"Discover performance - discover AMD", that was the slogan which adorned every single AMD PR kit at Computex 2002 in Taiwan. And there really were several things to be discovered - in the end even the Hammer itself.
AMD had already shown its CPU called "Hammer" on the fringes of Intel's IDF at the end of 2001. It shall be launched as desktop processor (code name ClawHammer) and as well as server version (product name Opteron), ClawHammer will be launched at first at the end of 2002.
Apart from AMD (8000 series) also Ali (M1687), NVIDIA (nForce2), SiS (755) and VIA (K8 series) are producing chipsets for this CPU. But the Northbridge has not a great deal to do because the memory controller is integrated into the processor. Thanks to this widespread support for ClawHammer AMD did not manage to prevent a hardware manufacturer from allowing tecCHANNEL to take a look at the processor.
Test Hardware
Maybe AMD has beaten the Hammer drum a bit too loud at Computex and has raised great expectations. The Texas-based company may propably not be too happy about these early benchmark results. But we still decided to publish them. On the one hand we were allowed to use the system a whole hour on our own and could equip it with the test components. Neither while configurating the system nor while testing it any problems whatsoever emerged. The prototype hardware was running absolutely steadily.
On the other hand, AMD itself has been making a fuss about the new CPU for a whole year - at least if the Hammer should be launched by December. For customers it will be a completely new platform and therefore the purchase decision should be well considered. But it is much too early to make a decision by now because the tested devices were mere prototypes like those engineers use for developing components.
The graphics card we used was a GeForce3 with Detonator driver XP 28.32. We used 256 Mbytes DDR333 memory. The Timing was conservative 2.5-3-3 clocks, for all other BIOS issues we used Default settings. Processors we compared with the system have been tested in a configuration comparable with our current test configuration.
The Tested ClawHammer
As test mainboard we used an AMD's Solo 2 with 8000 chipset, the processor was a ClawHammer at 800 MHz. It was equipped with 64 Kbytes L1 cache each for instructions and data, the L2 cache came with 256 Kbytes. AMD vice president Richard Heye himself had announced the cache sizes at the first day of Computex.
The clock rate that was gauged with tecMem and WCPUID meets the result which was circulating at Computex on the quiet - and also this specification has already been reported by Heye to the British news site The Inquirer. If the lock against overclocking the Inquirer quotes has been a part of our test system we could not check, time was too short.
According to the software we had at our disposal it was a CPU "Stepping 0". But we can't truly be sure about this because Windows XP Professional, the test OS and the other programs can't possibly recognize the CPU. According to experience ClawHammer should experience at least one more Stepping until its launching.
Benchmarks
The processor and mainboard tested by tecCHANNEL were prototypes. With the given configuration the real performance of the series CPU can only be assumed, last but not least because of the low clock rate of the prototype. Moreover, we only could run our tests with 32-bit software. The test does not give any results about a performance with 64-bit software.
The test results only allow to draw conclusions if one compares the performance of ClawHammer at 800 MHz with the performance of Athlon MP at 800 MHz with 32-bit software. It's safe to say that ClawHammer will not be available in the version we tested.
Quake III Arena 1.17
We decided to publish those results of our tests which we thought would make sense, so we only give details but rounded results for Quake III Arena 1.17.
CPU | Clock Rate [MHz] | Memory | Quake [fps] |
|---|---|---|---|
Configuration: Windows XP Pro, GeForce3, 256 Mbytes memory. The Athlon MP was tested at 133 MHz FSB clock rate. The Pentium 4 is a version with Willamette core at 400 MHz FSB. | |||
ClawHammer | 800 | PC333 2.5-3-3 | 183 |
Athlon MP | 800 | PC333 2.5-3-3 | 130 |
Athlon MP | 800 | PC333 2.5-2-2 | 133 |
Athlon MP | 800 | PC333 2.0-2-2 | 135 |
Athlon MP | 1667 | PC333 2.5-3-3 | 210 |
Pentium 4 | 800 | PC800-45 | 111 |
Pentium 4 | 1600 | PC800-45 | 182 |
Memory Performance
Here you find a comparison of cache and memory transfer graphs which we detected with our benchmark tecMEM. We do not publish the exact results of this highly exactly benchmark, to us it does not make much sense because of the prototype hardware. Please consider the results as prospects for a possible performance of the internal memory controller of the ClawHammer.
The specifications in the graphs for 32 and 128 bit refer to the used 32-bit or respectively multi-media instruction set. tecMEM is a 32-bit software that additionally can use 64-bit MMX and 128-bit SSE instructions but does not use the x86-64 extensions of the Hammer processors.
You may download the tecCHANNEL Benchmark Suite for free here using this link.
32-bit Transfer
In the case of 32-bit Load, Store and Move and the optimized computer instructions LODSD, STOSD and MOVSD come into operation.
128-bit SSE Transfer
The 128-bit tests were carried out with the SSE instruction MOVDQA. With this instruction data is read respectively written from the source buffer and the target buffer tecMEM built up.
Conclusion
ClawHammer owes the prototype's impressing results primarily to its integrated memory controller. The clock rate seems to be AMD's main problem. For comparison: first prototypes of Intel's Pentium 4 worked at 900 MHz, and lateron Intel changed the test systems' clock rates to 1 GHZ. Not until its 1.4 GHz version Intel was able to mass-produce big numbers of units, and the 1.5 GHz version has been very rare for a long time.
It seems AMD has been very conservative in regard of its first Hammer processors at 800 MHz. They are optimized on stability for hardware and software developers and not really on performance results. An indication: While testing the system the aluminium heatsink with its relatively quiet fan didn't even get hand-hot. That says something about the low power consumption of the rather small die.
At Computex rumours were afloat that the clock rate will even be 1.6 GHz at the launching. Should this really come true and should Intel really cope with the aimed at 3 GHz, then AMD has to do much awareness training in regard of its model number -3400+ rumours say. The launching of AMD's first complete design in three years time seems to become a marketing problem much more than a technical problem - our first beats on the Hammer drum have been successful by all means.
Unfortunately AMD did not want to comment on the tecCHANNEL tests. What difference there is between the ClawHammer we tested and its series version, this questions still remains open to all speculations. (nie/cvi/hal/mec/mje/bmu)