Javier D. Bruguera
Javier D. Bruguera has a total of 16 co-authors. Here are the top co-authors by number of shared publications:
- Elisardo Antelo
University of Santiago de Compostela
3 shared publications E.L Zapata
Dept. Arquitectura de Computadores, Fac. Informática., Univ. Málaga, 29013 Málaga, Spain
2 shared publications- Francisco Fernandez Rivera
Centro Singular de Investigación en Tecnoloxías Intelixentes (CiTIUS), University of Santiago de Compostela, 15782 Santiago de Compostela, Spain
2 shared publications - Ramón Doallo
University of A Coruña, Spain
2 shared publications - Anisha Saini
ARM Ltd., Austin, TX, USA
1 shared publication - David Exposito Singh
Universidad Carlos III de Madrid
1 shared publication
Latest Publications
- 25 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computers
- Vol. 75 (3) , 1171-1182
Fused FP8 Many-Terms Dot Product With Scaling and FP32 Accumulation
In the era of Deep Learning, hardware acceleration has become essential for meeting the immense computational demands of modern applications. In many Machine Learning applications, Generalized Matrix Multiplication (GEMM) with dot product is an ubiquitous and computationally intensive operation. This paper introduces two innovative microarchitectures for executing a fused FP8$m$ ‐way dot product with dynamic range scaling and FP32 accumulation. Both microarchitectures have been synthesized in a 3 nm technology node at 3.6 GHz, and were designed to deliver power- and area-efficient performance, targeting a 4-cycle latency for$m = 4,8$ and 5+ cycles for larger$m$ values. The first design – termed dot product with late accumulation – computes the dot product in the first cycles, then expands intermediate products to a fixed-point format (2 cycles for$m = 4, 8$ and 3+ cycles for$m \gt 8$ ), before using an additional two cycles for accumulation. This approach enables the reuse of a modified, FMA-capable FP32 adder. The second design – dot product with early accumulation – employs a dedicated FP8 datapath that concurrently computes the FP8 sum-of-products while aligning the FP32 accumulator, followed by the addition of the significands (2 cycles for$m = 4, 8$ and 3+ cycles for$m \gt 8$ ). This is then followed by two cycles for normalization and a single rounding operation. This design aligns addends (products and accumulator) from an “anchor” for efficient, arithmetically fused,$m$ -way FP8 dot product computation. Comparative analysis with previous proposals reveals that, despite challenges in establishing a fair comparison, our designs achieve significant area savings.- 25 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computers
- Vol. 75 (3) , 1171-1182
Fused FP8 Many-Terms Dot Product With Scaling and FP32 Accumulation
In the era of Deep Learning, hardware acceleration has become essential for meeting the immense computational demands of modern applications. In many Machine Learning applications, Generalized Matrix Multiplication (GEMM) with dot product is an ubiquitous and computationally intensive operation. This paper introduces two innovative microarchitectures for executing a fused FP8$m$ ‐way dot product with dynamic range scaling and FP32 accumulation. Both microarchitectures have been synthesized in a 3 nm technology node at 3.6 GHz, and were designed to deliver power- and area-efficient performance, targeting a 4-cycle latency for$m = 4,8$ and 5+ cycles for larger$m$ values. The first design – termed dot product with late accumulation – computes the dot product in the first cycles, then expands intermediate products to a fixed-point format (2 cycles for$m = 4, 8$ and 3+ cycles for$m \gt 8$ ), before using an additional two cycles for accumulation. This approach enables the reuse of a modified, FMA-capable FP32 adder. The second design – dot product with early accumulation – employs a dedicated FP8 datapath that concurrently computes the FP8 sum-of-products while aligning the FP32 accumulator, followed by the addition of the significands (2 cycles for$m = 4, 8$ and 3+ cycles for$m \gt 8$ ). This is then followed by two cycles for normalization and a single rounding operation. This design aligns addends (products and accumulator) from an “anchor” for efficient, arithmetically fused,$m$ -way FP8 dot product computation. Comparative analysis with previous proposals reveals that, despite challenges in establishing a fair comparison, our designs achieve significant area savings.- 25 May 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computers
- Vol. 72 (10) , 2990-3001
Radix-64 Floating-Point Division and Square Root: Iterative and Pipelined Units
Digit-recurrence algorithms are widely used in actual microprocessors to compute floating-point division and square root. These iterative algorithms present a good trade-off in terms of performance, area and power. Traditionally, commercial processors have iterative division and square root units where the iteration logic is used over several cycles. The main drawbacks of these iterative units are long latency and low throughput due to the reuse of part of the logic over several cycles, and its hardware complexity with separated logic for division and square root. We present a radix-64 floating-point division and square root algorithm with a common iteration for division and square root and where, to have an affordable implementation, each radix-64 iteration is made of two simpler radix-8 iterations. The radix-64 algorithm allows to get low-latency operations, and the common division and square root radix-64 iteration results in some area reduction. The algorithm is mapped into two different microarchitectures: a low-latency and low area iterative unit, and a low-latency and high-throughput pipelined unit. In both units speculation between consecutive radix-8 iterations is used to reduce the timing.
- 25 May 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computers
- Vol. 72 (10) , 2990-3001
Radix-64 Floating-Point Division and Square Root: Iterative and Pipelined Units
Digit-recurrence algorithms are widely used in actual microprocessors to compute floating-point division and square root. These iterative algorithms present a good trade-off in terms of performance, area and power. Traditionally, commercial processors have iterative division and square root units where the iteration logic is used over several cycles. The main drawbacks of these iterative units are long latency and low throughput due to the reuse of part of the logic over several cycles, and its hardware complexity with separated logic for division and square root. We present a radix-64 floating-point division and square root algorithm with a common iteration for division and square root and where, to have an affordable implementation, each radix-64 iteration is made of two simpler radix-8 iterations. The radix-64 algorithm allows to get low-latency operations, and the common division and square root radix-64 iteration results in some area reduction. The algorithm is mapped into two different microarchitectures: a low-latency and low area iterative unit, and a low-latency and high-throughput pipelined unit. In both units speculation between consecutive radix-8 iterations is used to reduce the timing.
- 16 October 2019
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computers
- Vol. 69 (2) , 274-287
Low Latency Floating-Point Division and Square Root Unit
- 16 October 2019
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computers
- Vol. 69 (2) , 274-287
Low Latency Floating-Point Division and Square Root Unit
- 1 June 2018
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 84-91
Radix-64 Floating-Point Divider
The following topics are dealt with: floating point arithmetic; digital arithmetic; IEEE standards; field programmable gate arrays; learning (artificial intelligence); cryptography; parallel processing; table lookup; multiplying circuits; mathematics computing.
- 1 June 2018
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 84-91
Radix-64 Floating-Point Divider
The following topics are dealt with: floating point arithmetic; digital arithmetic; IEEE standards; field programmable gate arrays; learning (artificial intelligence); cryptography; parallel processing; table lookup; multiplying circuits; mathematics computing.
- Open Access
- 26 January 2012
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 8 (4) , 1-20
FlexSig
With the advent of chip multiprocessors, new techniques have been developed to make parallel programing easier and more reliable. New parallel programing paradigms and new methods of making the execution of programs more efficient and more reliable have been developed. Usually, these improvements require hardware support to avoid a system slowdown. Signatures based on Bloom filters are widely used as hardware support for parallel programing in chip multiprocessors. Signatures are used in Transactional Memory, thread-level speculation, parallel debugging, deterministic replay and other tools and applications. The main limitation of hardware signatures is the lack of flexibility: if signatures are designed with a given configuration, tailored to the requirements of a specific tool or application, it is likely that they do not fit well for other different requirements. In this paper a new hardware signature organization, called Flexible Signatures ( FlexSig ), is proposed. FlexSig can change dynamically the resources assigned to a given signature and the number of signatures in the system, by redistributing the available hardware resources according to the system requirements. This allows higher flexibility than with traditional fixed-resources signatures based on Bloom filters, while maintaining a low false positive rate. FlexSig has been evaluated by comparing it with signatures based on parallel Bloom filters, and we conclude that FlexSig outperforms (in terms of false positive rate) conventional parallel Bloom filters in most cases, due to its ability to use all the signature resources available.
- 26 January 2012
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 8 (4) , 1-20
Open AccessFlexSig
With the advent of chip multiprocessors, new techniques have been developed to make parallel programing easier and more reliable. New parallel programing paradigms and new methods of making the execution of programs more efficient and more reliable have been developed. Usually, these improvements require hardware support to avoid a system slowdown. Signatures based on Bloom filters are widely used as hardware support for parallel programing in chip multiprocessors. Signatures are used in Transactional Memory, thread-level speculation, parallel debugging, deterministic replay and other tools and applications. The main limitation of hardware signatures is the lack of flexibility: if signatures are designed with a given configuration, tailored to the requirements of a specific tool or application, it is likely that they do not fit well for other different requirements. In this paper a new hardware signature organization, called Flexible Signatures ( FlexSig ), is proposed. FlexSig can change dynamically the resources assigned to a given signature and the number of signatures in the system, by redistributing the available hardware resources according to the system requirements. This allows higher flexibility than with traditional fixed-resources signatures based on Bloom filters, while maintaining a low false positive rate. FlexSig has been evaluated by comparing it with signatures based on parallel Bloom filters, and we conclude that FlexSig outperforms (in terms of false positive rate) conventional parallel Bloom filters in most cases, due to its ability to use all the signature resources available.
- 19 November 2002
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 172-179
Redundant Cordic rotator based on parallel prediction
- 19 November 2002
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 172-179
Redundant Cordic rotator based on parallel prediction
- 13 November 2002
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 85-90
The STEM-II air quality model on a distributed memory system
- 13 November 2002
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 85-90
The STEM-II air quality model on a distributed memory system
- 1 July 1993
- Published by Elsevier in Parallel Computing
- Vol. 19 (7) , 729-744
Design of a pipelined radix 4 CORDIC processor
- 1 July 1993
- Published by Elsevier in Parallel Computing
- Vol. 19 (7) , 729-744
Design of a pipelined radix 4 CORDIC processor
- 1 May 1990
- Published by Elsevier in Parallel Computing
- Vol. 14 (1) , 51-60
Gaussian elimination with pivoting on hypercubes
- 1 May 1990
- Published by Elsevier in Parallel Computing
- Vol. 14 (1) , 51-60
Gaussian elimination with pivoting on hypercubes