AMD's .09 micron A64 3500+: Overclocking, Thermals, & Power Consumption

Introduction & Test Methods

When Intel released the first batch of Pentium 4 processors based on their Prescott core, which were built using a .09 micron manufacturing process, analysts found that the CPUs generated more heat and consumed more power than similarly clocked Pentium 4 processors based on the .13 micron Northwood core. These findings flew in the face of tradition, as a die shrink usually yielded processors that would run at higher clock speeds and consume less power. But the Prescott core wasn't a simple die-shrunk Northwood.  The Prescott architecture consists of millions of more transistors than Northwood and it has a deeper pipeline. The combination of building a new core on a relatively new and unproven manufacturing process, and the increased power leakage that inevitably comes from shrinking the gaps between CPU traces are what resulted in Prescott's power hungry nature.

At about the same time that the news of Intel's difficulties with the .09  micron Prescott was spreading, IBM was contending with similar issues with the .09 micron PowerPC based G5 processors used in Apple's high-end Macintosh systems. Apple had claimed that the G5 would hit 3GHz within a year of its launch, but to this day IBM hasn't been able to deliver a CPU that runs reliably at over 2.5GHz for an extended period of time. And even at 2.5GHz Apple had to resort to using a custom water cooling solution to keep the CPU temperature under control. These developments led most people to believe that transitioning from .13 micron down to .09 micron was going to be troublesome for all chip manufacturers, regardless of architecture. But we still hadn't heard from AMD at that point.

AMD has since released some Athlon 64 processors built using their .09 micron Silicon On Insulator (SOI) manufacturing process. These CPUs are based on the Winchester core which is essentially nothing more than a die-shrunk Newcastle.  There were some rumors that AMD's first .09 micron processors would incorporate some enhancements to the memory controller and support for SSE3 instructions, but that doesn't seem to be the case. Winchester is just a "smaller" Newcastle.  We recently had the chance to test a .09 micron Winchester core based Athlon 64 3500+, so we decided to see how it matched up against a .13 micron Athlon 64 3500+ based on the Newcastle core.  We weren't interested in gaming or general performance though. Both of these processors perform at identical levels.  What we wanted to investigate was whether or not the Winchester core had a significant amount of overclocking headroom, and whether or not it ran hotter and consumed more power than a similar .13 micron CPU...

Specifications: Socket 939 Athlon 64 3500+
Shrinkin' it Down...
AMD64 - When utilizing the AMD64 Instruction Set Architecture, 64-bit mode is designed to offer:

_Support for 64-bit operating systems to provide full, transparent, and simultaneous 32-bit and 64-bit platform application multitasking.

_A physical address space that can support systems with up to one terabyte of installed RAM, shattering the 4 gigabyte RAM barrier present on all current x86 implementations.

_Sixteen 64-bit general-purpose integer registers that quadruple the general purpose register space available to applications and device drivers.

_Sixteen 128-bit XMM registers for enhanced multimedia performance to double the register space of any current SSE/SSE2 implementation.

Integrated DDR memory controller:
_Allows for a reduction in memory latency, thereby increasing overall system performance.

An advanced HyperTransport link:
_This feature dramatically improves the I/O bandwidth, enabling much faster access to peripherals such as hard drives, USB 2.0, and Gigabit Ethernet cards.

_HyperTransport technology enables higher performance due to a reduced I/O interface throttle.

Large level one (L1) and level 2 (L2) on-die cache:
_With 128 Kbytes of L1 cache and 512K of L2 cache, the AMD Athlon 64 processor is able to excel at performing matrix calculations on arrays.

_Programs that use intensive large matrix calculations will benefit from fitting the entire matrix in the L2 cache.

64-bit processing:
_A 64-bit address and data set enables the processor to process in the terabyte space.

_Many applications improve performance due to the removal of the 32-bit limitations.
Processor core clock-for-clock improvements:
_Including larger TLB (Translation Look-Aside Buffers) with reduced latencies and improved branch prediction through four times the number of bimodal counters in the global history counter, as compared to seventh-generation processors.

_These features drive improvements to the IPC, by delivering a more efficient pipeline for CPU-intensive applications.

_CPU-intensive games benefit from these core improvements.

_Introduction of the SSE2 instruction set, which along with support of 3DNow! Professional, (SSE and 3DNow! Enhanced) completes support for all industry standards.

_32-bit instruction set extensions.

Fab location:
_AMD's Fab 30 wafer fabrication facility in Dresden, Germany

Process Technology:
_.13 micron SOI (silicon-on-insulator) - Newcastle
_.09 micron SOI (silicon-on-insulator) - Winchester

Die Size:
_Newcastle Core - 144mm2
_Winchester - 84mm2

Transistor count:
_Newcastle Core - Approximately - 68.5 million
_Winchester Core - Approximately - 68.5 million

Nominal Voltage:
_1.50v (Newcastle)
_1.40v (Winchester)

Athlon 64


To compare the overclocking potential, thermal characteristics, and power consumption of our .09 micron - Winchester core based Athlon 64 3500+, to a .13 micron - Newcastle core based Athlon 43 3500+, we had to take some precautions and avoid a few pitfalls to ensure the comparisons would be legitimate.  First off, we tested each processor in the exact same test bed; only the CPU itself was swapped between tests. The room where the testing was conducted is climate controlled, and remained at a constant 70oF (21oC) throughout. We used the exact same heatsink and fan combo, and the same thermal compound on each CPU. And the fan mounted to the heatsink was connected directly to the system's PSU, to prevent it from thottling at lower temperatures. The tests were conducted with all components mounted in a mid-tower case, but with one side-panel removed.  The MSI K8N Neo2 Platinum motherboard we used for testing was updated to the latest BIOS (v1.4), all voltages and bus speeds were set manually, and Cool 'n Quiet was disabled.

Related content