Dell added some more suds to the long-running Intel-AMD soap opera last Friday. After being an Intel-only shop for 22 years, Dell let it be known that they've decided to use AMD processors in some of their products. The company announced plans to use AMD Opterons to power their high-end four-socket PowerEdge servers. The move to AMD has long been awaited by both Dell customers and industry analysts.
Apparently Dell finally got the message, although at this point the use of Opterons in their four-socket servers is probably the most conservative thing the company could have done. Without HyperTransport technology or something equivalent, Intel will have a hard time beating AMD in four-socket and above configurations in the 64-bit x86 market. So in a sense, Dell was just filling a hole in its product lineup.
The Dell news came at a somewhat embarrassing time for Intel, as the chipmaker started to roll out the red carpet for Opteron's rival, the new high-end Woodcrest processor, which is due to be released next month. On Tuesday, Intel announced some impressive benchmark results for Woodcrest, now called the Dual-Core Intel Xeon processor 5100 series. Intel claims to have set world records for 20 benchmarks for dual-processor machines.
For the supercomputing crowd, the HPC benchmark numbers are of special interest. In particular, the Linpack results showed Woodcrest-based systems with more than twice the performance of systems using the Opteron 275. The Fluent Computational Fluid Dynamics (CFD) and the LS-Dyna Finite Element Analysis (FEA) benchmarks showed Woodcrest with modest performance leads over the Opteron. Take a look at the Intel-published HPC results at http://www.intel.com/performance/server/xeon/hpcapp.htm.
However, there a few problems with the numbers.
One is the inherent problem in Linpack itself. It's notorious for not reflecting real-world application performance. And some feel Linpack favors machines with lots of cache at the expense of other hardware features. But the results are rather interesting in one respect. Not only does Woodcrest beat the Opteron in Linpack, but on a processor-to-processor basis, all other x86, RISC and Itanium chips as well. Again, keep in mind, Linpack's not a real-world test.
The CFD and FEA benchmark results showed the Opteron at 79 percent and 86 percent, respectively, of Woodcrest's performance. The main problem here is that Intel made the comparison against the 2.2 GHz Opteron 275 not the newer, faster 2.6 GHz Opteron 285 model. Well, I suppose that's not really Intel's fault; the newer AMD chips probably haven't been benchmarked yet. But the fact remains that Intel compared its not-yet-released chip to AMD's older chips.
I have a feeling the Operton 285 would break even with Woodcrest for CFD and FEA — or come pretty close. Not good news for Intel. Woodcrest is already using 65nm technology, while the 285 is on 90nm. As soon as AMD goes to 65nm later this year, they're likely to leapfrog Intel once again.
Also, some folks at AMD think that the benchmark results don't reflect the latest compiler technology for their architecture. From what I hear, the company intends to do some reality correction pretty soon. I wouldn't be surprised if the engineers at AMD put in some extra hours over the next few days to get their benchmarks up to speed.
Intel ran into another distraction this week — itself. The pre-Core architecture Dempsey chip was quietly announced this week. Subsequently, a bunch of server makers (less quietly) announced support for it. The problem with the Dempsey processor, now called the Dual-Core Xeon 5000 series, is that it suffers from the same ailments as many of the pre-Core Intel architecture chips, namely not great performance/watt numbers. When compared to the new Woodcrest, Dempsey is exposed for what it is — the previous generation. The word is that Intel will try to position Dempsey as a less expensive alternative to Woodcrest, but many analysts think its days are already numbered.
Not to be out-benchmarked, SGI released some numbers of its own for the Altix platform. The company claimed record-breaking SPECjbb2005 results for the Altix 3700 Bx2 server running the BEA JRockit 5.0 JVM. The SPECjbb2005 benchmark measures Java application performance as business operations per second (bops). The Altix managed a record 1,828,349 bops during testing. That's a lotta bops!
As SGI tries to reinvent itself, one of its main goals is to establish itself in the enterprise market. It'll be interesting to see if the company can position their Altix platforms in this new domain. There's a lot of competition out there for enterprise server dollars and it's going to take more than some impressive benchmarks to gain a foothold. Stay tuned.
As always, comments about HPCwire are welcomed and encouraged. Write to me, Michael Feldman, at [email protected].