For more than 30 years, Intel has made a name with its CPUs, which rule the PC and server markets. Chip advances have helped Intel make devices smaller, faster, and more power-efficient.
Computing is now spreading into cars, robots, drones, smart devices, and a wide range of other electronics. Chip requirements have changed with new hardware and applications like artificial intelligence and graphics.
Intel is preparing for the future and making big changes in the way it designs chips. It is realizing that CPUs aren’t a big enough business model and has acquired a wide variety of chips to put into devices. The company is even researching quantum computers and neuromorphic chips.
Intel intends to team up its CPUs with Altera FPGAs in cars and Movidius computer vision chips in drones. It will also use its Nervana deep-learning chips in servers and other devices. Intel will make chips for different types of computers but also offer faster throughput inside its chips with its new design approach.
The fresh approach to chip design is necessitated by a slowdown in Moore’s Law, and difficulties in making smaller chips. Some applications like AI also need superfast throughput inside chips.
One of Intel’s changes is a mix-and-match heterogeneous design where different types of cores can be put in a single chip package. Under the new design, it’ll be possible to mix different architectures on a single chip. Chip packages could also have cores made using different manufacturing processes.
Another technology improvement is the EMIB, or Embedded Multi-die Interconnect Bridge, which allows for the integration of multiple chips inside a single package using a much higher-bandwidth interconnect than alternative technologies. Intel is already using the technology on its Stratix 10 FPGA but will expand it to other chips.
With this approach, Intel could put a high-performance ARM and x86 core inside a single chip package. Intel has already been able to do that but the new approach will make communication between the cores much faster.
Intel so far has integrated chips into system-on-chip designs. But EMIB is a much faster interface and is also an efficient way to create chips for heterogeneous computing.
EMIB provides the opportunity to design chips for a much wider range of devices outside of PCs, said David Kanter, president at consulting firm Real World Technologies.
“What’s different is when you have different chips to connect together, now you have a way to do that,” Kanter said.
The alternatives, like silicon interposers, are expensive, Kanter said. EMIB bridges chips with more pins and wires, which gives it much wider bandwidth than a conventional PCI-Express throughput technology, which has power constraints.
The technology will come to PCs and servers but only in devices when needed. The throughput on EMIB could into the terabyte range, which most applications don’t need, Kanter said.
EMIB will also support multiple throughput technologies, paving the way for Intel to put Infiniband and silicon photonics throughput technologies inside its chips. Intel has said it wants to enable cores to communicate using lights and lasers via silicon photonics.
The new approach also allows Intel to use cores made of exotic material inside its chips, Kanter said. Intel could connect its x86 chips to a separate core made of III-V material, which could ultimately replace silicon.
Ultimately, the new approach to chip development comes down to economics, performance, and power budgets. Systems have thermal constraints on voltage regulation and cooling, and all of these packages need to be able to fit into the economics of chip-making.
The EMIB-like integration should work well for furthering AI hardware, assuming the thermals don’t constrain the approach. The design allows low-cost changes to the hardware.
AI workloads are especially complicated, with large datasets and unpredictable data access patterns. A huge amount of chip connectedness would be just the ticket to compensate for that.
For chip designers, the EMIB-like design is also an invitation to design for Intel’s x86 CPUs. Intel’s conventional x86 CPUs have been a barrier to entry, and it takes a serious investment to even think about trying to compete. Intel is showing signs that external IP can be easily attached to its chips.
But attaching external IP hasn’t always worked out in chip design, and it remains to be seen if Intel will actually be open to easy integration of external IP.