Pentium II around the corner), the Nx686 became the K6 we all know. The original K6 they had been developing before was, according to Wikipedia, anaemic. The K6 offered superior integer performance compared to the Pentium II, not as impressive FP performance (addressed by the oft-ignored 3DNow extensions), came out one month earlier than the PII, and unlike the PII, it worked with Socket 7, so the Pentium/MMX owners didn’t have to run out and buy a new Slot-1 motherboard. A few refreshes such as the K6-2 (3DNow) and K6-III (3DNow+) came out later, which later continued the K6’s tradition of beating Intel offerings up. The K6-III was especially potent when combined with a 1MB L3 cache (back then the L3 was on the motherboard), going head to head with the Pentium III Katmai (the Coppermine was out of its reach).

AMD K7 (1999)
The AMD Athlon was the bomb. Clock for clock, it was faster than the Pentium III (or should I say Pentium Pro ‘P6′ architecture) in every benchmark. That’s right, every benchmark. Q: When was the last time you saw AMD beating Intel’s offerings shitless? A: 1999. Quake 3 simply loved the Pentium 4, so the K8 doesn’t count. The fact that a certain Dirk Meyer came over from DEC’s Alpha design team (Alphas ran at 500MHz at a time when the fastest PII was at 300MHz, with superior IPC) to design the Athlon and its DDR FSB only added to its allure. For the first time, somebody other than Intel had the x87 FP performance crown. It also helped that Intel had some teething troubles with their process. Refreshed and refreshed again, AMD milked every last ounce of potential out of it, until the Pentium 4, now reaching ridiculous heights of 3.06GHz, threatened to take back the crown again. But that was a job for…

AMD K8 (2003)
The major changes: a K7 with an integrated memory controller and SSE2 support, and a modified platform layout that eschewed the bus in favour of serial interconnects. Everybody would cite its memory controller when looking at its off the chart results. Of course, there were many other improvements, but it seems that the IMC was responsible for all the great work this CPU was doing. SSE2 support was welcome, but then again, the Athlon’s x87 FP was very good already, so it looked like it wasn’t really needed after all. AMD probably made millions on this baby. There was a dual core variant, to which Intel tried pathetically to respond with the Pentium D. If you asked any guy who knew his shit, he would recommend AMD over Intel. Everybody would, and everybody bought into it. In the end, only the numbnuts who knew nothing would buy OEM computers with Pentium 4 stickers on the front, because that’s all they knew.
Still, there was a cloud on the horizon. There’s a certain Core Duo/Solo floating around in the mobile market… oh well, nothing to worry about as long as it stays in laptops, right? After all, AMD never did that well in the mobile segment. Anandtech did a quick comparison of the Athlon X2 versus a Core Duo in a 479->478 adapter, and concluded that Intel certainly had something big cooking.

AMD K10 (2007)
By the time the K10 arrived at the scene, the K8 had already been shamed countless times by Intel’s Conroe, its laurel wreath on the ground, shredded. Here’s the juicy part: the K10 couldn’t do anything about it, the reason being that it’s just a prettied up K8, which in turn was a dressed up K7, 8 years old already. There were lots of minor improvements, sure, but Intel’s tick tock strategy was going on full steam, while Barcelona suffered from a TLB error, the fix for which caused a major performance hit; Intel’s slight errata with the Conroe core could be fixed with an updated BIOS, at no performance hit. Then Penryn and Nehalem came around…

Let’s not go into the details as to why the K10 is slower. The K5 lost the battle because it was late and had heat+clock issues, all of which come from trying to squeeze too many transistors onto one chip with any one process. The K10 suffers from the exact same problems. 140W TDP Phenom, anyone? The puny L3 cache? Hard to overclock? Only when K10 made the transition into 45nm did they AMD enough space to afford a larger L3 cache, and a higher clockspeed. The difference here is that now there is no competing architecture to buy, and AMD had more than enough time to come up with a completely new architecture. And AMD seems to be pricing the K10 aggressively, which helps their prospects a little.

Hopefully AMD is hard at work on Bulldozer, because it absolutely needs to succeed. Nehalem is already miles ahead, and the desktop derivatives are coming out this year, while I hear that we should not expect Bulldozer until early 2011, which makes me worry for the company’s future.