The short version of the story is that AMD pulled their 10 core variant of Abi Dhabi from the roadmaps and put in an 8-core version. This chip is noting more than a mildly warmed over version of the current Interlagos/Valencia CPU with a Piledriver core instead of a bulldozer core. Like the consumer oriented Vishera, it is a yawner. Why? It brings nothing new to the table but incremental increases in energy efficiency and absolute performance. The Piledriver core is about 15% better than it’s predecessor and server parts might eke a bit more out of the total than that.
Unfortunately, AMD is woefully behind Intel with their stunningly good Romley platform, and 15% is not going to close the gap, not even close. To make matters worse, AMD did not touch the part that is farthest behind, the uncore. The dual die 16-core Abu Dhabi and it single die 8-core Valencia don’t have PCIe3, don’t have USB3 natively, and are, well, way behind the times. It is a stunning omission, but understandable.
Today AMD released 10 CPUs with four to 16 cores, 1.8-3.5GHz base clocks, 3.8GHz top bin turbo on a few models, and a TDP of between 85 and 140W. Prices start at a mere $293 for the 4-core 6320 and go up to $1392 for the fastest part, the 16-core 6386SE. All have the same 4-channel DRAM as their Interlagos predecessors because, barring the cores, they are exactly the same as Interlagos. Here is the full specs in handy table form.
Abu Dhabi is significantly more exciting a place than a chip
Back to the disappointment. Remember when we took AMD to task for “Project Win”? Well, this is what you get. When we said that you can’t delay a semiconductor product cycle to suit economic needs however imperative they are, we mean it. What helps make Wall Street happy 18+ months prior is an unmitigated disaster when the results meet silicon. Instead of a 10-core PCIe3 CPU, you get a mildly tarted up version of a design that was not competitive to begin with. This is not the tool you want to try to lead the charge to increase your 4% server marketshare with.
Is it better than Interlagos? Sure. Is it good enough in face of Intel’s year old parts? Not even close. Will it win any new customers? Not a chance. Will it tempt Interlagos customers to upgrade their CPUs? Even that is unlikely. The situation is dire, and Project Win did nothing to speed up future developments either. Why do we have such a dire prognosis? Take a look at the one, note the lack of pleural, benchmark vs Intel that AMD put in their presentation.
The results please
Update 11-5-12@4pm: OK we got this wrong. The graph shows absolute performance, not performance per dollar as it seems to strongly suggest. That said, on these three tasks, Abu Dhabi is competitive with the top end Xeon on performance, performance per watt, and performance per dollar. The last three paragraphs were completely redone in light of this news.
The Abu Dhabi and Seoul CPUs are better than their immediate predecessors, but still not competitive with their Intel competition. The Piledriver cores are indeed 10-15% faster than the Bulldozer cores they replaced, but still far from enough to win new customers. They may be a worthwhile upgrade for some existing Opteron customers, but there is precious little there to recommend to a new installation. Until AMD gets their uncore updated to modern technologies, this situation is unlikely to change.S|A
It looks pretty good, right? The 6380 is in a dead heat with Intel’s top bin Xeon E5-2690, and the AMD part costs $1088 vs Intel’s $2057, wow. Then you look at what the graph title is, and you realize that it is performance per dollar. This means that 16 AMD cores are half as fast as 8 Intel cores, in real world performance, and these are probably benchmarks favorable toward AMD. That is so far away from being competitive that it is absurd. Worse yet, any potential customer thinking of buying on price will be scared away by one other little problem, power use. The Xeon is a 135W part, the Opteron is 115W, so AMD is better right? Differences in how TDP is calculated between two companies aside, to do the same about of work, AMD needs two 115W CPUs for every Intel 135W CPU, not to mention the overhead of servers, switches, storage, and everything else. Power costs are a significant portion of TCO, and AMD is almost twice the power use of Intel. This essentially precludes AMD from any serious installations on energy grounds. In the end it is sad to say, but Abu Dhabi and Seoul have essentially nothing going for them other than being superior to their immediate predecessors. They are cheap to buy but not cheap to run on a TCO basis. They are performance per dollar competitive to purchase, but not performance per dollar competitive overall. By non-core performance measures AMD was way behind and with Abu Dhabi there is literally nothing new on this chip, nothing. The gap between Interlagos and Romley was yawning, with the new parts it is a little less yawning. As we said, in the semiconductor business you can’t do what AMD did and expect to not pay an untenable price.S|A
Author’s note: A lot of people blame Dirk Meyer and his team for this misstep. If you look at what happened, it is very clear that his regime was not to blame for this debacle. They had a 10-core Abu Dhabi with an updated uncore on the roadmaps. Once he was ousted, the new order, likely due to Thomas Seifert’s “Project Win”, castrated the 10-core part and did the minimum possible job for the lowest cost. This is the main reason we said that AMD’s management doesn’t understand semiconductors. The current regime has technical people in place that do understand semiconductors and are top notch technical people, but this was already set in stone before they were hired.
Latest posts by Charlie Demerjian (see all)
- VESA shows off working USB-C Alt Mode with Displayport - Mar 24, 2015
- UzBrain’s Rail Gun turns a toy weapon into an FPS controller - Mar 18, 2015
- Fotonation uses computational imaging to focus faces - Mar 17, 2015
- HSA foundation releases v1.0 of their namesake spec - Mar 16, 2015
- Mediatek tries to offer the complete device stack - Mar 16, 2015