To build a truly heterogeneous SoC, you need at least four processors
The announcement of ViriSilicon’s acquisition of Vivante heats up the IP space, and puts pressure on Synopsys (which recently acquire Virlogic, which previously acquired ARC) to get a GPU, and ViriSilicon to get a CPU, so they can compete across the board with ARM and Imagi-nation Technologies.
Synopsys has the ARC CPU, a RISC design not too dissimilar from ARM (and also developed in Cambridge). ViriSilicon has a DSP (LSI’s ZSP), but neither Synopsys, ARM, nor Imagination has a separate DSP core, although ARM (which added DSP extensions in the Cortex-M7), Imagination, and Synopsys have some DSP capabilities. And only Imagination and VeriSilicon have an ISP core, while Synopsys has an embedded vision processor as well as integrated ARC-based IP subsystems, and VeriSilicon has a new vision processor (VP), which is a new and growing cat-egory of dedicated processors for automotive and IoT platforms; see the table below.
Now, to be fair, I must acknowledge that is not necessarily true that an SoC has to have four processors. Many SoCs have fewer than four processors; many only one. The number of proces-sors needed depends on the application. So with my pixel bias, I am, of course, referring only to vision/video/graphics-capable SoCs, of the type that are typically found in a smartphone, maybe a car, or a tablet. I could also add that increasingly connectivity processors are in the mix (e.g., Imagination’s Ensigma programmable/configurable RPUs for Wi-Fi/BT, etc.) and security processors like ARM’s TrustCenter.
Synopsys says ARC is number two in terms of share of licensed microprocessor shipments, according to SemiWiki. Synopsys also claims to be the only company that licenses a pre-integrated processor IP subsystem.
As for VeriSilicon, they aren’t interested in the CPU business. They would rather partner with ARM (or X86, ARC, MIPS, PPC, Andes …) for the CPU and specialize in everything else in the SoC.
There are other companies that could be added to the list. AMD, for example, has sold IP for over a decade, and in their recent financial call, Lisa Su, CEO of AMD, made a point about how the company is pursuing embedded and strategic IP deals.
The IP business is a monster to keep track of. IP is available for everything you can name in a PC, TV, smartphone, or any other computer-based device. Our interest in things that generate or modify a pixel gets skewed a bit in the IP space because things get so interlinked (like DSP extension to a RISC CPU—there’s pixel manipulation in there, but it’s tricky to tease out).
And then there are specialty companies like DMP that offer IP for GPUs and special accel-erators, and ThinkSilicon, which offers an IoT GPU IP core, or SiliconArts’ ray-tracing core. Ca-dence’s Tensilica’s DSP with vision extensions is another prime example.
As the number of GPU and CPU companies consolidated, several silicon suppliers went into the IP business, extracting themselves from the costs of producing a chip. That reduced their ASP from $10 to $100, down to 10 cents or less, but skyrocketed their bottom line percentage as all the COG (cost of goods) disappeared from the P&L. The IP business, once scorned as being where the not real men had to go, be¬came very desirable to both chip build¬ers, who could offset some R&D cost, and other chip builders who wanted out of the rat race.
But it is impossible to define IP without a suffix; IP is like electricity, it—everywhere, and everything. Everyone thinks they know what electricity is—it makes the lights come on and the toast pop up. And everyone thinks they know what IP is—someone’s intellectual property, the stuff companies sue each other over. Except when it’s not, and it’s a product, or an RTL listing, or a function, or a core, or an element, etc., etc. It can even be just an idea expressed as “could be implemented as a …” IP; it’s ethereal, essential, and elemental.