Cavium Networks (NASDAQ:CAVM) has been getting a bit louder about its next-generation of microserver, networking, and storage system-on-chip products under the code name Project Thunder. While Cavium is doing some great things, it's important to not get lost in hyperbole and to keep expectations in check.
What's the deal with Project Thunder?
With its Project Thunder product, Cavium claims it will be able to pack in 48 custom-designed ARM (NASDAQ:ARMH) compatible processor cores (running at up to 2.5 GHz), lots of memory bandwidth, a mass amount of built-in Ethernet, PCIe, and a slew of dedicated accelerator hardware. The family of Thunder products is targeted at power consumption levels of 20-95 watts, and different flavors of these chips will be targeted at different workloads (storage, microservers, and networking).
You'll see a lot of people drawing comparisons to Intel's (NASDAQ:INTC) Xeon product families. In fact, chip analyst Linley Gwennap recently made the following fairly bold claim about Cavium's effort: "Cavium is really able to position this chip against the heart of Intel's Xeon product line, and that will be much more interesting to customers because that's where the bulk of the server market is."
However, while Cavium's products may compete in some of the same segments as Xeon, these chips actually seem fundamentally different in a number of key ways.
Intel's and Cavium's cores aren't comparable
While some may look at a 95-watt Cavium chip and conclude that magic is being done to fit 48 custom ARM cores at 2.5 GHz in the same power envelope that Intel fits four of its CPUs, it's important to cut through marketing mumbo-jumbo. Cavium's management will often multiply the core counts by frequency and claim something on the order of 120 GHz of performance. This could lead investors to look at an eight-core Xeon at roughly 3 GHz and conclude that Cavium is just slamming Intel, which would be misguided.
Cavium's cores are usually designed to be very narrow and aren't intended for monstrous per-core/per-clock (i.e., per GHz) performance. In fact, on numerous investor calls, the company has indicated that the stock ARM cores are actually overdesigned for the very throughput-oriented processors that Cavium builds. So for applications that require tons of per-core performance, the Intel Xeon is unlikely to be threatened by any of Cavium's offerings. Actually, it would be surprising if on a per-core basis, Cavium's custom ARMv8 cores were even comparable to Intel's latest Atom cores.
Cavium's secret sauce is integration
The "big deal" here is that Cavium's Thunder product family will be highly integrated and offer a bunch of specialized IP blocks to perform various functions. Intel's Xeon processors today are more focused on raw general-purpose computing performance. For networking-oriented workloads, Intel has traditionally offered special variants of its platform controller hub, or PCH, that integrate dedicated network acceleration engines -- but they're not single-chip solutions.
Intel, though, has signaled that in addition to driving integration of its low-power Atom-based microserver/networking/storage products, it will also build a new line of highly integrated system-on-chip products based on its "big" cores. Indeed, the first product here -- known as Broadwell-DE -- is slated to roll out either late this year or in early 2015. It should offer better direct competition to the Cavium parts at the high end -- although until we know just what that chip will integrate, it's too soon to tell how they will compare.
But how can it kill Xeon if first silicon isn't even back?
If you head over to Zauba, you will see that Intel has silicon back from the fabs, suggesting the company should be well on its way to a late 2014 or early 2015 launch:
According to a recent EETimes article detailing Cavium's Thunder, it is made clear that first silicon hasn't even come back from the fab yet and that sampling to customers will begin in late 2014. To put this into perspective, Intel will have an entire suite of 14-nanometer integrated system-on-chip products -- Denverton for the very low power space, and Broadwell-DE for higher performance -- available in the market well before Cavium's first 28-nanometer -- two generations behind what Intel has in terms of chip manufacturing technology -- Thunder products are available in high volume, which is likely to be sometime in late 2015 or early 2016
Cavium is great at what it does, and its growth over the last few years has been phenomenal. However, it is important to not be swept away by hyperbole. Intel's moat in the traditional enterprise server market is broad and deep, and it is very difficult to imagine that Cavium -- or any other alternative architecture chip vendor -- will be able to compete with the performance, power, and cost of Intel products. That said, when the first silicon is in and performance results are made public, we can do a deeper analysis of the competitive situation.