AMD unveils its first ARM CPU, the 64-bit 8-core Opteron A1100

• By Sebastian Anthony on January 29, 2014 at 9:55 am
• • 
  • • AMD has finally taken the wraps off its upcoming 8-core 64-bit ARM SoC, codenamed Seattle. Seattle (officially designated Opteron A1100) is a server-class clip, with four or eight 64-bit ARM Cortex-A57 cores. The part, which begins sampling in March, is aimed squarely at the low-power server market, where AMD hopes that the SoC’s low cost (about one tenth the cost of competing Intel Xeon parts) can wrestle some market share away from Chipzilla. Performance-wise, AMD only gives rough figures, but it appears that the top-end A1100 will be around 2.5x faster than AMD’s current low-power Jaguar-based server chip (the Opteron X2150), while maintaining the same TDP.
      AMD’s A1100 SoC is a bit of a beast. There’s the option of either four or eight CPU cores, “based” on ARM’s 64-bit Cortex-A57 design (it isn’t clear whether AMD has modified the core or not), clocked at 2GHz or higher. Each pair of cores have 1MB of shared L2 cache (up to a total of 4MB), and all cores will share 8MB of L3 cache. The main feature that AMD seems to have worked on is a new memory controller, which is 128 bits wide and supports both DDR3 and DDR4 RAM, in SODIMM or DIMM format (registered or unbuffered). If registered DIMMs are used, a single A1100 SoC supports up to 128GB of RAM. The chip will be built at GlobalFoundries’ on its bulk silicon 28nm node.
    • 
    • Rounding out the specs, there’s an on-die 8-lane PCIe 3.0 controller, an 8-port SATA 3 (6Gbps) controller, dual-10GbE, and ARM’s TrustZone tech. AMD assured Anand that the chip was capable of sustaining full bandwidth to all eight SATA ports.
      In short, except for its wimpy CPU performance, this is an impressive, server-class chip. With support for so much RAM, tons of hard drives, and two 10-Gigabit Ethernet connections, AMD is positioning the A1100 as the ideal chip for low-power, high-density server farms. There are many situations where the raw computational power of big, expensive x86 CPUs isn’t required. Serving websites, cold storage, and Memcached are all prime use-cases for AMD’s first ARM chip. AMD isn’t talking exact pricing yet, but the “total solution price” should be around one-tenth of a Xeon-powered server — or around $100 per Opteron A1100. The general idea is that you can take those huge cost savings and plow it into more storage and RAM — a very desirable solution for many enterprise customers, I assure you.
    • 
    • There’s no word on whether the A1100 will include a graphics core, but with a 25W TDP it probably won’t. Once AMD has the toolchain and developers to make heterogeneous computing a relaity, though, an ARM-based Opteron APU would be a definite possibility. (Read: AMD buys SeaMicro, starts selling Intel-based servers.)
      AMD says it intends to sample the Opteron A1100 in March, with a reference board available for those who want to try the architecture out. Servers based on the new chip are expected to be announced in Q4 2014. Over the next few months, we can expect to see some real benchmarks, power consumption figures, and so on. What remains to be seen, though, is whether ARM is truly viable as a server architecture. Just last month, Calxeda, one of the original ARM server startups, closed its doors. AMD definitely has the engineering skills to produce a strong chip, but actually getting big customers to move away from tried-and-tested x86 will be difficult — especially with ARM’s much weaker software ecosystem.
    • 
    • According to AMD, small and cheap chips always win out … so watch out, Intel!
    • 
    • AMD lays out its (hopeful) future in this slide
      Just as with Kaveri and its promise of heterogeneous computing, the impact of AMD’s ARM efforts may be significantly delayed. Both chips are exciting on paper, but in reality it’s very much a case of wait-and-see.