AMD Bulldozer Overclocking Event - PAGE 1

Enthusiasts all over the world have long been waiting for AMD's Bulldozer. Their hopes have been dashed on more than one occasion; although the architecture has been in talks for quite a long time now, the first realistic release date announced was June of this year. It was then further delayed because, supposedly, the Bulldozer chips would not go as high in frequency as desired. More recent rumors had placed it's release this September 19th, one week from now.

AMD held a conference on August 31st in Austin, Texas, where reviewers like us were told that that Bulldozer would eventually be coming a bit later. While a date cannot be divulged, readers are reminded of a blog entry written by John Fruehe, Director of Server Product Marketing at AMD. The main news there is that the revenue shipments of the server parts codenamed Interlagos have begun. These consist of two quad-module, octo-core dies slapped together in the G34 package. When asked about the client processors in the blog comments section, Fruehe pointed to a Q4 release. While it might sound like another delay, note that at this point in time Q4 begins in less than three weeks. It's a bit understandable that AMD would begin shipping the server parts first,  as they are definitely more profitable. With Interlagos eating up two dies per processor, it's not helping the availability of other models.

There was some kind of debate between enthusiasts as to whether Zambezi should be seen as a quad-core with a technology concurrent to Intel's Hyper-Threading, or as a native octo-core as AMD has called it since the beginning. The matter has obviously been brought to the table during the presentation, and from the most logical point of view, it will be an octo-core. The consensus here is that a core is hardware that runs a thread. In the case of Intel's Hyper-Threading, each such structure runs only one thread at a time. It's the hardware storing the physical state of the processor that is duplicated, so two threads can be scheduled at the same time. The main exectution resources such as the integer pipelines are still single, so both threads will compete for their utilization. The advantage is seen when threads are often I/O bound; when one is waiting after the network for example, another can start to run with minimized delay since its state is already decoded, thus minimizing the wasted cycles.

Now if all the data is contained in cache, I/O is minimized so there is not much to gain from Hyper-Threading, performance wise. That's why in many supercomputers, the technology is actually disabled. The difference with Bulldozer is that all threads can run simultaneously. What may have confused some people is that Bulldozer is composed of a modular design, called a module, containing two cores by default. For more technical details, Neoseeker will have to hold on until launch.

The conference at AMD Austin was accompanied by a demonstration of Zambezi's overclocking capabilities, though the available components weren't actually production samples. Having attended the AMD Phenom II overclocking event at the same place, I was expecting pretty much what I saw back then;  lots of liquid nitrogen and maybe even liquid helium. It would depend on whether Bulldozer is plagued by a coldbug like the original Phenom, or entirely free of it like the Phenom II. It turns out the liquid helium was put to good use. :)

Enthusiast Overclocking

The overclocking demonstration consisted of three distinct steps with evolving cooling systems. The first test showed what a "typical" enthusiast can achieve with Bulldozer using an easy to use aftermarket cooling solution. In this case, the popular Antec Kühler H₂O was chosen. The system ran an Unigine benchmark in loop mode for the whole day, without any hiccup, at a wonderful frequency of 4816MHz. This is pure eight-core overclocking. The voltage was set to 1.5V, which is where people would stop for 24/7 overclocks on Phenom II, using similar cooling. Would this end up too high for the new manufacturing process? Further experiments show some interesting numbers.

With a phase change cooling unit keeping temperatures around -37^oC, Zambezi could be clocked at almost 5.9GHz! Admittedly there was no benchmark running when this speed was recorded, but still it ran all day long without exhibiting any signs of instability. It's the kind of frequency one should expect to reach for benchmarking purposes. For this test however, three modules out of four had been disabled in the UEFI, leaving the processor in a dual-core state. The voltage was set to 1.632V.

Now see what is possible when temperatures get colder still.