Last fall, AMD quietly started shipping Athlon 64 processors built on a 90nm process developed with IBM. Using its own product naming system, the 90nm processors top out with the Athlon 64 3500+, which clocks at 2.2GHz and contains 512KB of L2 cache. Codenamed “Winchester,” the 90nm 3500+ is a scant 84 mm2, compared to 144 mm2 using the older 130nm manufacturing process.
A smaller die size lets AMD build more CPU dies onto a single wafer, which theoretically reduces the cost to manufacture the CPU. It’s never quite that simple, of course, because the cost of developing the new manufacturing process needs to be amortized.
Another theoretical advantage should be lower thermal dissipation at the same clock speed. As we’ve seen with the Prescott-based Pentium 4 processors, this isn’t always the case. At clock rates exceeding 3.2GHz, the Prescott’s thermal output climbs in a sort of hockey-stick function, going from 84W at 3.2GHz to 115W at 3.6GHz. Interestingly, AMD has not tried to push their 90nm CPUs past the 3500+ (2.2GHz mark) to date.
With the 3500+, AMD has managed to reduce overall thermal output by a noticeable margin. Winchester’s voltage requirements have been reduced slightly, to 1.4v from 1.5v for the 130nm part. The rated nominal thermal output is 67W, versus 89W for 130nm. But does this reduction translate into real world advantage? We dug out our Fluke digital voltmeter and a thermocouple and decided to find out.
We’ve been fans of small form factor PCs for some time now, and the recent release of the
| Component | Brand / Model |
| Case / Motherboard: | Shuttle SN95G5 (Socket 939 Small Form Factor) with Nforce3 Ultra Chipset ( |
| Processor: | Athlon 64 3500+ at 2.2GHz (130nm and 90nm) ( |
| Memory: | Corsair XMS Pro Low Latency, 2 x 512MB ( |
| Hard drive: | Maxtor 160GB DiamondMax 9 Plus (ATA/133) ( |
| Optical Storage: | NEC 16x DVD+/- RW ( |
| Graphics: | Nvidia GeForce 6800 Standard, 66.93 drivers ( |
| Audio: | Sound Blaster Audigy 2 ( |
| OS : | Windows XP Professional ( |
We used a Fluke 189 digital multimeter with the TK80 thermocouple attachment to take temperature readings. We mounted the thermocouple onto the external top heat sink surface, which allowed us some idea of the temperature of the CPU, plus some ambient surrounding temperature in the case. The SN95G5 offers pretty good heat dissipation for such a restricted package, but can still get pretty warm when running a CPU and graphics intensive game.
The thermocouple was attached to the heat sink, in the lower trough between two center fins, and attached with a piece of thermal tape. We used Arctic Silver 5 between the copper heat sink surface and the heat spreader on the Athlon 64. We were also careful to use the same amount of heatsink paste and spread it evenly. In addition, the heat sink and CPU surfaces were cleaned with solvent when we swapped in the CPU. The 130nm version was tested first, then the 90nm product. The thermocouple wire was routed through vent holes in the SN95G5 case, and the case top was installed as it would be under normal use. The ambient air temperature was 68.5 degrees F.
We measured the temperature under three conditions:
- Running idle, in Windows, until temperature readings stabilize.
- Running 3DMark05 CPU tests, looping ad infinitum until a stable temperature is reached.
- Running the Prime95 mixed mode torture tests, until temperature stabilizes.
Continued…
Testing took the better part of a day, as stable temperature readings would typically take about 30 minutes to an hour to achieve, and we took two sets of readings and report the average. Let’s take a look at the results.
It’s clear that the 3500+ on 90nm runs cooler, as the difference in rated thermal output would indicate. Looking at the temperature chart, it’s tough to really see what it means, so let’s now consider the percentage difference.
The 3DMark05 test ostensibly measures CPU differences using game-like 3D scenes. However, even the low-res CPU tests use pixel shaders when running the tests, so the GPU kicks in. The Prime95 tests, on the other hand, are heavily CPU bound. So it’s natural that a larger difference would show up when running those tests. There’s even a minor, but repeatable, difference running at idle. (Note that we didn’t have AMD’s “Cool and Quiet” technology enabled.)
So the 90nm Athlon 64 3500+ runs cooler than the 130nm variant. At first blush, that shouldn’t be a surprise. But when we recall concerns about Intel’s thermal problems with Prescott, maybe it should be a surprise.
Recently, Ed Stroglio at Overclockers.com speculated that AMD realized that pushing the clock rate higher might result in the same sort of hockey-stick thermal performance as seen with Intel’s 90nm Pentium 4’s. As evidence, he points to IBM’s troubles in pushing up the clock rate of the most recent 90nm Power PC processors. AMD’s process is based on IBM’s technology, so perhaps it’s no surprise that AMD is holding back the clock rates. Their second-generation 90nm process is slated to come on line in early 2005. According to AMD’s own roadmap, high-end CPUs manufactured at 90nm should ship sometime in the first half of 2005.
Today, however, the
| Product: | AMD Athlon 64 3500+ at 2.2GHz (90nm version) |
| Web site: | www.amd.com |
| Pros: | Exceptional PC gaming performance for the price; 90nm version runs cooler |
| Cons: | Trails the equivalent Pentium 4 in most other applications; 512KB L2 cache |
| Summary: | If you want a great CPU for PC gaming at a reasonable price, the 3500+ is worth a look. Great for small form factor PCs |
| Price: | $265 |
| Score: | |
Keep up to date on CPUs in our