# US 2026/0203018 A1 · Pipelined Fused Multiply-Add Processing Element for FP4 Weight and FP8 Activation Arithmetic with FP8 Accumulation

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2026/0203018 A1
- Document ID: US20260203018A1
- Status: Published application
- Kind: A1
- Application number: 19/433,862
- Filed: 2025-12-28
- Published: 2026-07-16
- Earliest claimed: 2024-12-29
- Inventors: Kia Silverbrook
- Web page: https://patentgrep.com/patent/US20260203018A1

## CPC

- H10D 89/10: Integrated device layouts
- B05B 1/34: designed to influence the nature of flow of the liquid or other fluent material, e.g. to produce swirl (designed to control volume of flow B05B1/30)
- G01R 1/06733: {Geometry aspects (G01R1/06727 takes precedence)}
- G01R 1/06794: {Devices for sensing when probes are in contact, or in position to contact, with measured object}
- G01R 1/07378: {using an intermediate adapter, e.g. space transformers (G01R1/07371 takes precedence)}
- G01R 31/2886: {Features relating to contacting the IC under test, e.g. probe heads; chucks (G01R31/2865 takes precedence, test connections, e.g. test sockets, or probes per se, G01R1/04 or G01R1/06)}
- G01R 31/52: Testing for short-circuits, leakage current or ground faults
- G01R 31/54: Testing for continuity
- G06E 3/006: {Interconnection networks, e.g. for shuffling}
- G06F 1/04: Generating or distributing clock signals or signals derived directly therefrom
- G06F 1/06: Clock generators producing several clock signals {(G06F1/08 - G06F1/14 take precedence)}
- G06F 1/08: Clock generators with changeable or programmable clock frequency
- G06F 1/10: Distribution of clock signals {, e.g. skew}
- G06F 1/20: Cooling means
- G06F 1/206: {comprising thermal management}
- G06F 1/26: Power supply means, e.g. regulation thereof (for memories G11C)
- G06F 1/3203: Power management, i.e. event-based initiation of a power-saving mode
- G06F 1/3237: by disabling clock generation or distribution
- G06F 1/324: by lowering clock frequency
- G06F 1/3296: by lowering the supply or operating voltage
- G06F 7/483: Computations with numbers represented by a non-linear combination of denominational numbers, e.g. rational numbers, logarithmic number system or floating-point numbers {(G06F7/4806, G06F7/4824, G06F7/49, G06F7/491, G06F7/544 take precedence)}
- G06F 7/485: Adding; Subtracting {(G06F7/4833, G06F7/4836 take precedence)}
- G06F 7/4876: {Multiplying}
- G06F 7/49915: {Mantissa overflow or underflow in handling floating-point numbers}
- G06F 7/49942: {Significance control}
- G06F 7/5443: {Sum of products (for applications thereof, see the relevant places, e.g. G06F17/10, H03H17/00)}
- G06F 8/76: Adapting program code to run in a different environment; Porting
- G06F 9/30014: {with variable precision}
- G06F 9/30174: {for non-native instruction set, e.g. Javabyte, legacy code}
- G06F 9/3869: {Implementation aspects, e.g. pipeline latches; pipeline synchronisation and clocking}
- G06F 9/3887: {controlled by a single instruction for multiple data lanes [SIMD]}
- G06F 9/5027: {the resource being a machine, e.g. CPUs, Servers, Terminals}
- G06F 9/5044: {considering hardware capabilities}
- G06F 9/547: {Remote procedure calls [RPC]; Web services}
- G06F 11/1423: {by reconfiguration of paths}
- G06F 11/202: {where processing functionality is redundant (redundant communication control functionality G06F11/2005, redundant storage control functionality G06F11/2089)}
- G06F 11/2028: {eliminating a faulty processor or activating a spare}
- G06F 11/2041: {with more than one idle spare processing component}
- G06F 11/2051: {in regular structures}
- G06F 11/2053: {where persistent mass storage functionality or persistent mass storage control functionality is redundant (error detection or correction in information storage based on relative movement between record carrier and transducer G11B20/18)}
- G06F 11/2242: {in multi-processor systems, e.g. one processor becoming the primary tester (G06F11/2736 takes precedence)}
- G06F 11/2247: {Verification or detection of system hardware configuration}
- G06F 11/3024: {where the computing system component is a central processing unit [CPU]}
- G06F 11/328: {Computer systems status display (G06F11/327 takes precedence)}
- G06F 12/10: Address translation
- G06F 13/368: with decentralised access control
- G06F 13/385: {for adaptation of a particular data processing system to different peripheral devices}
- G06F 13/4022: {using switching circuits, e.g. switching matrix, connection or expansion network (G06F13/4009 takes precedence)}
- G06F 13/4027: {using bus bridges (G06F13/4022 takes precedence)}
- G06F 13/4068: {Electrical coupling}
- G06F 13/4295: {using an embedded synchronisation}
- G06F 15/17325: {Synchronisation; Hardware support therefor (intertask synchronisation G06F9/52)}
- G06F 15/17387: {Three dimensional, e.g. hypercubes}
- G06F 15/17393: {having multistage networks, e.g. broadcasting scattering, gathering, hot spot contention, combining/decombining}
- G06F 15/7817: {Specially adapted for signal processing, e.g. Harvard architectures}
- G06F 15/7867: {with reconfigurable architecture}
- G06F 15/80: comprising an array of processing units with common control, e.g. single instruction multiple data processors (G06F15/82 takes precedence {; for correlation function computation G06F17/15})
- G06F 15/8015: {One dimensional arrays, e.g. rings, linear arrays, buses}
- G06F 15/8023: {Two dimensional arrays, e.g. mesh, torus}
- G06F 15/8046: {Systolic arrays}
- G06F 17/16: Matrix or vector computation {, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization (matrix transposition G06F7/78)}
- G06F 30/27: using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
- G06F 30/323: Translation or migration, e.g. logic to logic, hardware description language [HDL] translation or netlist translation
- G06F 30/392: Floor-planning or layout, e.g. partitioning or placement
- G06F 30/398: Design verification or optimisation, e.g. using design rule check [DRC], layout versus schematics [LVS] or finite element methods [FEM] (optical proximity correction [OPC] design processes G03F1/36)
- G06N 3/04: Architecture, e.g. interconnection topology
- G06N 3/044: Recurrent networks, e.g. Hopfield networks
- G06N 3/0442: characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
- G06N 3/045: Combinations of networks
- G06N 3/0455: Auto-encoder networks; Encoder-decoder networks
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/047: Probabilistic or stochastic networks
- G06N 3/048: Activation functions
- G06N 3/063: using electronic means
- G06N 3/065: Analogue means
- G06N 3/0675: {using electro-optical, acousto-optical or opto-electronic means}
- G06N 3/08: Learning methods
- G06N 3/082: modifying the architecture, e.g. adding, deleting or silencing nodes or connections
- G06N 3/084: Backpropagation, e.g. using gradient descent
- G06N 3/088: Non-supervised learning, e.g. competitive learning
- G06N 3/09: Supervised learning
- G06N 3/098: Distributed learning, e.g. federated learning
- G06N 3/105: {Shells for specifying net layout}
- G06N 5/01: Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound
- G06N 5/04: Inference or reasoning models
- G06N 7/01: Probabilistic graphical models, e.g. probabilistic networks
- G06N 10/00: Quantum computing, i.e. information processing based on quantum-mechanical phenomena
- G06N 10/40: Physical realisations or architectures of quantum processors or components for manipulating qubits, e.g. qubit coupling or qubit control
- G06N 10/60: Quantum algorithms, e.g. based on quantum optimisation, quantum Fourier or Hadamard transforms
- G06N 20/00: Machine learning
- G06N 20/10: using kernel methods, e.g. support vector machines [SVM]
- G11C 5/025: {Geometric lay-out considerations of storage- and peripheral-blocks in a semiconductor storage device (geometrical lay-out of the components in integrated circuits, geometrical lay-out of the components in integrated circuits H10D89/10)}
- G11C 5/04: Supports for storage elements {, e.g. memory modules}; Mounting or fixing of storage elements on such supports
- G11C 7/1039: {using pipelining techniques, i.e. using latches between functional memory parts, e.g. row/column decoders, I/O buffers, sense amplifiers}
- G11C 7/106: {Data output latches}
- G11C 7/222: {Clock generating, synchronizing or distributing circuits within memory device}
- G11C 29/18: Address generation devices; Devices for accessing memories, e.g. details of addressing circuits
- G11C 29/44: Indication or identification of errors, e.g. for repair
- G11C 29/4401: {for self repair}
- G11C 29/46: Test trigger logic
- G11C 29/702: {by replacing auxiliary circuits, e.g. spare voltage generators, decoders or sense amplifiers, to be used instead of defective ones}
- G11C 29/808: {using a flexible replacement scheme}
- G11C 29/816: {for an application-specific layout}
- G11C 29/848: {by adjacent switching}
- H02H 3/021: {concerning the disconnection itself, e.g. at a particular instant, particularly at zero value of current, disconnection in a predetermined order (disconnection at zero value in general H03K17/18)}
- H02H 3/027: with automatic disconnection after a predetermined time (H02H3/033, H02H3/06 take precedence {; timing in overcurrent protection circuits H02H3/093; in undervoltage protection circuits H02H3/247; staggered disconnection H02H7/30})
- H02H 3/05: with means for increasing reliability, e.g. redundancy arrangements {(for logic circuits H03K19/003)}
- H02M 3/1566: {with means for compensating against rapid load changes, e.g. with auxiliary current source, with dual mode control or with inductance variation}
- H03K 19/17736: Structural details of routing resources
- H04L 9/0819: {Key transport or distribution, i.e. key establishment techniques where one party creates or otherwise obtains a secret value, and securely transfers it to the other(s) (network architectures or network communication protocols for key distribution in a packet data network H04L63/062)}
- H04L 9/0825: {using asymmetric-key encryption or public key infrastructure [PKI], e.g. key signature or public key certificates}
- H04L 9/0852: {Quantum cryptography (transmission systems employing electromagnetic waves other than radio waves, e.g. light, infrared H04B10/00; wavelength-division multiplex systems H04J14/02; WDM arrangements H04J14/03)}
- H04L 9/0866: {involving user or device identifiers, e.g. serial number, physical or biometrical information, DNA, hand-signature or measurable physical characteristics}
- H04L 9/0891: {Revocation or update of secret information, e.g. encryption key update or rekeying}
- H04L 9/3247: {involving digital signatures}
- H04L 9/3271: {using challenge-response}
- H04L 9/50: {using hash chains, e.g. blockchains or hash trees}
- H05K 1/0203: {Cooling of mounted components (H05K1/0272 takes precedence)}
- H05K 1/0213: {Electrical arrangements not otherwise provided for}
- H05K 1/0245: {Lay-out of balanced signal pairs, e.g. differential lines or twisted lines}
- H05K 1/0262: {Arrangements for regulating voltages or for using plural voltages}
- H05K 1/0271: {Arrangements for reducing stress or warp in rigid printed circuit boards, e.g. caused by loads, vibrations or differences in thermal expansion}
- H05K 1/0298: {Multilayer circuits}
- H05K 1/0306: {Inorganic insulating substrates, e.g. ceramic, glass}
- H05K 1/09: Use of materials for the {conductive, e.g. } metallic pattern
- H05K 1/115: {Via connections; Lands around holes or via connections (H05K1/112 takes precedence)}
- H05K 1/144: {Stacked arrangements of planar printed circuit boards}
- H05K 3/0041: {by plasma etching}
- H05K 3/0076: {characterised by the composition of the mask}
- H05K 3/0094: {Filling or covering plated through-holes or blind plated vias, e.g. for masking or for mechanical reinforcement}
- H05K 3/26: Cleaning or polishing of the conductive pattern
- H05K 3/30: Assembling printed circuits with electric components, e.g. with resistors
- H05K 3/3426: {characterised by the leads}
- H05K 3/3436: {having an array of bottom contacts, e.g. pad grid array or ball grid array components}
- H05K 3/361: {Assembling flexible printed circuits with other printed circuits}
- H05K 3/4038: {Through-connections; Vertical interconnect access [VIA] connections (H05K3/403, H05K3/42 take precedence)}
- H05K 3/4046: {using auxiliary conductive elements, e.g. metallic spheres, eyelets, pieces of wire}
- H05K 3/423: {characterised by electroplating method}
- H05K 7/026: {Multiple connections subassemblies}
- H05K 7/20327: {Accessories for moving fluid, for connecting fluid conduits, for distributing fluid or for preventing leakage, e.g. pumps, tanks or manifolds}
- H05K 7/20345: {Sprayers; Atomizers}
- H05K 7/20809: {within server blades for removing heat from heat source}
- H05K 7/20836: {Thermal management, e.g. server temperature control}
- H10B 10/12: {comprising a MOSFET load element}
- H10B 80/00: Assemblies of multiple devices comprising at least one memory device covered by this subclass
- H10D 80/251: {FETs covered by H10D30/00, e.g. power FETs}
- H10D 80/30: the at least one device being covered by groups H10D84/00 - H10D86/00, e.g. assemblies comprising integrated circuit processor chips
- H10D 84/968: {Macro-architecture}
- H10D 84/974: {Layout specifications, i.e. inner core regions}
- H10D 88/00: Three-dimensional [3D] integrated devices
- H10P 14/412: {Deposition of metallic or metal-silicide materials}
- H10P 50/244: {comprising alternated and repeated etching and passivation steps}
- H10P 72/74: {using temporarily an auxiliary support}
- H10P 72/7416: {used during dicing or grinding}
- H10P 72/744: {Details of chemical or physical process used for separating the auxiliary support from a device or a wafer}
- H10P 74/203: {Structural properties, e.g. testing or measuring thicknesses, line widths, warpage, bond strengths or physical defects}
- H10P 74/207: {Electrical properties, e.g. testing or measuring of resistance, deep levels or capacitance-voltage characteristics}
- H10P 74/232: {comprising connection or disconnection of parts of a device in response to a measurement}
- H10P 74/273: {Interconnections for measuring or testing, e.g. probe pads}
- H10P 74/277: {Circuits for electrically characterising or monitoring manufacturing processes, e.g. circuits in tested chips or circuits in testing wafers}
- H10W 10/10: Isolation regions comprising dielectric materials
- H10W 20/023: {the interconnections being through-semiconductor vias}
- H10W 20/0245: {comprising use of blind vias during the manufacture}
- H10W 20/20: Interconnections within wafers or substrates, e.g. through-silicon vias [TSV]
- H10W 20/211: {Through-semiconductor vias, e.g. TSVs}
- H10W 20/212: {Top-view shapes or dispositions, e.g. top-view layouts of the vias}
- H10W 20/213: {Cross-sectional shapes or dispositions}
- H10W 20/42: Vias, e.g. via plugs
- H10W 20/427: {Power or ground buses}
- H10W 20/43: Layouts of interconnections
- H10W 20/481: {on the rear surfaces of the wafers or substrates}
- H10W 40/22: characterised by their shape, e.g. having conical or cylindrical projections
- H10W 40/226: {characterised by projecting parts, e.g. fins to increase surface area (leadframes for cooling H10W70/461)}
- H10W 40/228: {the projecting parts being wire-shaped or pin-shaped}
- H10W 40/235: {attached to package parts}
- H10W 40/242: {comprising thermal conductors between chips and the and the arrangements for cooling, e.g. compliant heat-spreaders}
- H10W 40/253: {Semiconductors}
- H10W 40/257: {having a heterogeneous or anisotropic structure, e.g. powder or fibres in a matrix, wire mesh or porous structures (H10W40/254, H10W40/251 take precedence)}
- H10W 40/30: wherein the packaged device is completely immersed in a fluid other than air, e.g. immersed in a cryogenic fluid
- H10W 40/40: involving heat exchange by flowing fluids
- H10W 40/43: by flowing gases, e.g. forced air cooling
- H10W 40/47: by flowing liquids, e.g. forced water cooling
- H10W 40/73: for cooling by change of state
- H10W 40/776: {Arrangements for jet impingement, e.g. for spraying}
- H10W 42/00: Arrangements for protection of devices (arrangements for thermal protection H10W40/00)
- H10W 42/121: {protecting against mechanical damage (H10W76/00, H10W74/00 take precedence)}
- H10W 44/501: {Inductive arrangements (H10W44/20 takes precedence)}
- H10W 70/095: {of vias therein}
- H10W 70/60: Insulating or insulated package substrates; Interposers; Redistribution layers (leadframes H10W70/40)
- H10W 70/611: {for connecting multiple chips together}
- H10W 70/63: Vias, e.g. via plugs
- H10W 70/635: {Through-vias}
- H10W 70/644: {Elastic or compliant interconnections, e.g. springs, cantilevers or elastic pads}
- H10W 70/652: Cross-sectional shapes
- H10W 70/655: Fan-out layouts
- H10W 70/66: Conductive materials thereof
- H10W 70/68: Shapes or dispositions thereof
- H10W 70/685: comprising multiple insulating layers
- H10W 70/692: Ceramics or glasses
- H10W 70/698: Semiconductor materials that are electrically insulating, e.g. undoped silicon
- H10W 72/00: Interconnections or connectors in packages
- H10W 72/01: {Manufacture or treatment}
- H10W 72/072: {of bump connectors}
- H10W 72/073: {of die-attach connectors}
- H10W 72/20: Bump connectors, e.g. solder bumps or copper pillars; Dummy bumps; Thermal bumps
- H10W 72/227: {Multiple bumps having different sizes}
- H10W 72/241: {Dispositions, e.g. layouts}
- H10W 72/247: {Dispositions of multiple bumps}
- H10W 72/252: {comprising solid metals or solid metalloids, e.g. PbSn, Ag or Cu}
- H10W 72/29: Bond pads specially adapted therefor
- H10W 72/823: {Interconnections through encapsulations, e.g. pillars through molded resin on a lateral side a chip}
- H10W 74/012: {of encapsulations on active surfaces of flip-chip devices, e.g. forming underfills}
- H10W 74/15: on active surfaces of flip-chip devices, e.g. underfills
- H10W 74/476: {Organic materials comprising silicon}
- H10W 80/00: Direct bonding of chips, wafers or substrates
- H10W 80/312: {characterised by the direct bonding of electrically conductive pads}
- H10W 80/314: {characterized by direct bonding of pads or other interconnections}
- H10W 80/327: {characterised by the direct bonding of insulating parts, e.g. of silicon oxide layers}
- H10W 90/00: Package configurations
- H10W 90/10: Configurations of laterally-adjacent chips
- H10W 90/20: Configurations of stacked chips
- H10W 90/22: the stacked chips being on both top and bottom sides of a package substrate, interposer or RDL
- H10W 90/26: the stacked chips being of the same size without any chips being laterally offset, e.g. chip stacks having a rectangular shape
- H10W 90/284: {characterised by structural arrangements for measuring or testing}
- H10W 90/288: {characterised by arrangements for thermal management of the stacked chips}
- H10W 90/291: {characterised by containers, encapsulations, or other housings for the stacked chips}
- H10W 90/297: {characterised by the through-semiconductor vias [TSVs] in the stacked chips}
- H10W 90/401: {characterised by multiple insulating or insulated package substrates, interposers or RDLs}
- H10W 90/701: {characterised by the relative positions of pads or connectors relative to package parts}
- H10W 90/722: {between stacked chips}
- H10W 90/724: {between a chip and a stacked insulating package substrate, interposer or RDL}
- H10W 90/734: {between a chip and a stacked insulating package substrate, interposer or RDL}
- H10W 90/792: {between multiple chips}
- H10W 90/794: {between a chip and a stacked insulating package substrate, interposer or RDL}
- B33Y 10/00: Processes of additive manufacturing
- B33Y 80/00: Products made by additive manufacturing
- G06F 7/49936: {Normalisation mentioned as feature only}
- G06F 2119/18: Manufacturability analysis or optimisation for manufacturability
- G06F 2200/201: Cooling arrangements using cooling fluid
- G06F 2201/805: Real-time
- G06F 2201/84: Using snapshots, i.e. a logical point-in-time copy of the data
- G06F 2209/509: Offload
- G11C 2029/0409: {Online test}
- G11C 2029/1806: {Address conversion or mapping, i.e. logical to physical address}
- H02M 3/003: {Constructional details, e.g. physical layout, assembly, wiring or busbar connections}
- H05K 1/181: associated with surface mounted components
- H05K 2201/0338: Layered conductor, e.g. layered metal substrate, layered finish layer or layered thin film adhesion layer
- H05K 2201/042: Stacked spaced PCBs; Planar parts of folded flexible circuits having mounted components in between or spaced from each other
- H05K 2201/064: Fluid cooling, e.g. by integral pipes
- H05K 2201/09063: Holes or slots in insulating substrate not used for electrical connections
- H05K 2201/09227: Layout details of a plurality of traces, e.g. escape layout for Ball Grid Array [BGA] mounting
- H05K 2201/0979: Redundant conductors or connections, i.e. more than one current path between two points
- H05K 2201/09854: Hole or via having special cross-section, e.g. elliptical
- H05K 2201/10159: Memory
- H05K 2201/10189: Non-printed connector
- H05K 2201/10265: Metallic coils or springs, e.g. as part of a connection element
- H05K 2201/10742: Details of leads
- H05K 2201/2009: Reinforced areas, e.g. for a specific part of a flexible printed circuit
- H05K 2203/016: Temporary inorganic, non-metallic carrier, e.g. for processing or transferring
- H05K 2203/0723: Electroplating, e.g. finish plating
- H05K 2203/1194: Thermal treatment leading to a different chemical state of a material, e.g. annealing for stress-relief, aging
- H05K 2203/143: Treating holes before another process, e.g. coating holes before coating the substrate
- H05K 2203/1572: Processing both sides of a PCB by the same process; Providing a similar arrangement of components on both sides; Making interlayer connections from two sides
- H05K 2203/162: Testing a finished product, e.g. heat cycle testing of solder joints
- H05K 2203/163: Monitoring a manufacturing process
- H10W 40/475: {using jet impingement (H10W40/776 takes precedence)}
- H10W 40/50: Arrangements for sensing temperature
- H10W 72/07236: {Soldering or alloying}
- H10D: INORGANIC ELECTRIC SEMICONDUCTOR DEVICES
- B05B: SPRAYING APPARATUS; ATOMISING APPARATUS; NOZZLES (spray-mixers with nozzles B01F25/72; processes for applying liquids or other fluent materials to surfaces by spraying B05D)
- G01R: MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES (indicating correct tuning of resonant circuits H03J3/12)
- G06E: OPTICAL COMPUTING DEVICES (digital storage using optical elements G11C13/04)
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G11C: STATIC STORES (semiconductor memory devices H10B)
- H02H: EMERGENCY PROTECTIVE CIRCUIT ARRANGEMENTS (indicating or signalling undesired working conditions G01R, e.g. G01R31/00, G08B; locating faults along lines G01R31/08; emergency protective devices H01H)
- H02M: APPARATUS FOR CONVERSION BETWEEN AC AND AC, BETWEEN AC AND DC, OR BETWEEN DC AND DC, AND FOR USE WITH MAINS OR SIMILAR POWER SUPPLY SYSTEMS; CONVERSION OF DC OR AC INPUT POWER INTO SURGE OUTPUT POWER; CONTROL OR REGULATION THEREOF (transformers H01F; dynamo-electric converters H02K47/00; controlling transformers, reactors or choke coils, control or regulation of electric motors, generators or dynamo-electric converters H02P)
- H03K: PULSE TECHNIQUE (measuring pulse characteristics G01R; modulating sinusoidal oscillations with pulses H03C; transmission of digital information H04L; discriminator circuits detecting phase difference between two signals by counting or integrating cycles of oscillation H03D3/04; automatic control, starting, synchronisation or stabilisation of generators of electronic oscillations or pulses where the type of generator is irrelevant or unspecified H03L; coding, decoding or code conversion, in general H03M)
- H04L: TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION (arrangements common to telegraphic and telephonic communication H04M)
- H05K: PRINTED CIRCUITS; CASINGS OR CONSTRUCTIONAL DETAILS OF ELECTRIC APPARATUS; MANUFACTURE OF ASSEMBLAGES OF ELECTRICAL COMPONENTS
- H10B: ELECTRONIC MEMORY DEVICES
- H10P: GENERIC PROCESSES OR APPARATUS FOR THE MANUFACTURE OR TREATMENT OF DEVICES COVERED BY CLASS H10
- H10W: GENERIC PACKAGES, INTERCONNECTIONS, CONNECTORS OR OTHER CONSTRUCTIONAL DETAILS OF DEVICES COVERED BY CLASS H10
- B33Y: ADDITIVE MANUFACTURING, i.e. MANUFACTURING OF THREE-DIMENSIONAL [3D] OBJECTS BY ADDITIVE DEPOSITION, ADDITIVE AGGLOMERATION OR ADDITIVE LAYERING, e.g. BY 3D PRINTING, STEREOLITHOGRAPHY OR SELECTIVE LASER SINTERING

## Abstract

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.

## Claims

### Claim 1

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.

### Claim 2 (depends on 1)

The processing element of claim 1 wherein the logic circuitry comprises a mantissa network producing a mantissa contribution derived from a bitwise exclusive-or of mantissa bits of the FP4 weight operand and the FP8 activation operand and a normalization carry derived from a bitwise and of selected mantissa bits, the normalization carry serving as a carry-in to an exponent adder.

### Claim 3 (depends on 1)

The processing element of claim 1 wherein an output sign bit is formed as a bitwise exclusive-or of a sign bit of the FP4 weight operand and a sign bit of the FP8 activation operand and an output exponent equals a sum of a weight exponent, an activation exponent, and a normalization carry.

### Claim 4 (depends on 1)

The processing element of claim 1 wherein the exponent adder comprises a compact binary adder with a single carry-in input and overflow saturation logic to constrain the exponent range.

### Claim 5 (depends on 1)

The processing element of claim 1 wherein the accumulator comprises an eight-bit adder employing bounded-carry partial-sum logic to complete accumulation within a timing budget for one pipeline stage.

### Claim 6 (depends on 1)

The processing element of claim 1 wherein a latch boundary between combinational logic regions defines a two-stage pipeline through which fused multiply-add operations propagate.

### Claim 7 (depends on 1)

The processing element of claim 1 wherein symmetric physical layout is employed to equalize propagation delay paths across the logic circuitry and the accumulator.

### Claim 8 (depends on 1)

The processing element of claim 1 wherein the processing element sustains a throughput of one fused multiply-add operation per clock cycle at multi-gigahertz clock frequencies.

### Claim 9 (depends on 1)

The processing element of claim 1 wherein sign and mantissa computations share a common logic path through dual-polarity encoding responsive to a sign bit of the FP4 weight operand and a sign bit of the FP8 activation operand.

### Claim 10 (depends on 1)

The processing element of claim 1 further comprising a bias-control device coupled to a local supply node of arithmetic circuitry and configured to modulate threshold voltage responsive to switching activity to maintain timing margins.

### Claim 11 (depends on 1)

The processing element of claim 1 wherein local clock and data-ready signaling are conveyed over a single conductor using a phase or edge polarity encoding to indicate timing of valid results.

### Claim 12 (depends on 1)

The processing element of claim 1 wherein a single clock edge advances the pipeline and captures results of a fused multiply-add computation.

### Claim 13 (depends on 1)

The processing element of claim 1 wherein the logic circuitry and accumulator are fabricated using standard CMOS devices without multi-phase clocks or dynamic logic.

### Claim 14 (depends on 1)

The processing element of claim 1 wherein reception of FP4 weight operands and FP8 activation operands and output of accumulated FP8 results occur under fixed-latency synchronization across a plurality of processing elements.

### Claim 15 (depends on 1)

The processing element of claim 1 wherein the logic circuitry and accumulator together comprise on the order of hundreds of transistors to enable dense tiling of the processing elements.

### Claim 16 (depends on 1)

The processing element of claim 1 wherein deterministic propagation delay is maintained by geometric symmetry of routing and device placement.

### Claim 17 (depends on 1)

A computing device comprising a plurality of the processing elements of claim 1 arranged in an array and a controller configured to supply FP4 weight operands within the shared exponent domain and FP8 activation operands to the array for parallel execution of pipelined fused multiply-add operations.

### Claim 18 (depends on 17)

A computing device according to claim 17 further comprising a broadcast network configured to distribute FP8 activation operands in parallel to multiple processing elements while stationary FP4 weight operands are held locally within the processing elements.

### Claim 19

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.

### Claim 20 (depends on 19)

The method of claim 19 further comprising operating the processing element at multi-gigahertz frequencies using a symmetric layout and latch-defined pipeline boundary to maintain deterministic propagation delay across process, voltage, and temperature variations.
