We first heard that Qualcomm was planning to enter the server market more than a year ago, but the company has been quiet about its intentions and roadmaps since that first announcement. Now that the curtain has lifted, we can see why the company wanted to keep things hush-hush. Unlike other ARM firms, which have entered the market with relatively modest chips or attempted to focus on specific market segments like high-density servers, Qualcomm is pulling out the big guns. The company’s new ARM server CPU is a vigintiquattuor-core (read: 24-core) chip.
Here’s the real kicker — that 24-core version is just a prototype. The final variant is expected to pack even more CPUs per core. PCIe and storage controllers are all integrated on-die, but Qualcomm isn’t giving out information on the architecture yet — just that it’s markedly different from the existing Snapdragon processor. The chip’s supposed codename is Hydra, and it’s been reported that Qualcomm could roll up to 64 cores into a single stack, though we don’t have confirmation of that yet.
Based on the size of the current CPU (as shown above), we’d bet that any future 64-core variant would be introduced at a smaller processing node. That’s a huge piece of silicon, and while Qualcomm’s CPUs might be smaller than Intel’s on a per-core basis, 24 cores is still a lot to pack into a single chip. When Intel designs its highest-end Xeon processors, it spends a great deal of time laying out the internal linkages and L3 cache to ensure that all of the cores have equal latency and that cache contention and thrashing are kept to a minimum. We don’t know how much cache Qualcomm has dedicated or much at all about the internals of this new core, but the fact that they’re committing to hardware on this scale means they want to compete for heavy-hitter business. Think Google, Facebook, and other top-end customers.
At Qualcomm’s demo, the company showed its new core running LAMP — Linux, Apache Web server, MySQL, and PHP, along with OpenStack’s cloud software. PCWorld reports that Mellanox is designing network cards for the SoC, while Xilinx is building FPGAs that will accelerate particular workloads.
Qualcomm is far from the only company moving in the ARM processor space, but it’s easily the biggest potential threat to Intel. The other companies in this market — Cavium, Broadcom, AppliedMicro, AMD, and several others — have products that range from niche applications to nonexistent (at least, in shipping commercial volumes). Qualcomm, of course, hasn’t shipped anything either, but it’s absolutely got the cash to bring hardware to market.
The question is, will customers bite?
According to data gathered by the Electronic System Technologies Conference, the cost of server hardware is 57% of the total. That’s a sizable percentage, but this data set doesn’t account for the cost of the tools and software that actually run on the server. That’s where things get interesting. Companies like Facebook and Google, who write their own software and build their own cloud deployments, can afford to retool their own workflows and software to run on ARM-based equipment. Swapping server architectures is still a major endeavor for these companies, but they’re in control of the entire stack.
What about other companies, that typically provide hosting to other firms? For the likes of VMWare, Amazon, and Microsoft, the picture is murkier. The overwhelming majority of server and workstation software is x86-based, and the multi-year nature of licensing agreements and the need for guaranteed uptime and stability could make it difficult to persuade vendors to integrate ARM support.
Back in 2011, when Windows on ARM still seemed like a good idea and Intel was still pretending that the original Atom core (Silverthorne) was the only low-power chip it would ever need to make, the time seemed ripe for ARM to challenge x86 across the entire market. We saw prototype ARM netbooks, codenamed “Smartbooks.” Nvidia and Qualcomm were talking about making a concerted push into the x86 consumer space. Intel, meanwhile, thought that its new mobile chip, codenamed Medfield, would set the stage for its own ascendance into smartphones and tablets.
What happened in the intervening four years? A whole lot of nothing. True, Intel now has a healthy share of the low-end tablet market, but the company earned that by shipping hardware contra-revenue and by partnering with manufacturers on lower-end Android devices. Intel’s original plan, which was to lock down the high-end of the tablet market with superior x86 solutions, never materialized. No one seriously brought a Smartbook to market, though Chromebooks arguably inherited part of that Mantle, Windows RT turned into a disaster, and Calxeda, once seen as the vanguard of the new ARM server invasion, folded and died.
One common argument is that Qualcomm could win this fight by using Intel’s own strategy against it. Just as Intel once won the datacenter with a succession of “good enough” solutions at attractive price points relative to the existing RISC competition, Qualcomm could win the datacenter by exploiting its own cost scaling and the explosion of mobile devices. It’s an argument that hangs on two assumptions: First, it assumes that Qualcomm leverages its mobile IP into server chips (or vice-versa), lowering its overall costs. That could happen, but based on what we know so far, Qualcomm’s new chip isn’t Snapdragon with a new coat of paint — it’s a fundamentally new architecture and new IP stack. Second, it assumes that Intel can’t pivot to meet a new competitor.
In the past four years, Intel has done quite a bit of pivoting. It’s easy to criticize the company’s smartphone failures and decision to keep modems at TSMC, but Intel has been pushing its mainstream silicon relentlessly in recent years. At 14nm, it deployed its first low-power processes with the Core M. It’s introduced multiple new low-power Xeon products since ARM products began threatening over the horizon. And despite the claims of armchair CPU designers and amateur experts, the best research we’ve seen on the topic suggests that unless you’re measuring embedded power consumption, ISA just isn’t that important between x86 and ARM.
Of course, there are counters to this view, including the fact that Intel’s much-vaunted process engineering advantage is being threatened for the first time in over a decade. TSMC and presumably Samsung are known to have superior cost structures to Intel, which is why Chipzilla had to ship Atom contra-revenue in the first place. There’s definitely opportunity for ARM hardware overall, but the shape and size of that market are hotly contested, and Intel isn’t going to sit idly by and watch a competitor eat its most valuable market.