Another good example of this is the attempt to cut PCIe out of the Atom line, recently forbidden by the AMD/Nvidia/FTC settlement. If Intel had been allowed to cut PCIe out of their chips, how attractive do you think it would be to a server vendor to have a CPU with no, I repeat NO I/O capabilities? Intel was saved from their own marketing stupidity by the FTC, despite their protests. Do you sense two competing camps at Intel? Any guess as to which one is winning?
Intel went so far out of their way to cripple their products that, well, no one wants them any more. They left the same sized hole in the market that AMD drove the A64/Opteron truck through a few years ago, and Transmeta almost drove their chips through. If Intel Israel hadn’t effectively gone around management wishes to bring out the brilliant Merom/Banias/CoreX chips, Intel would have been in even more hurt. AMD is in the early stages of repeating the truck driving with Ontario/Zacate/Fusion, and they are driving it through the same stupid hole that the ARM guys are eyeing.
Because of the time it takes to make the design level changes needed to combat ARM, it will be years before Intel has an actual, not PR based, response. In that time, ARM based server chips will have gotten a foothold in the data center, proliferated, and changed the data center admin mindset about what is acceptable. Intel has already lost, but that doesn’t stop the press releases and slideshows.
Remember when we broke the story about Facebook going to ARM based servers? We weren’t kidding, they still are, despite what they tell Intel, and despite what they say publicly. Go read the ‘denial’ that they made at the bottom of the article, but were unwilling to send to us, and parse it carefully. Game over for Intel at that niche.
These holes are self-inflicted, and Intel seems to not only repeat the process time and time again, the behaviors that lead to them are becoming more common in Santa Clara. They keep making chips that no one wants, forcing them into artificial niches to extract cash, and act surprised when a competitor exploits the niche. It doesn’t take a marketing genius to see this.
That brings us back to ARM. The recent news was about Calxeda, the company formerly known as Smoothe-Stone. They are putting 120 4-core ARM A9 nodes in an ~600W 2U package. Each node is based on an ARM A9 chip, and each one is designed to be a server part, and fully capable of playing the role. Please note, there is no mention of hamstringing the chips to keep Nvidia from attaching a GPU, or to keep vendors from making parts that might compete against a low end Sandy Bridge. There also doesn’t even seem to be any trace of extorting profits based on not doing those behaviors, instead, they are just going to be making the parts that customers want.
The very first, and far from the only, ARM server shot across Intel’s bow has hugely better price/performance, hugely better performance/watt, and about 5x the density of the best Intel solution on the market. It is a data center operator’s dream come true, with the usual caveat of, “if your app fits the solution”. The desperate problem for Intel is that with each speed bump, a much larger portion of the server workload fits the ARM solution, and less and less of it fits the Atom solution.
Intel seems to have finally noticed this little blip, but the truck has already driven through the front gates, and is starting to unload the first boxes at data centers. In a swift, as Microsoft would define swift, response, Intel is leaping into action, and promising…. well, if the press conference they held today is to be believed, nothing.
Yes, Intel’s panicked response to a failed Atom architecture and annoyed customer base is to promise that in 2012 they will have something competitive to what the ARM guys are shipping now. Maybe. But no promises. And half of the features may not be there unless you pay a painful premium. Intel roadmaps have 8-core Atoms slated for the 22nm-LP node, but remember, this is the -LP roadmap, and 32-LP has yet to ship. It is very likely that ARM will always be ahead of Intel in core count, as they are now, should the market care.
Latest posts by Charlie Demerjian (see all)
- Spin Transfer Technologies talks about their ST-MRAMs - Sep 27, 2016
- ARM adds CMN-600 interconnect and DMC-620 memory controllers - Sep 27, 2016
- Globalfoundries 7nm process isn’t even close to the name - Sep 26, 2016
- ARM upgrades realtime offerings to v8-R and adds Cortex-R52 - Sep 21, 2016
- Everspin and Globalfoundries team up for embedded ST-MRAM - Sep 15, 2016