News & Analysis
Comment
chanj
Up to 4GB RAM may be a limitation to support multiple VMs with sophisticated ...
bobbytsai
here are a few more publicly announced Cortex-A licenses: Broadcom , NEC ...
Marvell quad-core ARM targets x86 servers
Rick Merritt
11/8/2010 4:24 PM EST
SAN JOSE, Calif. – Marvell will try to thrust ARM into servers and networking gear with a 1.6 GHz quad-core Cortex A9 chip debuting at the ARM Technology Conference. The Armada XP aims at a broad range of systems from low power Web servers for business to network-attached storage and media servers for the digital home.
The Armada XP delivers up to 16,600 Dhrystone MIPS at 10W. It includes up to 2 Mbytes L2 cache and supports a 64-bit interface to DDR2/3 memory running at up to 800 MHz. The chip is an SoC that integrates support for four PCI Express Gen 2.0 links and 16 serdes channels.
Marvell tipped in May its plans to deliver a 40nm ARM chip for low power servers. At the time, Marvell promised a five-fold reduction in power consumption compared to typical x86 server processors.
Earlier this year, Dell, Hewlett-Packard and IBM all expressed interest in architectures such as ARM for low power servers, especially for use in large data centers where power consumption is a bottleneck. ARM worked with Marvell on a server design used to run part of the ARM Web site.
Marvell is initially targeting 32-bit applications. ARM rolled out extensions to support virtually some 64-bit capabilities as part of its next-generation A15 core announced in September, but many analysts said many server makers will wait for ARM chips with full native 64-bit support.
"You can still do a lot with 32-bit code which is what most of the world still uses," said Richard Doherty, principal of market watcher Envisioneering Group (Seaford, NY). "ARM is going to be a force [in servers]," he said.
At least one startup, Smooth Stone (Austin) is working on a multicore ARM chip for servers. Another, SeaMicro, announced in June a low power multiprocessing system based on 512 Intel Atom chips to create a low power server.
Prior to the XP chip, the Armada line focused exclusively on mobile systems with mainly single- and dual-core ARM chips running up to 1.2 GHz. Marvell announced its first quad-core ARM chip, aimed at consumer systems, in January, but it's not clear the status of the chip.
"I know of no company with a shipping quad-core ARM Cortex core," said Will Strauss of Forward Concepts (Tempe, Ariz.).
Competitors including Qualcomm and Texas Instruments are also focused on single- and dual core ARM Cortex A8 and A9 chips, typically topping out at 1.2 GHz.


rick.merritt
11/8/2010 4:45 PM EST
If you are doing something to push ARM into low power servers I'd love to hear about it.
Sign in to Reply
DF
11/8/2010 11:50 PM EST
Definitely interesting, but it seems problematic that you have four cores that can address only a total of 4GB of memory, e.g. 1GB per core. Might be limiting for a lot of server apps. In practice it might be worse, with the system only able to handle 2-3GB (that's been my experience with 32-bit, non-PAE Linux).
Sign in to Reply
pica
11/9/2010 3:47 AM EST
2 or even 4GB of RAM is sufficiant for many, many server tasks. I think a system based on that SoC is on par with the Pentium III Tualatin quad-processor systems build in 2001-2002. These system also had just 2 or 4 GB of RAM. But nevertheless served well.
Sign in to Reply
neuraxon77
11/9/2010 12:03 AM EST
Typo: 800 MHz not 8090 MHz. It's also ECC.
Sign in to Reply
bobbytsai
11/9/2010 12:18 PM EST
here are a few more publicly announced Cortex-A licenses: Broadcom , NEC Electronics, NVIDIA, STM, Toshiba, Mindspeed Technologies, Freescale, Matsushita, Samsung, PMC-Sierra, Ziilabs.
Why are there no solutions like Tilera's 36 or 64 core products. Are ARM's per core licensing costs too high ? 32 bit address limitations ? There are a lot of multi-core non ARM product our there, Cavium Networks (MIPS), Netlogic (MIPS), Freescale (PPC), Azul (Java VM), Plurality ...
Sign in to Reply
chanj
11/9/2010 12:23 PM EST
Up to 4GB RAM may be a limitation to support multiple VMs with sophisticated service. The power saving may come down to how many VMs that can be supported on a ARM based server vs that can be supported by a x86 based server.
Sign in to Reply