Tilera is finally shipping their 3rd generation TILE-Gx processor in volume, or at least half of the Gx family. Today they announced the immediate availability of the 16 and 36 core models along with a few customer wins.
TILE-Gx CPUs come in 16, 36, 64 and 100-core variants, with the first two being the focus of today’s attention. All of the chips share a unique architecture, you can read about it here, and have a very unique network architecture. Each core has six network links to the four adjacent tiles, 24 total, with each link dedicated to a specific function. Tilera calls this CPU + network node a tile, and it is the basic building block for their architecture. The two smaller and two larger chips each share a common set of uncore components.
With this many cores, you probably don’t need much convincing to believe that each core isn’t capable of much raw single threaded computation, but that isn’t the idea. Instead, the borderline silly amount of bandwidth to each core means that it can get almost any data it needs in and out of a core without being blocked, and with very low latency.
This adds up exactly what you want for network chips and multimedia processors. Strangely, Tilera lists both categories as wins for their chips, along with the more generic ‘cloud’ column which could be just about anything nowadays. In any case, two names, Mercury Computer and Harmonic were listed as customers, but the one we were expecting was absent. While the big G still might happen, Tilera wasn’t talking about it today.
The message is still the same across categories, it can replace racks of more generic, read Intel, CPUs with Tilera chips taking a fraction of the power, space, and other metrics that worry data center designers. As we are always saying, if your workload fits the paradigm, Tilera silicon should be the greatest thing since sliced amorphous silicon crystals. If not, you are much better served with just about anything else.
With a two named customers, 20 claimed design wins, and 80 engagements, it obviously fits enough customers to keep things interesting. With the official release today, you too can buy one if you want to kick the tires on the architecture without an NDA.S|A
Latest posts by Charlie Demerjian (see all)
- Analysis: What does the automotive market look like for Nvidia? - Aug 19, 2014
- Analysis: Nvidia refocuses it’s Tegra program - Aug 15, 2014
- AMD’s Fiji GPU not arriving soon enough - Aug 14, 2014
- Intel talks a bit about Broadwell-Y - Aug 11, 2014
- Intel talks about their 14nm process - Aug 11, 2014