With the release of the Barton a couple months ago, a lot of people were anticipating great things with the L2 cache getting a bump up to 512KB. If you've been paying attention, you'll know that the performance was not all that dramatic over their previous Thoroughbred cores. Fact is, both are the same CPU except for the added cache. Both are 166FSB parts (the latter TBreds that is), based on a 0.13 micron copper fab process. Many had hoped for the jump to 200FSB, but that is only available with the more recent 3200+.
The Barton 2500+
Packaging is the same as the Thoroughbred, which means it's based on the OPGA, or organic pin grid array, packaging technology. The benefit over ceramics (used in the Thunderbirds) is two-fold as it lowers the impedance and the cost of the CPU.
As you can see above, the markings are for a 2500+ CPU. Despite the bump in Performance Rating, this CPU has a lower clock speed than the 2400+. The 2500+ runs at 11x166 (1.83GHz), whereas the 2400+ runs at 15x133 (2GHz). Those numbers are rounded up, but basically, the 2400+ is about 170MHz faster.
So why did AMD feel fit to add 100PR? Well, as we've seen with the P4 Northwood, the extra 256KB of cache they added (to go up to 512KB) made quite a difference in performance. According to tests in their labs, AMD is making similar claims that the extra cache merits an increase in PR. We'll be sure to prove or disprove these claims in our tests.
To be quite honest, I was originally quite anxious on what kind of overclocking success we'd have. We had a lot of positive results from our 2400+, and being at the bottom of the food chain for the Bartons, I was hoping for similar results this time around. Sadly, a lot of the overclocks I managed before were not attainable this time around. Here's what did and didn't work…
Out of the box, the Barton 2500+ is clocked at 11x166, which equals 1826MHz. The first thing I did was to jump to 12x166, which the CPU handled without any problems. Next up was 13x166, and this is where we experienced our first lockup. Upping the voltage on our Abit NF7-M to 1.85v added some stability, but after an hour or so of benchmarks, the system would freeze up again.
We pushed the vCore up to 1.9v and everything worked out fine. Keep in mind that we were using water cooling (courtesy of the Swiftech H20-8500), but a heavy duty air cooler such as a Swiftech MCX462+ or Thermalright SLK-800 and a huge fan will probably net similar results. At 14x166, our system would not boot, no matter what voltage was applied. 13.5 worked at 2.0v, but the system wouldn't run stable enough to do any testing. Increasing the voltage to 2.1v did not help at all. With the multiplier reaching the limit, it was time to start working on the FSB.
At the stock multiplier of 11, we got the FSB as high as 203. This was with our vCore at 1.9v. I did not run any benchmarks at this time, but the system seemed to be running fine. The FSB wouldn't go any higher at 2.0v, so we lowered the multiplier to 8. Surprisingly, I had even less success, with the multiplier topping off at 200. I went into Windows, and began messing around, only to find out that it'd crash whenever I started 3D Mark 2003. I went back to 11x203, and this time, I couldn't even get into Windows (!). I double checked our voltage, and yup, it was at a 1.9v vCore.
With that kick in the funsack figured out, I went for the highest CPU OC I could manage. It took some time figuring out the best multiplier and FSB settings, but in the end I settled on 12x200. The vCore is 1.95v, which is pretty much in the danger zone without the right cooling. Don't go trying this if you have a stock cooler.
So what are my thoughts on Ocing? Well, I know we had a little more success than some people, and less than others. Simply put, Ocing the Barton is a real crapshoot, depending on the CPU you get. I borrowed a friend's 2500+, and I couldn't get past 196FSB, no matter the multiplier and voltage. Another friend can do 13x210 at 2.1v, but he has a VapoChill setup. What bothers me is the Barton 2500+ wasn't as flexible as the TBred 2400+ in terms of OC options, and it certainly wasn't able to match the 2400+ in the maximum MHz OC department.