The best of the four STREAM benchmark results for a variety of workstations and servers is presented in Figures 4 and 5.
Figure 4: Part 1: Sustainable memory bandwidth for some non-vector uniprocessors, generally servers and workstations. The value shown is the maximum of the four STREAM tests.
Figure 5: Part 2: Sustainable memory bandwidth for some non-vector uniprocessors, generally servers and workstations. The value shown is the maximum of the four STREAM tests. The IBM RS/6000-990 is retained to fix the scale to match that of Part 1.
Note that the scale is comparable to that use for the non-vector shared-memory machines, but that the ``typical'' values are somewhat lower than the single cpu versions of the shared memory machines. This is primarily due to age and cost considerations. The data in the table includes a large number of older workstations, while the shared-memory data is mostly for fairly new machines.
The obvious feature of the data is the very high performance of some of the IBM machines (notably those with the Power2 processor), and the remarkably similar performance of the rest of the field. Two other anomalies are the HP J-200 and the DEC AlphaStation 600-5/300. Both of these employ new cpus that allow significant overlap of memory traffic and cache miss latencies, though the implementation details are complex in both cases.
Some attempts to model these results will be presented in the Discussion.