Pipelined Fused Multiply-Add Processing Element for FP4 Weight and FP8 Activation Arithmetic with FP8 Accumulation
OwnerKia Silverbrookinventor
Life of this patent
Estimating the term…
Loading patent…
OwnerKia Silverbrookinventor
Estimating the term…
A pipelined fused multiply-add processing element for mixed-precision arithmetic using FP4 weight operands and FP8 activation operands is disclosed. The element forms a product of the FP4 weight and FP8 activation and accumulates the result into an FP8 accumulator using bounded-carry logic. The datapath includes a latch-defined pipeline having a two-cycle latency while sustaining a one-cycle initiation interval, allowing one new fused multiply-add operation to be accepted each clock cycle. Weight operands follow an FP4 (E2M1) format within a shared exponent domain, permitting omission of per-element exponent alignment and rounding logic. Symmetric physical layout and deterministic timing enable multi-gigahertz operation with extremely small area, supporting dense arrays of mixed-precision arithmetic units for AI inference accelerators.
A fused multiply-add processing element comprising logic circuitry configured to form a fixed-scale product of an FP4 weight operand and an FP8 activation operand, an accumulator configured to combine the product with an FP8 accumulator value without intermediate rounding or normalization, and pipeline control circuitry configured to complete the multiply-add operation with a two-cycle pipeline latency while sustaining a one-cycle initiation interval, wherein the FP4 weight operand conforms to an E2M1 format having a shared exponent domain and the processing element omits per-element exponent alignment and rounding logic for the weight operand.
A method of performing fused multiply-add computation comprising receiving an FP4 weight operand in E2M1 format and an FP8 activation operand within a shared exponent domain, generating a mantissa contribution from combinational logic responsive to mantissa bits of the FP4 weight operand and the FP8 activation operand, generating a normalization carry from the mantissa bits, adding a weight exponent, an activation exponent, and the normalization carry to form an output exponent, forming a sign bit as a bitwise exclusive-or of a weight sign bit and an activation sign bit, and adding a corresponding product to an FP8 accumulator using bounded-carry accumulation in a pipelined datapath having a two-cycle latency and a one-cycle initiation interval.
Dependencies are inferred from the claim text.