AMD’s Phenom Architecture
There is quite a bit of information available today on AMD’s new Phenom architecture, but I thought I’d give a quick overview of it regardless. Those who are already well informed on the subject can skip ahead to the next section.
AMD’s K8 or “Hammer” architecture was released back in 2003. It introduced some very advanced new features that really set it apart from the competition. The Phenom continues to build on many of these features. Rather than give a complete overview of the Phenom architecture, I’ll be providing some details surrounding the changes from the previous generation, and some contrasting comparison to Intel’s lineup.
Integrated Memory Controller
In the K8, AMD introduced an on-die memory controller, which is perhaps one of its most unique features. This controller ran at the same frequency as the CPU core, providing very fast, low latency access to the system memory. In contrast, Intel processors have an off-die Northbridge that is part of the motherboard to handle all of the memory I/O. It will not be until Intel’s new Nehalem processors that are due out sometime later this year, that we’ll see an on-die memory controller from Intel.
AMD’s Phenom continues to rely on on-die memory control, but there are some important differences. Firstly, there are now two independent 64-bit memory controllers in Phenom, as opposed to a single 128 bit interface in the K8. This is significant because the processor can handle two memory requests simultaneously across the two controllers allowing for improved performance in multi-threaded applications. The Phenom can also work in a ‘Ganged’ configuration for a single 128 bit controller if desired.
Secondly, the memory controller no longer operates at the same frequency as the CPU core. Most Phenoms have their memory controller operating at 1800MHz or 2000MHz in their higher end models. Although this may impact performance slightly in comparison to the K8, it finally allows the Phenom to run its memory at a set frequency. When I first took a look at AMD’s new AM2 processors back in 2006, it was evident that the memory did not always run at its rated frequency—DDR2-800 running at 750MHz for example with some X2 processors. With an independent memory controller clock speed, the memory no longer needs to be run at an integer divisor of the CPU clock speed. I outlined the issue in more detail in the AM2 launch review for anyone interested.
PC2-8500 support has also been officially added for the Phenom. Buyers can take advantage of 1066MHz memory frequencies without having to overclock their PC.
Hyper Transport 3.0
AMD utilized ‘Hyper Transport’ for their K8 processors, a very high bandwidth system bus that connects the processor to the chipset. It ran at a frequency of 2000MHz and provided about 8GB/s of throughput. AMD has adopted version 3.0 of the Hyper Transport standard in the Phenom that provides it with 14GB/s of throughput at an operating frequency of 3600MHz. Phenom is also backwards compatible with Hyper Transport 2.0, allowing it to run in some older AM2 based motherboards.
Although Intel’s Core2 processors are incredibly fast, their aging FSB is much more congested than AMD’s implementation of Hyper Transport. Not only does Hyper Transport provide a faster interconnect, AMD’s on-die memory controller keeps all memory traffic off of this bus as well. Intel’s new Nehalem processors will use a similar, but proprietary technology in the future.
A Native Quad Core
Native Quad core has been a big marketing ace up AMD’s sleeve. AMD advertises the Phenom X4 processors as ‘real’ quad core CPUs. Keeping all four cores on a single die has some definite advantages like being able to have all cores share certain resources on the CPU—like memory controllers, HyperTransport connection and L3 cache. Intel’s current quad core CPUs, like the Q6600 for example, are literally a pair of dual core processors in a single package. Each of the two pairs can not access each others shared L2 cache.
This native design also allows AMD a great deal of flexibility in producing other non-quad core processors from the same Phenom silicon. AMD can easily ‘shut down’ one or more cores to produce Phenom X3 or even Phenom dual-core processors.
Cool’n’Quiet 2.0
AMD’s Cool’n’Quiet system has also been refined in the Phenom. Each of the four cores can have their frequency adjusted independently.
Other Efficiency Improvements
Aside from the interconnect improvements and native core design of the Phenom, various efficiency improvements have been made to make each core faster than an equally clocked Athlon 64. I’m not going to go into detail, but the Phenom now has wider internal data paths, improved prefetching and more efficient integer and float point units. There are also changes that will improve Phenom’s performance in virtualized environments over the Athlon 64.