Released in early 1998, the K6-2 CPUs were an enhancement of the original K6. Cache size remained the same, however the K6-2 saw the introduction of 3Dnow!, a SIMD instruction set akin to MMX. Some games and drivers (especially 3Dfx) were patched or updated to take advantage of 3Dnow!. Quite a few K6-2 models ran using a 100MHz FSB which gave a boost of 5%-10% depending on the game. K6-2 processors are used quite often in Super Socket 7 machines thanks to the unlocked multiplier.
The K6-III was released in 1999 , just a few months before the Athlon, although only in small volumes. It's only advantage over the K6-2 is the inclusion of full-speed L2 cache on the chip, where as K6 and K6-2 processors used motherboard cache as L2. This has the additional benefit of increasing the cacheable area. This is significant, as many Socket 7 and Super 7 motherboards could only cache 64MB or 128MB (only a few motherboards could cache more memory). When using a K6-III the motherboard cache acts as L3 cache. K6-III processors come in 2 different voltages, namely 2.2V and 2,4V. While the 2.2V parts produce only slightly more heat then the Pentium MMX parts do, the 2.4V chips ran rather hot (about 20W max for the 2.2V part and almost 30W max for the 2.4V part). In terms of performance, certain games such as MDK2 love the full-speed L2 cache and benefit from it, although most games don't benefit nearly as much. K6-III CPUs should generally be avoided and replaced with K6-2+/III+ CPUs (if the motherboard supports it) which are much cooler and clock higher.
K6-2+/K6-III+ CPUs were released in 2000 as cheap upgrades to owners of K6-2 and K6-III CPUs. They were manufactured at a smaller 0.180nm process thus consuming far less power. Despite the name, both processors feature full-speed ondie L2 cache, with the K6-2+ featuring 128KB and the K6-III+ 256KB. The K6-2+/III+ are perhaps the most favoured chips from the K6 line of CPUs, due to their low heat output and minimal power requirements, most of which allow them to clock as high as 600MHz, which generally is an easily attainable speed. At that speed, the performance of the K6-II+/III+ is around that of Pentium II/III 450 and Pentium II/III 500, depending on the game.
In 1999 AMD left Super 7 behind and created the Athlon processor for Slot A. This was a vast improvement when compared to it's earlier offerings and made AMD a genuine competitor to Intel, which started the famous "1GHz race". Later AMD replaced Slot A with Socket A around the time it introduced Athlon's second core, the Thunderbird. Thunderbird at first made use of a 200MHz FSB (using 100MHz SDRAM) but this was later raised to 266MHz (using 133MHz SDRAM). A budget version of Thunderbird was also made, called the Duron. The Duron is basically a Thunderbird with 3/4's of it's L2 cache cut away, yet it still remained competitive compared to Intel's Celeron's. Thunderbird went from 600MHz to 1400MHz.
When AMD introduced Thunderbird's successor, it had to compete with Intel's netburst CPU's, which inherently had much higher clock frequencies compared to Thunderbird. It was then that AMD reinstated the PR rating, starting with Thunderbird's successor, the Palomino.
The Palomino was basically an optimized Thunderbird (something which can be seen as the die is more square compared to the rectangular die of the Thunderbird CPU's) build in the same manufacturing process. Another difference is that Palomino CPU's were made with a plastic package instead of the purple ceramic package used by Thunderbird.
Both Thunderbird and Palomino produced a large amount of heat (sometimes in excess of 70W, compared to the 30W to 35W of Pentium 3) and needed beefy coolers in order to be kept cool and were relatively easy to overheat and burn out if the CPU cooler was either not installed properly or if it had moved while the entire system was in transit (like in the back of a car on a bumpy road).
The Palomino is otherwise better known as "Athlon XP" and, as mentioned before, used the PR rating in it's model number instead of using it's real clock speed like Thunderbird did. Palomino was also generally the first Athlon to use DDR memory instead of SDRAM (though both can work with either type of memory, it is best to use DDR memory instead of SDRAM as SDRAM created quite a memory bandwidth bottleneck). Palomino went from 1333MHz (1500+ rating) to 1733MHz (2100+ rating).
Palomino was succeeded by Thoroughbred, which was basically a die-shrink of Palomino and the processor die regained it's rectangular shape. Thoroughbred had some initial problems when AMD tried to scale Thoroughbred and a second revision of Thoroughbred was made to solve that problem successfully. Thoroughbred was also the first Athlon to reach a 333MHz FSB. Thoroughbred scaled from 1400MHz all the way to 2200MHz
Thoroughbred was again succeeded by the Barton core, which doubled the L2 cache from 256KB to 512KB, somewhat improving performance. Barton was also the first Athlon to reach a 400MHz FSB. And even though Barton was barly any faster then the fastest Thoroughbred which it replaced, it still performed better thanks to it's increased L2 cache.
The 2500+ Barton chip was very popular with overclockers as it was often possible to overclock the chip to 3200+ speeds.
Today: Try to get the fastest CPU that will run on your motherboard. Theres basically little difference in heat output between the various models of Socket A Athlons, but the performance difference between the older models and the newer ones is significant. The Thunderbirds are very easy to unlock though, one only needs a pencil. But there are also mobile Socket A Athlons which come unlocked in the first place and may consume somewhat less power then the desktop models. When considering building a Socket A system, one should keep in mind that Socket A Athlons draw a lot of power from the 5V line from the PSU, which is something modern PSU's are having trouble with.