The Barcelona processor from AMD keeps the tempers in the processorwereld been around for almost two years. The initials and on lwa le hype, and the subsequent problems and disappointments, I have already summarized on the occasion of the new year, so here I can be brief by saying that there long bad has seen for the product.
Even today with the K10 yet, not anywhere up to the expectations. The Phenom 9850, 2.5 GHz chip based on the bugvrije B3-stepping, appears in a review of AnandTech, on average, about ten percent slower than the 2.53 GHz Core 2 Quad Q9300. We compare the top model of AMD, with the slowest 45nm-vierkopper of Intel. Who can get a Phenom choose will also be dozens of watts extra power consumption into the bargain.
With this information in mind, it is not surprising that the Phenoms are so cheap. Some people find that a sign that AMD is doing it well’, because they are the consumer with a small(er) budget control. However, it seems strongly that this strategy is born out of necessity, and that he is for the company itself is not really beneficial extract.
The situation for servers is in stark contrast with those in desktops. Now the B3-stepping is off, and the quad-core Opteron for the first time is normally available, there are finally a number of good standardized benchmark results of which appeared to know SAP-SD 2-tier and TPC-C. Let’s look at how the scores relate to each other:
TPC-C (4 sockets)Power64,2GHz256GB$1.566.380
629.159Xeon2,93GHz256GB$1.337.458
516.752Xeon2,93GHz128GB$694.335
407.079
Barcelona2,3GHz128GB$502.836
402.234Itanium1,6GHz192GB$673.551
372.140Opteron3,2GHz128GB$549.416
262.989
The score that HP drop for Barcelona, with a new Proliant DL585 G5 is over fifty percent higher than the next best AMD result on the list, which still dualcorechips is achieved. The nearest Xeon-score uses the same software and the same amount of memory, but is just completed. n percent higher. The second entry for Xeon is a much more churlish system with twice the memory and a price tag that You say. Also is this on IBM DB2 rotated in place of Microsoft SQL Server 2005, what the figures even less comparable. In short: 402.234 against 407.079 is the fairest contest. Let’s look at the second benchmark look:
SAP-SD (4 sockets)Power64,7GHz64GB 
4010Xeon2,93GHz64GB 
3780
Barcelona2,3GHz64GB 
3524Itanium1,6GHz48GB 
2150Opteron3,0GHz32GB 
2100
The advantage of the Xeon is here with seven percent slightly larger, but very convincing, it is still not. This is not good news for Intel. The Barcelona has a 95W tdp, cost 1019 dollars and used ‘just’ registered ddr2-bars. The Xeon on the other hand has a 130W tdp, cost 2301 dollar, and only works in combination with relatively expensive and warm fbdimm memory. That is a high price for less than ten per cent extra performance.
What is there to the hand which Intel almost effortless wins in desktops, but in servers with four sockets on all sail to a small victory books? The answer lies in the bandwidth. A machine with four Opterons can, in theory, 42,4 GB/s from or to the memory pumps, while a machine with four Xeons up to 32GB/s. That last is the sum of 21.3 GB/s read and 10,6 GB/s write, which is actually especially reading meeweegt. If we look at what’s on a per-socket is possible, then it becomes even worse: a Xeon MP can hold up to 8.5 GB/s by fsb presses, while an Opteron through its local memory to combine with HyperTransport links to 34.6 GB/s can handle.
There are a lot of hooks and eyes to this calculation, in the first place, the fact that the other processors and their bandwidth is largely z lwa, lf necessary. The fact remains, however, that systems with direct interconnects and geheugencontrollers in practice, very have proved successful, both for AMD in x86-land, companies such as DEC, HP and IBM which has been previously used in their high-end server products.
The bad news
The current situation will not be long: at the end of the year will be Intel’s Nehalem and Tukwila catching up on the area of bandwidth. The Xeon goes up to 12.8 GB/s over the front side bus to 70,4 to 83,2 GB/s from three ddr3 channels two QuickPath connections, depending on the selected QPI speed of 4.8 GT/s or 6.4 GT/s). In practice, it is said to be at least a factor of four improvement. However, that is still child’s play compared with Tukwila: the Itanium is of 10.6 GB/s to 147,2 GB/s from four fbd2-channel, four full and two half QuickPath links, an increase by a factor of fourteen. The planned for 2009 octalcore Xeon will be on the same infrastructure to respond.
AMD is lucky that Intel is always slow with the release of Xeons that are suitable for four sockets. Not that the for this year planned Dunnington with six cores and 16MB L3-cache, no better results will show, but because he is still stuck at a pathetic-afstekende bus 8.5 GB/s will be Barcelona or otherwise, Shanghai is still in a near – even if only in certain benchmarks, or within a particular budget.
On the whole, however, it will still have to swallow for AMD. Five years in the can enjoy of ge lwa in geheugencontrollers and interconnects while Intel but with his old-fashioned buses remained mired in recession. If that advantage disappears, it will be a lot more difficult to keep up in the benchmarks. AMD had less than a year can enjoy this situation as the Barcelona at the time had come, but by the TLB-bug is a large part of that time is wasted. Now is the case to have a lot to benefit from it, because as soon as the big hand of the ‘tick-tock’ at Nehalem state, everything will change.