Sunnyvale (CA) – The introduction of AMD’s quad-core Opteron processor with the 45 nm Shanghai core may be five months late, but the CPU builds on hopes of an AMD that has recovered from a disastrous TLB bug that was discovered way too late in the first generation 65 nm quad-core processor. At least from today’s perspective, Shanghai looks like a solid improvement, even if benchmark numbers may raise some doubt, and we wonder: Is this the quad-core CPU Barcelona should have been?   

A little over a year ago, AMD introduced the Barcelona quad-core Opteron processor with much fanfare at a fancy event in San Francisco – and ended up with a painful bug that delayed the delivery of first Barcelona server systems with a “fixed” processor by eight months. This time around, the new generation - the 45 nm Shanghai CPU - is launched in a much more subdued environment that not only fits to the AMD as we know it today, but it is probably also the kind of processor we expected Barcelona to be: A processor that restores AMD’s competitiveness and enables the company to regain market share.    

So, how is Shanghai different from Barcelona?

In plain performance numbers, AMD claims that the 45 nm chip is up to 35% faster than the 65 nm part and consumes up to 35% less power in idle mode. Also, the CPU is promised to be up to 40% faster in virtualization environments. Shanghai launches in nine different versions – five (2.3 – 2.7 GHz) dual-socket and four (2.4 – 2.7 GHz) quad-socket versions.     

In direct comparisons, Shanghai is priced much more aggressively than Barcelona as the 45 nm 2.5 GHz chip is priced against the 2.3 GHz 65 nm version. In comparison with Intel’s recently released 45 nm Dunnington quad- and six-core processors (Xeon 7400 series), Shanghai is substantially cheaper: A quad-core Dunnington processor will cost at least $1980 and a six-core CPU at least $2301; AMD’s Shanghai 8000 series starts at $1165 and tops out at $2149. Intel’s outgoing 45 nm Tigerton processors (Xeon 7300 series)  are priced between $856 and $2301.

The dual-socket Shanghai sells from $377 to $989, while Intel’s Harpertown (Xeon 5400 series) chips sell between $209 and $1493 (Intel offers 17 different Harpertown SKUs).

In terms of new technology, Shanghai introduces more cache (now 2 MB L2 and 6 MB L3), enhanced pre-fetching for DDR2-800 memory and smart fetch technology that results in power savings. There are also virtualization improvements with a faster world switch. AMD said that the transition to immersion lithography allowed the company to squeeze more clock speed out of the CPU and launch Shanghai at a higher speed than previously anticipated. There are two more features, which, however, are not enabled yet. Hypertransport 3.0 (HT3), which is scheduled to be activated in CPUs scheduled for a Q2 2009 launch and L3 Cache Index Disable, which is planned for a “2009” availability and “select operating systems”. AMD claims that HT3 will result in about 3% performance increase in dual-socket systems and provide about 10% more speed in quad-socket versions.   

All initial Shanghai processors are released as 75 watt “average power consumption” processors, which translates to 95 watt TDP CPUs.

Availability

Even if we haven’t seen any Shanghai processors and have not seen actual server systems in the market, AMD claims that all new Shanghai processors are available “immediately”. Lower power processors (55 watt) and faster (105 watt) CPUs will follow in Q1 2009, the company said. Of course, we are still remembering Barcelona’s TLB bug and we are sure that this new CPU will be scrutinized in detail. AMD’s John Fruehe told TG Daily that AMD has made adjustments in its validation process that are designed to avoid similar bugs in the future. Fruehe also told us that Shanghai silicon looked “very strong” from the beginning and reached production quality much faster than any other AMD processor before.

Fruehe noted that AMD has done “all the heavy lifting” with Barcelona and was able to apply that knowledge to Shanghai.   

 
Performance

The performance of server processors is a double-sided sword and no matter what we or any other publication will write, either AMD or Intel will be unhappy. The reason for this simply is that there are dozens of different benchmarks and even those may not indicate the actual performance of a processor as the provided speed and power efficiency will actually depend very much on a specific environment and needs. However, here are a few guidelines on what to expect.

AMD claims that a 2.7 GHz dual-socket Shanghai will beat Intel’s 3.0 GHz Harpertown E5450 in integer throughput, Java server and HPC server performance. The company also promises that a quad-socket Shanghai will beat Intel’s E7450 Dunnington processor in Java server performance and almost match it in integer performance. These results provide the overall impression that Shanghai will be a very competitive offering, especially if we hear that Shanghai has lots of headroom for clock speed increases.     

But there are also notes that AMD isn’t playing entirely fair with its published benchmarks. The Inquirer, for example, found that Shanghai may actually be slower than Intel’s Harpertown in some key benchmarks, and that AMD may have configured Intel systems in a way to gain an unfair advantage.

While we do not want to take sides here and simply do not have the benchmark number to prove one side or the other, there are more benchmarks that indicate that Intel’s fastest server processors still have an edge in certain applications, especially in terms of performance. For example, recently released SAP application benchmark results show that the 2.7 GHz Shanghai CPU slightly beats Intel’s fastest Tigerton processors, but clearly trails the fastest 6-core Dunnington CPU. However, there are, of course, always two sides to the story and the picture changes if we compare the power consumption levels. The 75 watt (95 watt TDP) Shanghai beats a 130 watt Tigerton processor with a score of 7010 vs. 6615, but trails the 170 watt Dunnington CPU, which hits a score of 9200.

The takeaway here may be that the actual performance and the answer to the question which may be the better processor will entirely depend on the needs of the customer. If pure performance is required, AMD admits that it cannot beat Intel at this time. However, it appears that AMD has made a big step in terms of power efficiency. It would be interesting to see how Intel’s low-power (65 watt) 6-core Dunnington processors (2.13 GHz) perform against the fastest 75 watt Shanghai CPU. We are not aware of any benchmark that pitches those two processors against each other.

However, even without that knowledge, Shanghai is set for a decent start and in fact looks like the CPU Barcelona should have been. And even if AMD is trailing Intel in terms of performance, it appears that AMD has begun regaining some ground from Intel. IDC’s latest market share numbers indicate that Intel has lost 0.6 market share points to AMD in Q3 of this year in the server segment. Intel is currently estimated to hold 85.6% of the x86 server market and AMD the remaining 14.4%.
   

More 45 nm

AMD will continue its tradition and introduce the desktop part of a new technology about 60 – 90 days after the server CPU. The firm said that 45 nm desktop processors are scheduled for a Q1 2009 introduction. The Dragon platform will include the Deneb processor, which will officially be called as well as a 700 Series chipsets and ATI Radeon HD 4000 graphics products.