I said at the time that the Opteron launch was essentially a server product launch, as there were no workstation motherboards. I also said in response to the article's headline that it was "an assured yes". It looks like that label can also be applied to Opteron's workstation performance as well.
In the November issue of Cadalyst Magazine, nine dual powered workstations were reviewed. Two of those were AMD Opteron-based, one was AMD Athlon MP powered, and the remaining six were Intel Xeon configured. For its roundup requirements, Cadalyst asked each vendor to send in a representative system that was equipped with a 3D OpenGL graphics card. So that wasn't a situation, which has been seen in so many reviews before, where a consumer level graphics card had been used instead. Cadalyst's review tests were run on systems that were properly configured for workstation use.
Cadalyst used various benchmarks for its workstation tests. For the AutoCAD tests, it ran its Cadalyst C2001 benchmark test using AutoCAD 2004 . The proe-01 (Pro Engineer) suite from SPEC ViewPerf v7 was used as well. MaxBench 4, a 3D Studio Max benchmark, was the other test suite used. All machines were configured with Windows XP Professional.
Cadalyst used a weighted index to derive its performance results. I don't know how those numbers were arrived at, but I'm sure that benchmark performance and system cost weighed in very heavily.
The overall performance winner was @Xi Computer's Xi MTower 2P64 workstation, which was Opteron-based. Its lead over the second placed Polywell Xeon system was almost 7%. The other Opteron-based workstation, which was supplied by Monarch, came in a close third. After fourth place, the rest of the field were also-rans. The Athlon MP system came in a very poor last, but it was the only system that didn't support AGP 8x graphics. What that last result shows though is the huge performance improvement that Opteron brings to the table.
Looking at the winning system's performance in detail: For the AutoCAD results, which were the highest that Cadalyst had ever seen, its lead was about 10% better than its nearest competitor. The MaxBench 4 results were marginally better than the other systems, but the ViewPerf proe-01 results were worse, but not to a significant degree. The winning system also scored five straight A's for performance, cost, warranty, expandability, and return policy. When all those attributes were added together, Cadalyst could only give @Xi's top performing MTower 2P64 its highly recommended award.
Even memory handicaps couldn't prevent the
Opteron system from winning
It should be pointed out that the winning Opteron system was somewhat hobbled. It currently isn't the best Opteron workstation system available. The victorious system was configured with an Arima HDAMB motherboard (4 DIMM slots), which only had one of its two processors directly connected to memory. The other processor connects to memory via its twin sibling. This is called a 4+0 setup, as the second processor doesn't have its own dedicated memory. So that not only cuts the memory bandwidth in half, it also adds undesired memory latency that the second processor has to live with. But even with those handicaps, which many have criticized on message boards, it was still the best performing workstation.
Tyan's K8W (2885) dual Opteron workstation motherboard connects both processors directly to memory - four DIMM slots per processor (4+4). On @Xi's MTower 2P64 configuration page, that board is now available as a $90 upgrade option. So it would be interesting to see how much better that system would perform when the memory constraints have been removed.
Registered DDR400 EEC memory will soon enter the market in volume. So when Opteron gains official support, a 20% boost in theoretical memory bandwidth is going to stretch Opteron's legs even further.
Xeon technology still stuck in the dark ages
Going back to my earlier piece, I said at the time, "Xeon is hemorrhaging performance because its northbridge-based architecture is out of date. Opteron's ability to scale has made today's Xeon yesterday's technology."
Today's up to two-way Xeon platforms are limited to using DDR266 memory. The up to four-way Xeon MP is hobbled with the slower 200 MHz RAM. When DDR400 Opteron support is officially confirmed, the Xeon platform will be up to three memory speed grades behind.
Six months ago in April, Intel blessed its P4 with a phenomenal 800 MHz FSB. Today's up to two-way Xeon is still saddled with the slower, previous bus (533 MHz), and the Xeon MP has never graduated beyond the original 400 MHz bus.
So why are Intel's dual processor platforms still stuck with slower FSB and memory technologies? Are faster solutions just too hard to implement?
We shouldn't forget that Intel has lined up DDR2 for its next generation memory. But as reported here, a major RAM manufacturer suggests that DDR2 would offer little performance gain, but would serve as a product differentiator for Intel.
If that should prove to be the case, and Intel doesn't fundamentally change its current design philosophy, how can it ever hope to make inroads into AMD's technological lead.
If Intel wants parity in the bandwidth stakes, it will have to follow AMD's example and incorporate an on-die memory controller and something equivalent to the HyperTransport bus. If the chip giant continues down the road of the past, no amount of advertising will compensate for poor performing designs.
When one considers Prescott's alleged thermal and FSB problems, because of apparent 90 nm transition issues, AMD appears to be holding all the technological aces:
1. SOI technology for lower power devices - less heat from AMD's process.
2. Onboard memory controller - no FSB issues to worry about.
3. HyperTransport interconnect technology - copious amounts of bandwidth and scalability.
4. AMD64 instruction set - AMD64 customers only need one platform to efficiently run both 32 and 64-bit code. With AMD64, you can say good bye to Heath Robinson computing.
Will Intel ever match AMD64? Intel produces the fastest frequency processors in the world, and the cache in those chips are probably the most dense in the semiconductor industry. Also, Intel arguably has the best manufacturing process in the known universe. But all those individual achievements don't matter a dime if the platform design can't fully exploit the best-of-breed technology.
When one looks at the AMD64 technology that has been brought together thus far, one has to say that the grand sum of it all puts AMD at least a generation ahead of Intel.
I've said this before, but I'll say it again:
"If Intel's 90 nm Netburst products don't pack anything more substantive than more cache and higher frequency, then from where I'm sitting, Opteron's server performance lead could be maintained for at least the next year and through the 90 nm generation."
What I said four months ago is still true today. And now that the Opteron platform has added workstation performance leadership as well, AMD's overall performance lead has been extended once again.
AMD64 technology has truly been combined into a carrier-grade architecture that fully projects its credentials. Intel's disparate technologies perform well in isolation, but when they're brought together as a whole, the total sum doesn't hold a candle to what AMD64 can and will deliver. µ