AMD formally announced its new Naples server CPU today, with up to 32 cores, eight memory channels, and its new, 128-lane Infinity Fabric. Ryzen’s consumer performance may be impressive, but this is a straight server play, with AMD hoping to capitalize on strong performance in the 1-2 socket market.
The company’s strategy is similar to the one it adopted eleven years ago. Back in 2006, AMD’s server business was one of the company’s biggest success stories. Beginning in 2001, with the launch of Athlon MP and the 760MP chipset, AMD had slowly taken market share aware from Intel, particularly once Opteron launched. Its share of the worldwide server business peaked in Q2 2006, at just under 25% of the market. Today, AMD has less than 1% of the market, which is one of the reasons it’s so eager to rejoin that particular fight. Intel’s data center business was worth $17.2 billion in 2016, up 8% from 2015. If AMD can capture 5% of that business, that’s $860 million of additional revenue per year, and it’s revenue that comes with a nice margin.
AMD’s Naples (that’s the design name, not the marketing brand) server CPU is a multi-chip-module design rather than a unified 32-core chip. This makes smart sense for a number of reasons. First, GlobalFoundries has no experience building a 32-core processor on 14nm, and the sheer size of the die would be a bunker-buster. Instead, AMD has glued together multiple four-core CCXs into at least two multi-chip-modules (MCMs). Each MCM has four memory controllers and there are two groups of four in a 32-core chip, which is how AMD hits its 8-channel memory target. Total memory bandwidth per CPU is 170 GB/s.
At its Ryzen tech day, AMD showed demos of its Naples CPU platform performing a sparse matrix calculation on a 3D data set. When both AMD and Intel chips were held to the same configuration — 44 cores, DDR4-1866 — the Intel CPU took 35 seconds to complete a workload where Naples took 18 seconds. When given its head and allowed to run all 64 cores and DDR4-2400, AMD cut another four seconds off its time. That’s fairly significant performance, though we have to note that AMD released no details on how its Naples CPU was clocked, what clock speeds the chip will target, or even how the two servers were specifically configured. All such benchmark results should therefore be taken with a grain of salt.
Then again, after seeing Ryzen’s performance in desktop applications, the idea that it could serve up substantially better performance in memory bandwidth-bound applications isn’t surprising. Any test that can leverage memory bandwidth is going to love Naples’ 170GB/s of connectivity.
Why AMD is focusing on the server market
It may seem odd for AMD to prioritize breaking into servers over mobile, given that the mobile market’s revenue is much larger than the data center space, but there are some specific reasons for AMD to target this market.
First, Ryzen / Zen is unabashedly a server and workstation CPU. All of our benchmarks show that Ryzen is most competitive in computational workloads, particularly rendering tests. AMD is working on a mobile Ryzen, to be fielded with Raven Ridge later this year, but we still don’t know much about the GPU solution that chip will use or what configurations AMD will offer. Bringing up the CPU and GPU together on 14nm is more complex than launching a CPU on 14nm, so it makes sense for AMD to target the easier lift.
Second, the data center market is the only PC space that’s really expected to grow in the next few years. Consumer PC sales have fallen every year since 2011. The best forecasts for this year are either flat or predict a smaller decrease than previously. AMD is going to enter the mobile market with Ryzen, but it doesn’t make sense to prioritize launches in a contracting business.
Third, AMD has an opportunity to shine with Naples that it may lack in mobile. It is unlikely, for example, that any mobile APU will offer the same advantages in core count and memory bandwidth the server part can achieve.