BTC 72,807.00 +6.62%
ETH 2,134.55 +7.66%
S&P 500 6,869.50 +0.78%
Dow Jones 48,739.41 +0.49%
Nasdaq 22,807.48 +1.29%
VIX 21.15 -10.27%
EUR/USD 1.09 +0.15%
USD/JPY 149.50 -0.05%
Gold 5,151.60 +0.33%
Oil (WTI) 76.11 +1.94%
BTC 72,807.00 +6.62%
ETH 2,134.55 +7.66%
S&P 500 6,869.50 +0.78%
Dow Jones 48,739.41 +0.49%
Nasdaq 22,807.48 +1.29%
VIX 21.15 -10.27%
EUR/USD 1.09 +0.15%
USD/JPY 149.50 -0.05%
Gold 5,151.60 +0.33%
Oil (WTI) 76.11 +1.94%

AMD’s K6-III ‘Sharptooth’ debuted this week in 1999 with on-die L2 cache to savage the Intel Pentium II — it also held the line against the Pentium III

| 2 Min Read
AMD launched its first processors with on-die L2 cache this week in 1999, heralding the beginning of the Super Socket 7 era.

AMD’s K6-III ‘Sharptooth’ debuted this week in 1999 with on-die L2 cache to savage the Intel Pentium II — it also held the line against the Pentium III

Delidded K6-III
(Image credit: Fritzchens Fritz)

AMD launched its first processors with on-die L2 cache this week in 1999, heralding the beginning of the Super Socket 7 era. With its new K6-III CPUs, AMD presented a major architectural leap above the popular K6-2 line, and knocked the Intel Pentium II 450 off its fastest processor perch. Intel would quickly follow up with its first Pentium III Katmai CPUs, but the K6-III could still outclass its pricier foe in cache latency-sensitive apps.

The Sharptooth scene

The new AMD K6-III with on-die cache launched in 400 and 450 MHz SKUs on Feb 22, 1999. AMD positioned it as a pre-prepared answer to Intel’s upcoming Pentium III (Feb 26 launch). Reviews and comparisons from the time show that AMD’s gambit largely paid off – maintaining the red team’s price-performance competitive edge until the K7 Athlon series tech could trickle down to the mainstream.

AMD K6-III

(Image credit: AMD)

The notable change arriving in the K6-III was the L2 on-die cache, which replaced the slower ‘backside’ L2 cache used by its direct predecessor and rival Intel Pentium II/III chips.

Swipe to scroll horizontally
Key specs of the AMD K6 III chips were as follows:

Technology

K6 NextGen architecture

Instruction sets

x86, MMX, 3DNow!

Cores

1C / 1T

Node

0.25um

Clocks

400 and 450 MHz

Caches

64KB L1, 256KB L2 on-die

Super Socket 7 era

AMD’s open, inexpensive Socket 7 era ended on a high note with the Super Socket 7 motherboards. These offered features like a 100 MHz front side bus, AGP support to elevate them from their legacy roots, and more flexible voltage controls.

In 2026, good condition Super Socket 7 boards are prized by retro enthusiasts for their wide-ranging CPU support, spanning Intel Pentium P54 and P55C, AMD 5k86, AMD K5, AMD K6, AMD K6-2, and K6-III (and +) chips, as well as rarer x86 artifacts like the Cyrix MII, IDT WinChip 2, and Rise mP6. They also deliver excellent DOS compatibility, AGP support for legendary GPUs like the Voodoo3, TNT2, and others, plus flexibility with RAM (EDO, SDRAMM) and storage controllers.

AMD’s successor to the K6-III, the Athlon K7, arrived at 500 MHz later in the same year. However, the K6-III would remain available through 2003. That’s quite an overlap, but keeping the older chip/motherboards available for mainstream users, and the new Athlons and Socket A motherboards for high-end desktop and workstation users, helped smooth the transition for AMD.

Google Preferred Source

Follow Tom's Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds.

TOPICS
Mark Tyson
News Editor

Mark Tyson is a news editor at Tom's Hardware. He enjoys covering the full breadth of PC tech; from business and semiconductor design to products approaching the edge of reason.

  • wussupi83
    I'm pretty sure this is what the first PC my family went out and bought had in it. I can't remember which clock speed version it was. Back then games were more hit or miss with compatibility with your hardware and drivers so it couldn't run everything but when it all worked it was a beast.
    Reply
  • usertests
    I have one of the last APUs with 3DNow!
    Reply
  • democog
    Hmm looks like L1 and L2 cache sizes have not changed since that era, which looks kinda weird I may say :) There is only L3 being introduced in bigger sizes..
    Reply
  • abufrejoval
    I believe I pretty near had every leading edge CPU that would fit into socket 5 and 7, some with the adapter that allowed you to run socket 7 CPUs on a socket 5 mainboard. And since they came out with a new one every few months, that was quite a few. I can't really remember what I did with the older ones... each upgrade was relatively cheap, but they accumulated.

    Just like I had every CPU that would fit into socket 3, including the overdrive variant (and an overdrive CPU to go with it): my first Intel was a 80286 overclocked to 8 MHz (1 MHz 6502 before that).

    On one hand they were super cool, because just swapping out the CPU gave you an extra boost. Had a K5, might have gone directly to a K6-II, but with nearly every speed rank, up to 550MHz. And that was a bit of a bummer, because switching to the K6-III meant you had to trade clocks for the on-chip L2 cache (only 450MHz on the 250nm process size), something Intel needed a "CPU slot" for. My final K6-III+ (back to 550MHz), became a bit unstable at the speed, it was finally time to change.

    The 32-bit wide RAM had limited bandwidth, Northbridge chips RAM capacity limitations, and on the I/O side there was an 8-MHz 16-bit ISA bus, that went with those mainboards... Eventually socket 7 just became too many bottlenecks, especially since I started to play with GUIs, graphics accelerators, and virtual machines on VMware (1.0, which used code morphing, since hardware support came later). I believe 64MB RAM capacity was a hard limit with all my socket X boards, enough to run NT 3.51 with a Linux VM or vice versa, but not for long.

    I sat out the whole slotted CPU era (where both Intel and AMD used CPU boards to fit SRAM caches), and then went to socket 754 and 939 very quickly to leave all those tiny bottlenecks behind. EISA, VESA local bus and AGP had been a real pain, micro channel obviously worse, PCI ok but PCIe finally so much better, even if RAM became a complete write-off far too frequently, considering the price.

    Since the Pentium Overdrive at 83MHz they all needed a fan (unlike all earlier CPUs from 6502 to 80486DX4), but those were still tiny, nothing you'd have to budget your power supply for, or select a chassis for aero dynamic qualities. In fact systems back then often had so many add-in cards and cables, they were an authentic jungle, but heat was rarely an issue... at least after those first-generation BiCMOS Pentiums at 800nm, which tended to unsolder themselves at 66MHz and 5 Volt power.

    Looking back it seems I was under the hood of my PCs pretty near every day, while today some systems may not be openend for months and then perhaps more to clean out dust than swap out parts. But I also had mostly one, perhaps two, not dozens of physical systems in the home-lab/family-home.
    Reply
  • razor512
    I have one of their ancient K6 II CPUs, I took a pic of it. While an extremely low 533MHz, it handled windows 98 pretty decently back then, though they really skimped on the cache back then.

    https://i.imgur.com/PJRKIZm.jpeg
    Reply
  • vinay2070
    I wanted to upgrade my pentium 100 to a K6 2 3D now, but being a school kid and lack of funds, I had to hold on till Duron came out. Duron was a nice upgrade along with a color monitor :)
    Reply
  • Ogotai
    i still have a K6 2 and k6 3, and both work just fine still... too bad i dont have room to keep them usable, currently sitting in the mobo box...
    Reply
  • thesyndrome
    Once upon a time a company could just decide they were going to clone a popular CPU that would work in the same socket, and they got away with a slap on the wrist and went on to become one of the biggest names in PC hardware.

    Now we are stuck in monopolies of what might as well be the hardware equivalent of a "two-party system", with trademark and copyright rules so strict that if anyone even ATTEMPTED the same in the modern age, they would be demonised by the industry and financially ruined by legal fees.
    Reply
  • abufrejoval
    razor512 said:
    I have one of their ancient K6 II CPUs, I took a pic of it. While an extremely low 533MHz, it handled windows 98 pretty decently back then, though they really skimped on the cache back then.

    https://i.imgur.com/PJRKIZm.jpeg
    533MHz isn't slow, it is quite near top clock for the chip (550MHz). And please remember that the 32-bit x86 era started at 16 MHz on an 80386 in 1985, already enough to punch out quite a few multi-user VAXes.

    With quite a few IPC and ISA enhancements and caches to take advantage of these, after little more than a decade it was probably 100x faster than the initial class-defining CPU on anything intenger and logic.

    Add floating point, especially the AMD vector variant instead of the obsolete 8-bit "optimized" 287 stack ops from Intel, and you can add two extra orders of magnitude!

    On single threaded integer code things might have improved only 20x nearly 30 years later. Pumping out 512 bit vector results on 192 cores at every clock cycle at 10x clock, obviously yields a bit more HPC, but does little for Word on Windows 95.

    And it was the on-die L2 cache the K6-3 added only a year later, much like the V-cache more recently, which helpled with many, but not all workloads.

    Calling it "skimping" is had to swallow: those were made in 250nm, not 2.5nm and that's 100X on both the X and the Y axis for a factor of 1:10,000 in density or simply available transistors: please, show some respect!
    Reply
  • abufrejoval
    democog said:
    Hmm looks like L1 and L2 cache sizes have not changed since that era, which looks kinda weird I may say :) There is only L3 being introduced in bigger sizes..
    Printed on paper, the CPU design literature on that topic might easily balance a sky scraper. It's only weird, until you've started reading, but the simple truth is: cache costs more than just transistors for the data, it needs to be found, which requires searching and management, which requires time, not just extra transistors for that logic. The bigger the cache, the higher the processing overhead, and no, that's not just linear, that's the sky-scraper part, trying to fight the diminishing returns.

    Have a go, find the easy video lectures, lean back and enjoy the hard work which made modern performance levels possible. In those "CPU wars", far fewer people ever died.
    Reply

Comments

Please sign in to comment.
Rampagefang Market Intelligence