Cadaveca offers the uninitiated a trip down Route 64 with his article which investigates AMD's 64-bit platform. There's a few overclocking tips to be had too!

A Trip Down Route 64 

A new era is upon us...and it's called 64.

A few have travelled down its path, and today we get to ride along.

What is 64? Why, an Athlon 64, of course. Although they have been on the market for quite some time, price has been a major factor in preventing people from "going 64". Recently,  prices have dropped dramatically, and new features and chipsets have been added. Not only that, but the whispers in the IT hallways are that Windows 64 is soon to hit shelves and Intel has responded it's own way.

The biggest attraction to Athlon 64 is what sets it apart from the previous generations of Athlon; hinted in it moniker itself: "64". The latest breed of Athlons brings us 64-bit computing, and a possible 100% performance boost from the current 32-bit standard.

Most of you already know all this, and some of you have even taken the plunge already, but what makes an A64 really different? Plain and simple, the memory controller has moved home from the northbridge to the CPU die. For overclockers, this is even doubly important, as in all seriousness...it's just something else to go wrong.

With that all said, let's take a look at each of the individual parts of the Athlon 64 processor and explore their overclocking abilities. Read on if you dare!

When AMD moved the memory controller from the northbridge to the CPU, they did not a little thing. Firstly, they changed the FSB to HTT. What is HTT? Going to google will give you numerous and many varied results, none of which you find on the first page will be the right answer. Not on page 2, either. But on page 3, we get this result:

"HyperTransport™ Consortium - Press Releases

Recent Press Releases. For more information about HyperTransport press releases and press related materials, please contact: Constance Sweeney - (650) 358-9119. ... www.hypertransport.org/consortium/cons_news.cfm - 25k - Cached - Similar pages"

HyperTransport is a whole different beast from the FSB. The HyperTransport Consortium says this:

"HyperTransport interconnect technology is a high-performance, high-speed, high-bandwidth, point-to-point link that provides the lowest possible latency for chip-to-chip links. HyperTransport technology provides a flexible, scalable interconnect architecture designed to reduce the number of buses within the system, provide a high-performance link for applications ranging from embedded systems, to personal computers and servers, to network equipment and supercomputers."

"HyperTransport technology's aggregrate bandwidth of 22.4GB/sec represents better than a 70-fold increase in data throughput over legacy PCI buses. While providing far greater bandwidth, HyperTransport technology complements legacy I/O standards like PCI as well as emerging technologies like PCI-X and PCI Express."

HyperTransport™ Technology Overview

So now we know what they call it, it's really the communication channel between the CPU and the northbridge, and will be labelled "LDT" in your bios. With the memory controller residing on the CPU die, passing info from memory to the other components in your system is much, much faster, as it no longer goes memory, northbridge, CPU, northbridge, components, but memory, CPU, northbridge, components. That's one whole step gone, and a large reason for AMD's strength in the gmaing market. What the technology poses for the future could be discussed quite thoroughly, probably taking up weeks and weeks, but all we want it is to overclock it. Lets get to it, shall we?

First, we start here, taken from the same page as the last info from the HyperTransport Consortium:

1.2-V Low-Voltage Differential Signaling (LVDS) with a 100-ohm differential impedance

So it starts off at 1.2v, but in the technology white papers, it says 1.14-1.26v max., and more specifically:

Measured at the external connection to the HyperTransport device package. The VLDT as measured on the die should maintain a 1.1V to 1.3V range under all conditions. This ±100mV variation at the die is considered when defining the DC output characteristics in this specification.

That means that the 1.25v setting for LDT that you most often see in your bios...just leave it. Even signalling up to 2.8gb/s, and as low as 400mb/s, all require the same voltage. Going faster may nessecitate more, but seeing how most chips now are only in the 2.0gb/s range, there's losts of headroom. There's even more signal loss at the hihgher frequencies, so i don't think that giving it any more will really make a difference.

Your chipset basically sets the base speed, and currently some of the Via chipsets allow up to 2.0gb/s, to coincide with the Wincester A64's, but most chips, either Clawhammer- or Newcastle-core based both say 1600mb/s, at least from AMD, so keep this numbers in mind for you ultimate goal. Some chips will go higher than what they are rated for, but it's best to be -/+ 50mb/s. And yes, i know it says mhz in your bios, but ignore that, and read the whitepapers located at:

HyperTransport Link Specification.

to get the real story.

But, thankfully the new A64 is in control of my system's RAM. Setting up the memory controller for the A64 is fairly easy. Most BIOS'es will have options to set up the memory according to the specifications from the manufacturer, this is because most manufacturers are aware that SPD settings sometimes do not work. Want to take the HTT higher? Lower the memory speed in the BIOS, generally with selections ranging from 200 to 400 MHz, in 200, 266, 333, and 400 MHz options. If you set it to 333 MHz, your memory should hit 200 MHz around 250 MHz HTT.

I myself have acouple of A64 systems up and running. and true to my nature, each has had an attempt at overclocking. In pure bang for your buck, it seems running memory and HTT at the same speed will get you the best results. Setting the memory controller to a slower spped tends to affect performance a little, because going from 333 MHz to 400 MHz with an HTT overclock wil overclock the memory controller as well. You can take my word for it or give it a try for yourself. Even very low latency ram does not affect it too much.

When choosing RAM for an A64 try to get some that is rated higher than PC3200. Even dropping a multi, and increasing the HTT so that you keep the CPU running at it's default speed is beneficial, 8x 250 MHz performs way better than 10x 200MHz.

With the HTT replacing the FSB, the terminology is different, but the technology is basically the same. The CPU itself still has a FSB, if you will, as the multiplier has to multiply something, doesn't it? The difference is that the new FSB, or HTT, can be set independantly of both the memory-CPU link and northbridge-CPU link speeds.

There are a few adventurous souls who have gone beyond the 300mhz HTT mark, which although it may seem a far way away, it is obtainable with some of the chipsets out there. Unfortunately, this speed is still limited by the motherboard that you use, although most of the current releases reach 285 with ease. However, like i mentioned earlier, you get the best performance by having memory that will run at the same speed. If you really want a reason as to why, it lays in the signal conversion from one speed at one CPU part to the other speed on the second CPU part. The conversion is naturally going to add some latency.

There are other factors that play a role in your overclocking, like what power supply you use, videocard, and miscellaneous other components, but how they affect an A64, in comparision to any other CPU, is no different. Most A64 chipsets out there feature an AGP/PCI lock, and because the issue of having such a thing has been well discussed in the past, I will not go over it again. THere is however, one more very important thing...and that's the actual CPU core that you end up with.

Currently, there are only 2 major cores out there, in a myriad of versions. The Newcastle features a 1600 MHz LDT link, and 512 KB of on-die memory. The Winchester features 2000 MHz LDT link, and the same on-die memory, 512 KB. The third core, the CLawhammer, features the same 1600 MHz LDT link of the Newcastle, but has double the cache, for a total of 1 MB on-die. THe Winchesters (90nm) are only available on the S939 platform, and tend to overclock better, as is par for newer chips. If you want the easiest overclock, Winchesters are the way to go.

The Newcastle can be found in both the S754 and S939 socket line-ups, and in a 0.5 or a 0.6 revision. The 0.6 revision features a 2000 MHz LDT to coincide with the Winchester and only differs from the 0.5 revision in this regard. For overclockers this means that you can take the newest revision a little higher that the previous, assuming you are willing to forget about the higher LDT speed. Deciding which revision to get is as simple as trying to get the newest possible chip, although it seems there are still quite a few 0.5 chips out there. The newest has been reported as an easy overclocker, although the one that I got my hands on seems far from it. How high you really get just seems to be luck of the draw, sometimes.

The ClawHammer used to be found largely in the S754 segment, but recently has been relegated to only the FX line. This means not only do you get the benefits of being able to address more memory with the 1 MB cache, but you also get an unlocked multiplier; in the case of the FX line, you can go both up and down, while other chips can only go down from thier stock multiplier. Being able to add 4 double-sided memory sticks does not seem to be as much of an issue with the Clawhammer, thanks to it's larger cache, and extra HTT link available (A64's feature one, the FX two, the Opterons three). But, that extra memory can also prevent this chip from going as high as some of it's counterparts, unless you go to the extreme high-end cooling.

So, just like last week we are left with the same question, which one to choose? Fortunately, it seems that once again budget is always going to be the deciding factor. However, there are many options either way, now that prices have fallen. Socket 939 offers the best upgrade path as well as possible future technologies that have yet to be released. S939 also brings dual channel RAM functionality and PCI-E which brings nVidia's SLi into the fray. So, really it is not any easy choice.

The 754 segment has been around for a while, and is possibly a bit better refined than 939, as it has had time to age. As well, some of the cores that have full 939 functionality are ending up in the 754 ranks, which could possibly mean higher clocks, with the extra features not in use. I went S754 myself, recently, and still manage to bench with the dual channel systems, and some of them a full 200 - 300 MHz faster. You can even find some of the S939 chipsets on S754 boards, which possibly adds even more of a performance boost.

Regardless of what you end up chosing, if you are making the upgrade from a socket A platform even going with the lowest A64 will garner a performance boost, albeit small. Load up Windows 64 and everything seems lightning fast. My recommendation would be S754, if you don't plan on any major upgrades or, like me, plan on a whole new system sometime in the near future. In the end the choice is up to you.

Dicuss further here.