You are currently viewing SemiWiki as a guest which gives you limited access to the site. To view blog comments and experience other SemiWiki features you must be a registered member. Registration is fast, simple, and absolutely free so please, join our community today!
Faster and cheaper.... that's a worthy business model. Now let's see some revenue! This heated race is going to leave the ARM server folks in the dust?
ARM servers should be even cheaper plus with better TCO (thanks to faster transitions to new process nodes).
You have to take into account, that x86 processors are extremely overpriced today. Most expensive x86 CPU cost similary to fully equipped SPARC server or 2x dual socket (4 CPU total) POWER 8 server. It is CPU vs server cost comparison...
ARM servers should be even cheaper plus with better TCO (thanks to faster transitions to new process nodes).
You have to take into account, that x86 processors are extremely overpriced today. Most expensive x86 CPU cost similary to fully equipped SPARC server or 2x dual socket (4 CPU total) POWER 8 server. It is CPU vs server cost comparison...
I do not think Arm processors will ever be able to compete with X86_64 (and other architectures designed for speed). ARM was designed for low power. One example that comes to mind is the bits in instructions to dynamically turn off execution.
ARM cores as provided by ARM at the moment are designed for low power. I am not really sure that the ISA per se (especially in the v8 incarnation and SVE extensions) has anything that makes it inherently suitable only for low-power and not for higher performance systems. I believe that it is just that noone in the ARM space has ever designed a really high performance core and it takes time and multiple iterations to succeed (and a lot of optimized software to be written). But things are changing. I guess the first change will come with Fujitsu presenting their ARM-based processor for the next super that they are designing and Qualcomm entering the server race. I saw a convincing first attempt of a Cavium server at a demo in the recent ARM Research Summit in Cambridge where for a certain biomedical application that was memory bound, it seemed to perform favourably compared to a Xeon v4. We'll have to wait and see. If I had to make a buying decision at the moment, I would not buy an ARM server unless I had an application that could be proven to work very well with the ARM-based offerings (like the application I saw at the demo). If I needed a more general-purpose yet high performance solution but not x86, then I would look at POWER. Hopefully the POWER9 seems to be really nice and I am really interested in the CAPI interfaces as I work with FPGA accelerators. And generally I think that AMD's come back is going to shake the x86 world for the better as well.
Conditional execution was removed from ARM in 64 bit instruction set. It is significantly rethought.
Mostly it will be a battle of out of order, support for acceleration of math and AI, cache size, memory bandwidth, interconnect fabric, transition to PCIe gen4 or CCIX, etc. It is not really about instruction set. It is about the ecosystem.
