Sure, as that's true of any industry. The "problem" is we don't know which it will be, and they'll have to battle it out. I don't really see Qualcomm as a server company, though. I don't think they have their hearts in it. Nvidia, Applied Micro and either Calxeda or AMD will be the third.
But again, it's very premature to call it out now. We still have at least 5 major players (Samsung, Qualcomm, Nvidia, Mediatek and Apple) in the mobile chip market, and it's a much more mature market than the ARM server market, so we're a long way until we know for sure.
What exactly is an ARM server chip? ARMv8 with 10G networking? Seems like a good combo for embedded networking applications also. Seems to me the best approach is still to provide a selection of generic, expandable interfaces, and most roadmaps I've seen (and I've seen quite a few) the chips will offer PCIe, USB3, SATA and 1G in addition to 10G or more.
I think you may be right as far as the number of vendors who end up inside the bulk of server shipments, but again, what exactly is an ARM server chip anyway?
The ARM site lists 49 Mobile SOC IP vendors, that would be too many by how much in your opinion?
That's the ARM way, lots of vendors, competing often by carving out a unique niche in the ecosystem.
I guess maybe twice to three times, 16-32 ARM server vendors is possible. Think of the different things servers may do. The silicon needed for a render farm is different to a mass video encoding (YouTube) . Different again to a highly transactional database (banking), again to Data Wharehouse analytics. All with different compute, data latency, data throughput requirements.
Efficiency through specialisation.
ARM's A9 & A15 cores will do just fine for :
HTC = High Throughput Computing
and not really for :
HPC = High Performance Computing
A15 need to put a cap on its peak power consumption in the range of less than 2 watts/core should be the target in the longer run.
Many-many, too-many cores will help to build next-gen data centre ARM server processor.
Data centre server processor should handle maximum small threads in the range of 16/32/64/128 simultaneously.
Data centre software threads are not compute (math) intensive and hence having many-many cores to handle many-many threads on a single chip will lead to increased energy efficiency and lower operating cost.
I am guessing he is about 3 too high on the count of survivors. What exactly can an ARM server core do than an undervolted x64 can't? And all that software that needs to be rewritten, retested, and recertified. It's not exactly that Intel will allow itself to be seriously underpriced, and this removes the only possible way these things can make inroads. Plus finfets should let them drop supply voltage even further. I smell an Itanium.
APM, Calxeda, Cavium, HiSilicon/Huawei, Marvell, Nvidia, Qualcomm, Samsung, TI are addressing ARM-derivative Server.
Three original contenders; Calxeda, Marvell, APM were well categorized; low, mid, high perf segments.
Calxeda & Marvell initial efforts have been shielded from Intel on unique 32 bit engines which Intel does not address in portfolio. Switch in SOC this analyst suspects from Intel during 26 nm gate length run end before true 22 nm shrink; exposes Intel to political issues among NIC, switch & route compliments in Intel server environment.
Nvidia & TI appear uniquely segmented into HPC, where TI w/DSP accelerator and Nvidia w/GPU ala CUDA on ARM control plane address FLOPs instead of MIPS and IOPS.
Analyst suspects Nvidia large multicores v Intel Atom Octa and bridge power reduction between discrete CPUs on blades.
Needless to say end market is pressing for ARM 64-bit, and as usual, ARM willingly licensees all comers to see what sticks.
ARM under financial pressure in slow market seeks a royalty hedge; insurance. Where Intel has won 3 out of first 4 ARM offensive moves into notebook, Smart TV & Signage where Intel won without a fight, where in server Calxeda does reach first base, only A15 in phones and tablets brought in ARM base holders.
Intel has ARM cluster surrounded in its strongholds and ARM Ltd. is feeling the squeeze.
Intelís crush campaign targeting ARM and AMD has been especially effective dumping 24,993,201 Sandy Bridge Celeron at cost last & current quarter.
Celeron dumping will be followed by 25 million dumped ATOMs & Ivy Bridge sludge through first half of 2013.
Finally, Intel confidence agents have been effective curbing ARM ability to compete by disabling ARM Ltdís strategic marketing functions on all the old hazards of Intel relations and the politics of Intel Imperial Court.
Blog Doing Math in FPGAs Tom Burke 15 comments For a recent project, I explored doing "real" (that is, non-integer) math on a Spartan 3 FPGA. FPGAs, by their nature, do integer math. That is, there's no floating-point ...