Tilera is a small CPU design firm that first attracted attention back in 2007, when it debuted its TILE64 architecture. The company's tech is designed to offer a grid of CPU tiles. Each tile contains a very simple CPU core, its cache, and a router. All of the processors are attached via mesh networking. Each tile has its own L1 and L2 cache. If a given CPU has a local L2 cache miss, it can reach out and search the combined L2 cache of the entire processor cluster. Tilera refers to this as a "very large, effective L3 cache."
Today, the company demonstrated a 100-core processor it claims is capable of competing with the best Intel and AMD have to offer. "The reason we can go against Sandy Bridge architecture is [Intel's range] was designed for general-purpose [applications], so it has to account for single-thread performance and power-point performance and Windows," Ihab Bishara, Tilera's head of marketing, told ZDNet UK. "What we're targeting here is a very specific [high-throughput] application... If we compare our chip to Sandy Bridge in the standard enterprise application, we will not do well."
Tilera's updated 100-core processor
Tilera's design shares a number of similarities with Intel's 2006 Teraflops Research Processor and the "Single-chip Cloud Computer" the company demo'd back in 2009. According to company executives, Intel's decision to focus near-exclusively on general purpose x86 processors make it impossible for the CPU giant to maximize performance for any given workload. The company implies that by explicitly targeting cloud computing, its own tile mesh architecture will deliver both raw performance and performance-per-watt that'll flatten the best Intel has to offer.
The problem with Tilera's argument is that it's not too different from what big iron RISC vendors were saying in the early 1990s. Back than, Intel's x86 architecture wasn't seen as a credible threat to the various RISC architectures that powered the servers, workstations, and big iron deployments of the IT industry. Intel proved them wrong in part because it managed to aggressively scale the clockspeed and performance of its x86 CPUs while simultaneously taking advantage of economies of scale. x86 solutions of the day didn't necessarily offer best-in-class performance or stability, but they hit "good enough" criteria straight on.
Tilera's new 40nm process will help the company compete effectively in its chosen niche market, but it's facing a slog. There's already a great deal of chatter concerning the low power CPUs under development at ARM, Intel, and AMD. Intel's Knight's Corner isn't a manycore mesh processor like Tilera's new Gx3036 (36-core), Gx3064 (64-core), and Gx3100 (100-core) products, but it's built to confront some of the same bottlenecks. Tilera, like Intel's ill-fated Itanium, depends heavily on its compiler to squeeze maximum performance out of its program code.
We don't doubt the company's performance claims as such, but we're not certain its mainstream enough to acquire its own developer community.