Speedcore eFPGA IP architecture incorporates many architectural enhancements that dramatically increase performance, reduce power consumption, and shrink die area. When selecting a Speedcore eFPGA, designers can select the optimal mix of architectural elements including:
- Logic – 6-input look-up-tables (LUTs) plus integrated wide MUX functions and fast adders
- Logic RAM – 2 kb per memory block for LRAM2k, and 4kb per memory block for LRAM4k
- Block RAM – 72 kb per memory block for BRAM72k, and 20kb per memory block for BRAM20k
- DSP64 – 18 × 27 multiplier, 64-bit accumulator and 27-bit pre-adder per block
- Machine learning processors (MLP) – 32 multiplier/accumulators (MACs) per block, supporting integer and floating point formats
Reconfigurable Logic Blocks (RLB)
- Logic – 6-input look-up-tables (LUTs) that implement all functions with as many as 7-inputs and some 8-input functions in a single level of logic. Reducing the need for multiple logic levels improves performance.
- 8:1 Muxes – New, dedicated 8-to-1 multiplexers dramatically increase logic performance.
- Shift chain – Double the number of registers compared to the original Speedcore architecture plus optimized routing for shift chains.
- ALU – A larger ALU now supports 8-bit operations for addition, counting, comparison, and maximum functions.
- LUT-based multiplication – Efficient, LUT-based multipliers require half the on-chip resources compared to other leading FPGA products: A 6 × 6 multiply requires only 11 LUTs and runs at 1 GHz. An 8 × 8 multiply requires only 18 LUTs and runs at 500 MHz.
- The LRAM2k implements a 2,304-bit memory block configured as a 32 × 72 simple dual-port (one write port, one read port) RAM. The LRAM2k has a synchronous write port. The read port is configured for asynchronous read operations with an optional output register.
- The LRAM4k implements a 4,096-bit memory block configured as a 128 × 32 simple dual-port (one write port, one read port) RAM. The LRAM4k has a synchronous write port. The read port is configured for asynchronous read operations with an optional output register.
- The BRAM72k primitive implements a 72-kb simple-dual-port (SDP) memory block with one write port and one read port. Each port can be independently configured with respect to size and function, and can use independent read and write clocks. The BRAM72k can be configured as a simple dual port or ROM memory.
- The BRAM20k implements a dual-ported memory block where each port can be independently configured with respect to size and function. The BRAM20k can be configured as a single-port (one read/write port), dual-port (two read/write ports with independent clocks), or ROM memory.
- The DSP64 blocks include multiple/accumulate and associated logic to efficiently implement math functions such as finite impulse response (FIR) filters, fast Fourier transforms (FFT), and infinite impulse response (IIR) filters. The DSP64 blocks are optimized to operate with the logic fabric and LRAM blocks to implement math functions.
Machine Learning Processor (MLP)
The new MLP in Speedcore eFPGA IP is a complete AI/ML compute engine. Each MLP includes a cyclical register file that leverages temporal locality to reuse stored/cached weights or data, thus boosting performance by significantly reducing data movement for a variety of calculations. The MLPs are tightly coupled with their neighboring MLPs and larger memory blocks to maximize processing performance and to deliver the highest number of operations per second with the lowest power profile. The MLPs support fixed-point and floating-point formats (Bfloat16; 16-bit, half-precision; and block floating point). Users can trade off precision versus performance by selecting the optimal data precision on the fly, as required by each application.
|Configurable multiply precision and count||Trade off performance/power vs. precision - Increasing multiplier count for lower precision functions.|
|Cyclical register file||Double compute performance - Similar to a cache function in that data is saved for efficient reuse by the MLP.Optimized for AI/ML functions.|
|Column bonding and MLP cascade paths||
Higher performance - Hard paths between memory and other MLP blocks enable high-performance functionality while freeing up general-purpose routing.
|Multiple number formats||Flexibility - Supports mainstreams fixed- and floating-point formats and frameworks.|
|Rounding and saturation||System performance - Support for multiple rounding formats and saturation that would otherwise need to be implemented in LUTs.|
- Dedicated buses – A first in the eFPGA industry! High-performance, bus-grouped routing channels, separate from the standard eFPGA routing channels, ensure that there is no congestion between bus-oriented data traffic — common with memories — and other types of data traffic routed over the eFPGA’s standard, bit-oriented channels.
- Bus muxes – Another first in the eFPGA industry; bus muxes allow users to efficiently create bus mux functions without consuming any LUTs or standard routing. This capability effectively creates a giant, distributed, run-time-configurable switching network that is separate from the eFPGA’s bit-oriented routing network.