Alpha Floating Point Sinks

In our recent reviews of Alpha NT systems we tested several NT workstations with the Digital 21164PC processor using AIM Technology's Workstation Benchmark for Windows NT. The 533MHz Alpha systems scored far below the Dual 300MHz Pentium II systems. In fact, the Alpha's overall system performance was much more comparable to that of a Single 300MHz Pentium II system.

In search of the cause of this disparaging performance, I used the AIM Subsystem Benchmark for Windows NT to target CPU performance: in this case, floating-point performance.

533MHz Alpha 21164PC WorkstationBased upon published SPEC95 floating-point values, we expected to see the Alpha values at roughly twice the Pentium II values. What we found was floating-point performance at nearly one half the Pentium II values.

In the charts to the right, floating-point performance tests from AIM Certified Reports for double and single precision addition, division and multiplication are represented in thousands of operations / second.

The 300MHz Pentium II clearly exceeds the 533MHZ Alpha. Part of the explanation for this unexpected result may lie in AIM's benchmark methodology.



300MHz Pentium II Workstation
AIM believes that in practical application no single component ever stands alone. In the case of CPU performance, AIM ensures that the benchmark instructions do not reside entirely in on-CPU cache. This could explain some of the differences between the AIM results and the SPEC95 findings. However, if the Alpha is being hampered in some way, why is the Pentium II not? The problem could be in the CPU, the cache, the motherboard, the benchmarks or the O/S.

Discussion with an engineer at AIM revealed that poor floating-point performance is consistent with all Alpha NT workstation results published by AIM Technology. Not only with 21164PC processors, but with 21164 processors as well. This eliminates any cache-related issues. While the 21164PC has no on-CPU 2nd level cache, the 21164 does have a 96K 2nd level cache on the CPU, yet both are equally afflicted with the sinking performance.



AIM then directed me to their Certified Server results. The floating-point performance on an Alpha NT server does not follow this sluggish trend. This is intriguing when you consider that the source code for the floating-point subsystem tests is the same in AIM's Workstation Benchmark for Windows NT and AIM's Server Benchmark for Windows NT.

533MHz Alpha 21164 Server
The graphs to the right represent floating point performance consistent with all 533MHz Alpha NT servers and 300 MHz Pentium II NT servers. (AIM's Subsystem Benchmarks are single threaded, therefore no advantage is gained in these charts through multiple processors). There is almost no difference in the Pentium II performance over the workstation configuration, but the difference in the Alpha's performance is astonishing. Like the SPEC95 results, AIM's benchmark now shows the Alpha to have twice the floating-point power of the Pentium II.



300MHz Pentium II Server
The power of the Alpha is revealed in overall system performance as well. Single processor 533MHz Alpha NT servers reach comparable levels of performance to Dual 300MHz Pentium II NT servers with similar memory and peripheral configurations. This would seem to eliminate any problem with the CPU and/or the benchmark.

The majority of the differences in the O/S's (Windows NT server and Windows NT workstation) are network related. The chances of this being a Microsoft problem are pretty slim.

There is very little difference between a high-end Pentium II workstation and a Pentium II server. The technology of the motherboards, memory, and peripherals are very similar. Alpha servers and workstations, however, are a different breed. Every Alpha workstation tested in this study used an SX or LX series motherboard. The servers utilize an optimized AlphaServer system bus. All other things being equal, the Alpha processors are clearly being choked by the SX and LX series motherboards.