Athlon Overclocking - The AfterBurner

NoWhere Man writes "Over at RB Computing (an AMD-only shop in Ottawa, Canada), they are distributing the AfterBurner, an Athlon Overclocking card, developped by Golden Fingers. It offers on-the-fly frequency and core voltage modifications, that is a reasonable alternative to building your own, as shown at Tom's Hardware Guide. "

Comments

Review

[Fjord]

Here's the Hard OCP review of the Afterburner (from the link).

Crippling the L2 for the sake of the core?

[DiningPhilosopher]

The cache divider is not controllable through the Athlon's edge connector. In some cases, higher overclocked speeds may be possible by changing the cache divider from its default setting of 1/2 to 1/3.

I'm no overclocking master, but are they suggesting you cut the L2 cache speed from 1/2 core to 1/3 core? Why on earth would you do that? Let's say your core frequency is 800 MHz, and your L2 runs at 400. If you overclock it to 900 but your L2 cache is only running at 300, surely you're getting worse performance overall than you were before...

Is this just another example of the blind worship of the almighty MHz? I think this is the first time I've seen anyone sacrifice performance for higher core processor frequencies...

Or have I just forgotten everything from my architecture class? :-)

Re:Crippling the L2 for the sake of the core?

[Wakko Warner]

According to Anandtech, who did a review of overclocked Athlons (including ones with the L2 cache switched to 1/3), you still get a performance boost if you up the MHz and change the L2 cache multiplier to a lower setting. Granted, it's not linear, but you still *do* see a significant boost. Here's the article -A.P.
--

"One World, one Web, one Program" - Microsoft promotional ad

SMP not what it's cracked up to be..

[HomerJ]

I jumped on the BP6 bandwagon when they were released with a couple of celeron366's at 458. BTW I got off-week celery's at a good price and hoped for the best for 504 and didn't work.

q3a supports SMP, only in WinNT. I bought q3a expecting SMP code to be in there for linux, but Carmack doesn't think it's nessecary to have SMP in linux. That rant is for another day though =) I grabbed Win2k early to get a natvie DirectX and SMP support. The only thing I saw was complicated aps taking 50% cpu power. I never got the voodoo 3 driver to work in WIn2k quite right, so I didn't benchmark q3a with SMP(the drivers from NT game palace is you must know)

Kernel compiles in linux only take a couple minues, but that's about the only real use I get out of SMP. Linux distributes the processes wekk enough, but I rarely do something that really taxes the machine, besides q3a and compiling kernels. And GCC is the only thing that uses both. About the only REAL benafit, is that I can run GCC with only one job, and it doesn't tie down the system.

With just about every OS now supporting SMP, including WIn2k, OS X, linux, etc., when will companies start writing apps that take advantage of it? Is Win9x holding SMP back because it doesn't support it?

Re:SMP not what it's cracked up to be..

[costas]

I will repeat this until I am soar, 'coz, well, it's a soar point with us in the Beowulf community: Linux SMP sucks. If you want good SMP performance, you're better of with NT or Solaris.

The big problem with Linux SMP, IMNSHO: NO CPU affinity. Which means roughly this: processes are rotated thru all available CPUs, instead of being assigned to one CPU and then being dynamically balanced (new jobs sent to the lightest-used CPU, when CPUs are imbalanced by some threshold %, move 1-2 smaller jobs that will balance them out).

What does this mean? well, CPU cache is practically useless. Makes all that dough spent on Xeons instead of Celerons seem wasted --and it is.

Don't get me wrong; I am all for Linux, and I am sure SMP will catch up pretty soon. But don't go spending $$$ on SMP machines expecting (n-1)*100% increase in performance.


engineers never lie; we just approximate the truth.

Re:Overclock Wars?

[Jeffrey Baker]

Actually the PIII is WAY overclockable. The PIII-550E can be clocked up to 825 MHz according to some reports. Consider that a PIII-550E costs $350 and an Athlon 800 costs $904. An 825 MHz x86 processor cannot even be purchased on the open market yet. Anand's has a pretty good report on overclocking the PIII.

-jwb

Overclocking, a Silicon Valley perspective

[Animats]

Classically, overclocking was a terrible idea, because speed was a part selection. In other words, parts were manufactured to run at the highest speed, then tested. If a part was available in 100MHz, 150MHz, and 200MHz, a part marked 100MHz had failed the acceptance test at 150 and 200. But typically, only a few gates on the chip failed at the higher speed, so the chip almost worked at the higher speeds. And if its temperature was kept well below the upper limit of the rated range, it might work consistently. But in general, overclocking meant a system with a substantially higher error rate.

As the fab for a given process became more mature, the defect level usually decreased. So at the beginning of a product cycle, you got more of the slow parts and fewer of the fast ones, and over time, more parts were produced with the higher speed ratings. Over time, then, the price of the high-speed parts declined.

Then Intel reinvented itself as a consumer products company, and started pricing ICs the way GM prices cars. In the auto world, a luxury car costs maybe 30% more to build than an economy car, but sells for perhaps 3x as much. Intel started doing this for processors, with advertising-promoted brands at different points in the speed spectrum. The interaction between this policy and the way fabs actually work resulted in some deliberately undermarked chips, and the rebirth of "overclocking" as a semi-respectable enterprise.

Then some distributors started shipping systems with overclocked CPUs. Some even printed fake part numbers on the chip package. This led to trouble. Intel may have lost some revenue, but worse, they were getting a reputation as an unreliable IC supplier. So they added holograms on chips, part ID info readable from software, and speed-checking (which is hard; CPU chips ordinarly lack an on-chip timebase.)

Today, IC fab yields are so good that the part-selection approach is rare. If parts are failing, the fab has a problem. CPU speed and model has become mostly a market positioning thing.

In the industrial computer world, underclocking is common; the temperature margins improve, and so does reliability.

At this point, Intel and AMD are competing so hard on speed and price that neither can afford to undermark. So overclocking is a marginal idea at best. Gamers are probably better off getting a new graphics board.