AI chip uses special multi-port memory

November 18, 2021 // By Jean-Pierre Joosting
AI chip uses special multi-port memory
sureCore designs special high performance multi-port memory for Semidynamics RISC-V-based, tensor processing chip for AI applicatiions.

Embedded memory specialist sureCore has designed a power and area efficient, high performance, multi-port, embedded memory for Semidynamics’ new RISC-V-based, tensor processing chip. Semidynamics is developing a high bandwidth, vector processing unit optimised for tensor processing aimed at AI applications. In order for AI to realise its full performance potential, the speed at which tensors are mathematically manipulated is critical. To enable this, the vector compute units must be tightly coupled to a high-performance register file. Whilst this can be implemented via a standard synthesised approach, the outcome is often inefficient in terms of power and area. The sureCore solution delivers optimal and predictable performance metrics for all key parameters.

Semidynamics’ CEO, Roger Espasa, explained that sureCore was able to deliver performance exceeding 2.5 GHz with compelling area and power characteristics. “Using a standard, synthesised, physical implementation flow would deliver the performance we need but at the expense of power and area,” he explained. “Achieving the optimal configuration of Power, Performance and Area (PPA) is always the key goal in realising a chip design with the right market value proposition. This is why we went to sureCore to exploit their embedded memory expertise as this is fundamental to meeting our performance goals. In fact, it is a central feature of our highly parallel architecture that necessitates multiple instantiations of the high-performance register file to deliver the performance. Effectively we de-risked a critical part of the design.”


Vous êtes certain ?

Si vous désactivez les cookies, vous ne pouvez plus naviguer sur le site.

Vous allez être rediriger vers Google.