History of Overclocking Part Two

Athlon 64, Athlon 64 X2
(2003-present, 130nm-65nm process, 1800-3000MHz)

AMD’s response to the increasing Pentium 4 challenge was to get 64-bit to the desktop market first, but with 32-bit compatibility as well. In September 2003 the Athlon 64 came, and AMD architecturally diverged from Intel again.

The memory controller had been taken off the northbridge and incorporated into the CPU, which reduced latency remarkably. One unfortunate side effect was that it would also be subject to whatever voltage you fed the CPU – this wouldn’t be fixed until the release of Socket AM2+.

As such, the FSB no longer existed and was replaced by the HyperTransport bus (HTT). This originally ran at 200MHz, and was pumped by an HT multiplier (originally called the LDT, after HyperTransport’s original name of Lightning Data Transport) of up to 4x, giving an effective 800MHz bus. Since this was a DDR system, this would give an effective speed of 1600MHz.

The general rule of thumb among overclockers was to keep the expected HyperTransport frequency the same – so if you were meant to run on a HyperTransport frequency of 1GHz and needed to ramp the HTT, you dropped the HT multiplier accordingly. In fact it was recommended at first to stay under the recommended HyperTransport frequency due to stability issues with early motherboards. The nForce3 in particular kept the multiplier at 3x for this reason.

A heat spreader was introduced to the CPU so cores would no longer be easily crushed.

The nForce3 allowed the the PCI and AGP buses to finally be decoupled from the FSB, so they could run at a native 33MHz and 66MHz – although this would be slightly broken until its second generation. This same advantage would crop up in the Pentium 4.

The Athlon 64 allowed the CPU multiplier to be bumped down, meaning HTT could be ramped brutally within the limitations of your RAM.

To compete with Intel’s Extreme Edition the Athlon FX was released. The only difference from normal chips tended to be a slightly higher clock speed and an unlocked multiplier – although the first FX chip required buffered memory.

The Athlon 64 started on Socket 940 for the FX and Socket 754 for the desktop edition, but in remarkably Intel-like behaviour this was quickly shifted to Socket 939 (bringing dual channel memory and a HyperTransport bus increase from 800MHz to 1GHz courtesy of a 5x multiplier) six months later for desktop, while 754 was relegated to the new Sempron, essentially the replacement Duron. The FX would stay on 940 for a while, but this would be revised to use unbuffered memory – and would eventually transgress to the 939 platform. The nForce4 was released offering PCI Express, and as a consequence SLI
was reborn.

Tower HSFs became popular such as the Scythe Ninja, and would approach 1kg in weight – the 120mm fan significantly quieter than the howlers of old. Suddenly we were no longer willing to accept excruciatingly loud coolers in the name of overclocking – they had to be quiet too.

Venice and San Diego revisions were released at 90nm, which featured SSE3 (the original Athlon 64 supporting SSE2 only) and an overhauled memory controller.

The Athlon 64 X2 brought dual core to the masses, and the masses said ‘Thank you’.

In 2005 AMD released the X2 – the first dual core CPU. It was released at around the same price as high-end single core chips, while the single core prices dropped, assuring quick adoption. All things seemed rosy, and in 2006 AMD released Socket AM2 Athlon 64s, its major advantage being DDR2 support and virtualisation. This would prove to be underwhelming in the face of Intel’s mighty comeback, which started with a small processor called ‘Core’ on the notebook platform…

Pentium D
(2005-2006, 130-90nm process, 2660-3730MHz)

To fend off the X2 attack, Intel released the Pentium D, codenamed Smithfield. HyperThreading was disabled on all but the Extreme Editions. Initially released at slower clock speeds than the Pentium 4, the low-end Ds were heavy overclockers, the Prescott architecture they were born from finally showing maturity. This would be followed by Presler, using two Cedarmill cores, which also showed good overclocks.

Intel was finally starting to outperform AMD again in some areas, and the rumbling rumours of a chip name Conroe were starting to be felt.

Core 2
(2006-present, 65nm, 1060-2930MHz)

Conroe became Core 2 Duo, and Intel bit back hard. Clock for clock Core 2 seriously outperformed the Athlon 64, and saw the retirement of the thirteen-year old Pentium name. A complete redesign, the majority of Core 2 chips ran at 65W rather than the previous 130W, allowing for plenty of overclocking headroom in the heat stakes.

Core 2 operated on a 1066MHz bus, although an 800MHz budget edition was later released. Called the E4300, it happily reached Extreme Edition speeds of 3.0GHz from its original 1.8GHz by ramping the FSB. People started getting stable FSB overclocks close to 500MHz – that’s an effective 2000MHz bus speed – as the Intel and DDR2 platform reached maturity, in no small part thanks to the P965 chipset.

AMD bought ATI, in an effort to compete with Intel at the platform level. This would result in a big slow down as the companies attempted to integrate, ATI’s DirectX 10 part ‘R600’ being delayed many times.

NVIDIA and Intel just smiled.

At the same time AMD tried to counter with its ‘4x4’ concept, aimed at enthusiasts and would require FX chips. After a lot of FUD (Fear, Uncertainty and Doubt) it was revealed to be nothing more than a workstation board that didn’t require ECC memory, and was released eight months too late by a single vendor using the nForce 680a platform, to the fanfare of a single mouse squeaking. No one else noticed.

Before this NVIDIA released its 680i board, for the first time offering RAM speeds unlinked from the FSB – although this was simply implemented as a huge array of ratios hidden from the user. Still, this allowed greater overclocking flexibility than ever before, as RAM was no longer as crucial a component in the process.

Determined to press the advantage, Intel dropped prices rapidly and released quad core chips. The benefits were spurious to the home user, however they retained the overclocking capability of their Core 2 heritage. Not finished with digging the boot in, Intel announced its 45nm process, with no doubt even more overclocking headroom.

The future
Well there we are – a (not so) short history of overclocking up to this point in time. If you missed the first part, have a look here.

Now that you understand the development of the CPU, you can see what’s going on when you overclock and how the systems that have been implemented over time have matured to get us where we are.

Interestingly, methods of cooling haven’t changed much – they have simply become more accessible to the public. Air, water/coolant, peltier, phase change and LN2 cooling have all existed for quite some time – although admittedly they have been refined.

There are numerous other avenues that could have been pursued throughout the course of this article, but if we followed all traces, forum posts and offhand references we’d be here until the word ‘CPU’ had no meaning.

Which probably isn’t that far away, considering the inexorable march of general purpose GPU...


For all references used in this article, please check www.atomicmpc.com.au/overclockhistory.