facebook rss twitter

Review: AMD Athlon 64 Model 3000+

by Tarinder Sandhu on 5 March 2004, 00:00

Tags: AMD (NYSE:AMD)

Quick Link: HEXUS.net/qawn

Add to My Vault: x

What is it?

We'll turn to our usual table format to highlight the differences between the 3000+ model and others in the range.

CPU AMD Athlon 64  Model 3000+ AMD Athlon 64  Model 3400+ Pentium 4 3.4GHz Prescott Pentium 4 3.4GHz Northwood
Clock speed 2000MHz 2200MHz 3400MHz 3400MHz
L1 cache 128kb 128kb 20kb 20kb
L2 cache 512kb 1024kb 1024kb 512kb
Memory bandwidth 3.2GB/s (DDR400) 3.2GB/s (DDR400) 6.4GB/s (i875P) 6.4GB/s (i875P)
FSB 2000MHz (core speed) 2200MHz (core speed) 200MHz double pumped 200MHz quad pumped
Integer pipeline length 12 12 32 20
CPU Die Size 193mm² 193mm² 112mm² 131mm²
Transistor count 105.9 million 105.9mm² 125 million 55 million
Manufacturing process 0.13-micron SOI 0.13-micron SOI 0.13-micron 0.13-micron
Memory support DDR400 Single Channel DDR400 Single Channel DDR400 DC DDR400 Dual Channel
OS Support 32/64-bit 32/64-bit 32-bit 32-bit
Voltage 1.5v 1.5v 1.65v 1.525 - 1.55v
Form Factor 754 754 462 478


You may also realise that the Model 3000+ runs at exactly the same clock speed as the 3200+ model, namely 2GHz. So why is it labelled as a supposedly inferior CPU?, Take a look at the differences between the characteristics between the current range-topping Model 3400+ and the 3000+. Save for clockspeed the major difference is in the amount of L2 cache, which is halved from the Models 3200+ and 3400+ 1024kb to 512kb. Other than that, there's no difference as far as we can tell. The Model 3000+ benefits from the Athlon 64 S754's on-die memory controller, albeit single channel, HyperTransport link to the chipset, Silicon-On-Insulator (SOI) Technology, SSE2 support, AMD64's Industry Standard Architecture, and, of course, 64-bit support.

The question we pose ourselves is why AMD didn't decide to keep to the Model 3200+ and 3400+'s architecture to the letter and simply reduce clock speed to, say, 1.8GHz?. That would tie-in with the model numbering scheme. A clue is with the Model 3000+'s die size. AMD reckons it to be the same as the 3400+'s, which, as mentioned, is a staggering 193mm². Reasoning tells us that a large proportion of that die is used to house the millions of transistors needed for 1024kb of L2 cache, so reducing this number to half (512kb) will bring about a smaller die size. Smaller die sizes translates to more dies per wafer and lower costs per given die.

Our reasoning concludes that a 'true' Athlon 64 CPU with 512kb of cache should occupy a die size that's considerably less than 150mm². Yet AMD is adamant that all Athlon 64 Clawhammers share a basic manufacturing process. The simple explanation seems to lie with the fact that AMD disables half the L2 cache from a standard Model 3200+ CPU and rebrands it as a Model 3000+. It makes sense if you carefully consider the efficiency of the manufacturing process. It's not possible to have consistently perfect wafers every time. Those that fail initial testing often do so for L2 cache reasons. AMD needs to ensure that the whole 1024kb of L2 cache can run at the processors internal clock speed. That's millions and millions of transistors that have to pass at 2GHz and above, without fail. Defects are more likely to occur in this area than anywhere else on the die, so AMD has decided, it seems, to re-use dies that have been marked as semi-cache failures. They can't be used in the 3200+ and 3400+ models, evidently, but enough can be salvaged for 512kb usage. That's all the Model 3000+ is, that is, a die originally destined for a Model 3200/3400+ CPU that didn't make the grade for L2 cache reasons. Disable half of the cache-related transistors, keep the clock speed the same, and, voila, a new CPU is born. AMD isn't new to this. The AthlonXP Thorton core follows the same die recycling principles, and ATI has undertaken a similar approach with the 9800SE and 9800Pro VPUs.

The recylcing of select dies lets AMD produce a cheaper CPU that's still based on the performance Athlon 64 core. The question is just how much negative performance impact with half the cache have?. The answer largely depends on how cache- and memory-dependant a particular activity is. If it isn't, you've literally lost no performance when compared to a Model 3200+ CPU.