For a product whose success determines whether AMD lives or dies, there's been surprisingly little said about Shanghai of late. Granted, there has been no shortage of semiconductor news, economic blues, various lawsuits, and the company's self-division to occupy the digital press, but when all is said and done, AMD's future rests significantly on Shanghai's ability to compete in the server market. Server processors typically carry much higher premiums than their desktop counterparts; the revenue-per-CPU that AMD derives in this market is extremely important to the company's bottom line. With Nehalem-EP on the way, and Shanghai now established and available, what's the consenus on the core?
To bench a server

The first rule of understanding server benchmarking is that it's almost entirely different from the desktop/workstation markets. There are some task overlaps, to be sure—3D rendering and HD video encoding come to mind—but even then, the scale and complexity of the tasks performed on a home system vs. a render farm are immense. Even after eight years of experience benchmarking desktops and workstations, I've still had to learn about server performance measurements from scratch. Server evaluation introduces multiple levels of variability into an already complex process. Server/enterprise benchmarks are built to an entirely different scale than a potential desktop counterpart. A desktop program capable of scaling up to octal-core effectively is a rare animal; a server application that can't scale past 64-cores is of limited use indeed.

The application mix for servers is quite a bit different from desktops; there's a notable difference between two socket and four socket purchases as well. Both of the images below are uwed, with permission; originals appeared on Anandtech.com




Building the infrastructure for proper server evaluation is no simple task; routers, storage solutions, and system interconnects must all be built and tested to ensure that the system is running at peak performance. The best analogy I can think of is a beehive. Testing the performance of a single desktop or workstation is equivalent to testing one particular bee, finding out what it is and isn't capable of, and how good it is at finding pollen. When we step up to server testing, the goal is to take snapshots of how the entire beehive performs as it goes about its day. Server tests, like their desktop counterparts, place different load stressors in different places.

In writing this article, I consulted reviews and articles written here at Ars, as well as those published at Tech Report, Anandtech, and TechRadar. Official results from AMD and Intel were also considered, though I put less emphasis on these for obvious reasons.

Performance Today: Shanghai vs. Harpertown/Dunnington

Tech Report and Anandtech have both performed extensive Shanghai vs. Xeon comparisons, but the two websites focused on different types of benchmarks with virtually no overlap between them. The differing results between Intel and AMD in what are supposedly the same benchmarks were interesting (if expected). In some cases, official manufacturer benchmark reports map well against individual reviews—but I wouldn't expect it.

If the phrase "AMD server products" conjures images of a sweating man in a seersucker suit frantically waving a few precious graphs where the company has eked out a win, it's time to reboot and reevaluate. The positive preview and launch data on Shanghai's performance and performance-per-watt has held true even under the most exacting microscopes that could be brought to bear. Shanghai doesn't win every test, but certain tests show the 2.7GHz quad-core ranked just behind Intel's six-core Dunnington. Overall, Shanghai performs best in true enterprise situations.

It's hard to find a benchmark where Shanghai entirely collapses. Even when it's decisively beaten by Intel's various Xeons (and this does happen), AMD's 45nm beauty counters with a strong power-per-watt ratio or high price/performance value. If Shanghai had launched a year early (displacing Barcelona entirely), we'd be talking about AMD's fabulous server positioning and how/when the company might introduce higher-clocked chips to gain ground in those tests where it still fell behind Intel's Xeon line.

Unfortunately for AMD, it's 2009, not 2008, and the light at the end of the Sunnyvale tunnel looks an awful lot like a train.
Nehalem-EP cometh

It hasn't actually launched yet, but everything points towards a 2S (octal-core) Nehalem-EP launch within a matter of weeks. Anandtech ran some previews of what a Nehalem vs. Shanghai matchup might look like, and the results are a punch in the gut. Johan benched a hypothetical Nehalem-EP (quad-core, 1S) against the Opteron 8384—that's AMD's current high-end baby at 2.7GHz. Even in its strongest tests, Shanghai is blown away by Nehalem; the performance gap between the two is far greater than anything AMD can match with clockspeed boosts.

These results actually mirror our own Phenom II vs. Nehalem match-up quite well. Even when we took a Phenom II up to 4.2GHz on freon, it could not decisively beat the Core i7-965 and was challenged at times by the Core i7-920. In that review, I noted that AMD would have to innovate their way out of the problem—clockspeed jumps alone were never going to give Shanghai/Deneb processors the grit they need to challenge Nehalem. What we see now is that the same problem exists in the server world—once Nehalem-EP launches.

At least one website, TechRadar, claims it has had the opportunity to benchmark one. Pre-launch benchmarks should always be taken with a double handful of salt, but other sources have confirmed the authenticity of these results. TechRadar describes their testbed as follows: "dual-socket Nehalem EP platform...based on essentially the same Tylersburg chipset as used with the first 'Bloomfield' Core i7 desktop chips. For our testing it was configured with 24GB of 1,066MHz DDR3 memory."

TechRadar "managed to squeeze in a single full run of arguably the most important test of processor performance - SPECfp_rate_base2006." According to the website, the Nehalem-EP test system scored 160 in the above SPEC test; they quote AMD scores in the same test as "just 105" and dual-socket Penryn 3.4GHz chips "fail to hit 90."

After seeing SPEC results, the folks at TechRadar also ran the Stars Euler3D benchmark. Again, the site claims extremely impressive results: "it's twice as quick as a pair of 2.7GHz Shanghai processors (14.34 seconds to complete five instances versus 30.32 seconds)...the platform's memory bandwidth is borderline biblical, clocking up 35GB/s in SiSoft Sandra's quick and dirty bandwidth benchmark."

AMD's response, the road ahead

AMD is not without response to Nehalem-EP's imminent appearance. The company recently demonstrated its own hexa-core Constantinople Istanbul processor. AMD's "probe filter" (now called HT Assist) technology isn't supposed to debut until the Maranello chipset (currently scheduled for 2010), but the Instanbul chips Sunnyvale showed off made use of the tech. The feature cuts down on the number of probe requests aimed at any given processor and dramatically boosted performance in certain benchmarks. Istanbul is also positioned as a drop-in upgrade for anyone currently using Shanghai or possibly even Barcelona systems.

Even with these enhancements in the pipeline, AMD admits that it won't be able to realistically match Nehalem's performance. The company's plan, therefore, is to fight back on upgrade potential, price/performance, performance-per-watt, and possibly price. That last is a road neither CPU manufacturer is going to want to go down. Intel will most certainly want to keep its Nehalem server parts priced at a premium level, while AMD needs every dollar it can get.

Based on available data, AMD's current position in the server market could be characterized as sunny skies today with thunderheads looming on the horizon. There's a significant chance those storm clouds could evolve into severe weather—but it's not a guarantee. With Shanghai, at least, AMD has some room to manuever—Barcelona gave it none.