IP Library Granted Patent US 12,436,896
Granted Patent B1
US 12,436,896 · App. 18/989,492 · Granted Oct 7, 2025

AI accelerator integrated circuit chip with integrated cell-based fabric adapter

Inventors: Gary S. Goldman (Los Altos, CA); Ramalingam K. Anand (Los Altos Hills, CA); Kalyana S. Venkataraman (San Jose, CA); Berend Ozceri (Los Gatos, CA); Pradeep R. Joginipally (San Jose, CA); Chung Y. Lau (Milpitas, CA); Jigar K. Savla (San Jose, CA); Ashwin Radhakrishnan (Fremont, CA); Michael Davie (St Augustine, FL); Shijun Li (Southborough, MA)
Assignee: Recogni Inc.
G06F12/1081G06F13/4068G06F13/409
View Patent ↗
Loading inventors, assignments & file history…
Monitor This Case
Get email alerts when status or documents change.
Order Certified Copies
Most orders are placed with the USPTO same day — all within 24 business hours.
Order via The Patent Place →
Pre-filled with this patent's details
Quick Facts
Patent No.
US 12,436,896
App. No.
18/989,492
Granted
Oct 7, 2025
Kind
B1
Abstract

An integrated circuit formed on (i) a single semiconductor die or (ii) a plurality semiconductor dies that are integrated into a single package. The integrated circuit may include a communication interface including a serializer/deserializer (SerDes) interface; a fabric adapter communicatively coupled to the communication interface; a plurality of inference engine clusters, each inference engine cluster including a respective memory element and/or memory interface; and a data interconnect communicatively coupling each respective memory element and/or memory interfaces of the plurality of inference engine clusters to the fabric adapter. The fabric adapter may be configured to facilitate remote direct memory access (RDMA) read and write services and/or datagram communication over a cell-based switch fabric to and from the respective memory elements and/or memory interfaces of the plurality of inference engine clusters via the data interconnect.

Claims (34)

1. An integrated circuit formed on (i) a single semiconductor die or (ii) a plurality semiconductor dies that are integrated into a single package, the integrated circuit comprising:

a communication interface;

a fabric adapter communicatively coupled to the communication interface;

a plurality of inference engine clusters, each inference engine cluster including a respective memory element; and

a data interconnect communicatively coupling each respective memory element of the plurality of inference engine clusters to the fabric adapter,

wherein the fabric adapter is configured to facilitate remote direct memory access (RDMA) read and write services over a cell-based switch fabric to and from the respective memory elements of the plurality of inference engine clusters via the data interconnect,

wherein the cell-based switch fabric comprises a plurality of cell-fabric switch chips that are disposed external to the integrated circuit, and

wherein the fabric adapter comprises a virtual output queue (VOQ) subsystem.

2. The integrated circuit of claim 1 , wherein the fabric adapter comprises a transmit fabric adapter configured to transmit data to the cell-based switch fabric and a receive fabric adapter configured to receive data from the cell-based switch fabric.

3. The integrated circuit of claim 2 , wherein the transmit fabric adapter comprises a transmit fabric interface adapter and a transmit fabric interface.

4. The integrated circuit of claim 2 , wherein the receive fabric adapter comprises a receive fabric interface and a receive fabric interface adapter.

5. The integrated circuit of claim 3 , wherein the transmit fabric interface adapter comprises a packet generator, a packet table, a transmit direct memory access (DMA) read engine, a transmit buffer, a packet header encapsulator, and a transmit cellifier.

6. The integrated circuit of claim 3 , wherein the transmit fabric interface comprises the VOQ subsystem and a fabric transmit data pipe.

7. The integrated circuit of claim 4 , wherein the receive fabric interface comprises a fabric receive data pipe and a fabric receive control.

8. The integrated circuit of claim 4 , wherein the receive fabric interface adapter comprises a receive cell buffer, a receive decellifier, a packet header decapsulator, a receive direct memory access (DMA) write engine, a receive packet linked list manager, an output queue (OQ) subsystem and a packet reader.

9. The integrated circuit of claim 3 , wherein the transmit fabric interface adapter comprises a transmit job manager, a packet generator, a packet table, a transmit direct memory access (DMA) read engine, a transmit buffer, a packet header encapsulator module, and the VOQ subsystem.

10. The integrated circuit of claim 3 , wherein the transmit fabric interface comprises a fabric transmit control and a fabric transmit data pipe.

11. The integrated circuit of claim 4 , wherein the receive fabric interface is configured to deliver completed packets to the receive fabric interface adapter.

12. The integrated circuit of claim 4 , wherein the receive fabric interface adapter comprises a packet header decapsulator, a receive direct memory access (DMA) write engine and a receive job manager.

13. The integrated circuit of claim 1 , wherein each inference engine cluster includes a plurality of inference engines, each of the inference engines configured to execute neural networks, including one or more of convolutional neural networks (CNN), recurrent neural network (RNN), transformers, or large language models.

14. The integrated circuit of claim 1 , wherein the data interconnect comprises one or more of a crossbar switch, a ring, a torus interconnect or a mesh.

15. An integrated circuit formed on (i) a single semiconductor die or (ii) a plurality semiconductor dies that are integrated into a single package, the integrated circuit comprising:

a communication interface;

a fabric adapter communicatively coupled to the communication interface;

a plurality of inference engine clusters, each inference engine cluster including a respective memory interface; and

a data interconnect communicatively coupling each respective memory interface of the plurality of inference engine clusters to the fabric adapter,

wherein the fabric adapter is configured to facilitate remote direct memory access (RDMA) read and write services over a cell-based switch fabric to and from the respective memory interfaces of the plurality of inference engine clusters via the data interconnect,

wherein the cell-based switch fabric comprises a plurality of cell-fabric switch chips that are disposed external to the integrated circuit, and

wherein the fabric adapter comprises a virtual output queue (VOQ) subsystem.

16. The integrated circuit of claim 15 , wherein the memory interface comprises a dynamic random access memory (DRAM) interface and controller.

17. The integrated circuit of claim 15 , wherein each of the inference engine clusters includes a memory cache.

18. The integrated circuit of claim 15 , wherein the fabric adapter comprises a transmit fabric adapter configured to transmit data to the cell-based switch fabric and a receive fabric adapter configured to receive data from the cell-based switch fabric.

19. The integrated circuit of claim 18 , wherein the transmit fabric adapter comprises a transmit fabric interface adapter and a transmit fabric interface.

20. The integrated circuit of claim 18 , wherein the receive fabric adapter comprises a receive fabric interface and a receive fabric interface adapter.

Assignments (11)
CHANGE OF NAME Recorded Sep 11, 2025
From: RECOGNI INC.
To: TENSORDYNE, INC.
Reel/Frame 072859/0058 →
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Jan 3, 2025
From: DAVIE, MICHAEL
To: RECOGNI INC.
Reel/Frame 069742/0366 →
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Jan 3, 2025
From: GOLDMAN, GARY S.
To: RECOGNI INC.
Reel/Frame 069742/0369 →
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Jan 3, 2025
From: JOGINIPALLY, PRADEEP R
To: RECOGNI INC.
Reel/Frame 069742/0376 →
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Jan 3, 2025
From: LAU, CHUNG Y.
To: RECOGNI INC.
Reel/Frame 069742/0379 →
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Jan 3, 2025
From: ANAND, RAMALINGAM K.
To: RECOGNI INC.
Reel/Frame 069742/0337 →
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Jan 3, 2025
From: OZCERI, BEREND
To: RECOGNI INC.
Reel/Frame 069742/0387 →
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Jan 3, 2025
From: SAVLA, JIGAR K.
To: RECOGNI INC.
Reel/Frame 069742/0396 →
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Jan 3, 2025
From: RADHAKRISHNAN, ASHWIN
To: RECOGNI INC.
Reel/Frame 069742/0399 →
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Jan 3, 2025
From: VENKATARAMAN, KALYANA S.
To: RECOGNI INC.
Reel/Frame 069742/0416 →
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Jan 3, 2025
From: LI, SHIJUN
To: RECOGNI INC.
Reel/Frame 069742/0384 →
Continuity (1)
Provisional Application 63694397 · Sep 13, 2024
References Cited (30)
US 8145785B1 · Finkelstein · 2012 [cited by examiner]
US 10721187B1 · Goldman · 2020 [cited by examiner]
US 11025544B2 · Marolia et al. · 2021 [cited by applicant]
US 11194753B2 · Marolia et al. · 2021 [cited by applicant]
US 11916800B2 · Arditti Ilitzky et al. · 2024 [cited by applicant]
US 20020006110A1 · Brezzo et al. · 2002 [cited by applicant]
US 20110219208A1 · Asaad et al. · 2011 [cited by applicant]
US 20130117766A1 · Bax · 2013 [cited by examiner]
US 20150055649A1 · DeCasatis et al. · 2015 [cited by applicant]
US 20190042518A1 · Marolia et al. · 2019 [cited by applicant]
US 20200304427A1 · Sandler et al. · 2020 [cited by applicant]
US 20200387564A1 · Simpson · 2020 [cited by examiner]
US 20200412659A1 · Arditti Ilitzky et al. · 2020 [cited by applicant]
US 20210117246A1 · Lal et al. · 2021 [cited by applicant]
US 20210117360A1 · Kutch et al. · 2021 [cited by applicant]
US 20210334184A1 · Menon · 2021 [cited by examiner]
US 20220276973A1 · Wei et al. · 2022 [cited by applicant]
US 20230231811A1 · Dalal · 2023 [cited by applicant]
“AI Networking”, ARISTA, White Paper, Jun. 5, 2024, 14 pgs. [cited by applicant]
Biglari; et al., “Designing Reconfigurable Interconnection Network of Heterogeneous Chiplets Using Kalman Filter”, Cornell University, arXiv:2406.00568v1 [cs.AR] Jun. 1, 2024, 6 pgs. [cited by applicant]
Gao; et al., “Customized High Performance and Energy Efficient Communication Networks for AI Chips”, IEEE Access, May 17, 2019 (revised Jun. 10, 2019), vol. 7, 69434-69446. [cited by applicant]
Mabavinejad; et al., “An Overview of Efficient Interconnection Networks for Deep Neural Network Accelerators”, IEEE Journal on Emerging and Selected Topics in Circuits and Systems, Sep. 2020, 10(3):268-282. [cited by applicant]
Rettkowski, Jens, M.Sc., “Design and Programming Methods for Reconfigurable Multi-Core Architectures using a Network-on-Chip-Centric Approach”, Technische Universitat Dresden, dissertation, Sep. 9, 2021, pp. 1-246. [cited by applicant]
“Season 3 Ep 11: Cell-based AI Fabric”, DriveNets, CloudNets Video accessed Jul. 1, 2024, transcription, 4 pgs. [cited by applicant]
Final Office Action dated Jul. 2, 2025, for U.S. Appl. No. 18/989,525, filed Dec. 20, 2024, 23 pgs. [cited by applicant]
Notice of Allowance mailed Jul. 22, 2025, for U.S. Appl. No. 18/989,508, filed Dec. 20, 2024, 9 pgs. [cited by applicant]
Non-Final Office Action dated May 2, 2025, for U.S. Appl. No. 18/989,508, filed Dec. 20, 2024, 19 pgs. [cited by applicant]
Amendment filed May 27, 2025, for U.S. Appl. No. 18/989,525, filed Dec. 20, 2024, 10 pgs. [cited by applicant]
Amendment filed Jul. 3, 2025, for U.S. Appl. No. 18/989,508, filed Dec. 20, 2024, 9 pgs. [cited by applicant]
Non-Final Office Action dated Feb. 27, 2025, for U.S. Appl. No. 18/989,525, filed Dec. 20, 2024, 26 pgs. [cited by applicant]
Cited By (1)
US 12,664,413