`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
` UNITED STATES PATENT AND TRADEMARK OFFICE
`
`
`
`
`
`
`BEFORE THE PATENT TRIAL AND APPEAL BOARD
`
`
`
`
`
`
`APPLE INC.,
`Petitioner,
`
`v.
`
`Zentian Limited
`Patent Owner.
`____________________
`
`Case IPR2023-00037
`Patent No. 10,971,140
`____________________
`
`
`
`PATENT OWNER’S RESPONSE
`[CORRECTED]
`
`
`
`
`
`
`
`
`
`
`
`IPR2023-00037
`PATENT OWNER’S RESPONSE
`
`TABLE OF CONTENTS
`
`
`
`
`I.
`
`II.
`
`Introduction
`
`Background of the ’140 Patent
`
`III. Person of ordinary skill in the art
`
`IV. The POSA would not have had a reasonable expectation of success with
`respect to the Petition’s combination of Jiang and Chen
`
`1
`
`3
`
`6
`
`7
`
`The Petition and Mr. Schmandt fail to demonstrate a motivation to combine
`V.
`Jiang with Chen
`16
`
`VI. The Petition fails to show obviousness as to limitations 1(d) and 1(e)
`
`VII. Conclusion
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`- 1 -
`
`27
`
`35
`
`
`
`
`
`
`
` IPR2023-00037
`PATENT OWNER’S RESPONSE
`
`
`EXHIBIT LIST
`
`
`
`Exhibit No.
`
`Description
`
`2017
`
`Deposition of Christopher Schmandt dated September 6, 2023
`
`2018
`
`Intentionally left blank
`
`2019
`
`Hennessy & Patterson, Computer Architecture, A Quantitative
`Approach, Third Edition (2003) (“Hennessy & Patterson”)
`
`2020
`
`Declaration of David Anderson, Ph.D
`
`2021
`
`Binu K. Mathew et al., A Gaussian Probability Accelerator for
`SPHINX 3, (“Mathew I”)
`
`
`
`
`
`
`
`
`
`
`- 2 -
`
`
`
`
`
`
`
`IPR2023-00037
`PATENT OWNER’S RESPONSE
`
`I.
`
`Introduction
`
`The Petition proposes that an ordinary speech recognition artisan would
`
`have been motivated to implement Jiang’s speech recognition teachings in Chen’s
`
`clustered processor and memory architecture with a reasonable expectation of
`
`success.
`
`But neither Jiang nor Chen enables such a combination, so the Petition must
`
`prove enablement through evidence outside those references. The Petition and Mr.
`
`Schmandt, however, fail to make that showing. Indeed, Mr. Schmandt admitted
`
`under cross-examination that he has never built or designed the processor to
`
`memory architecture for any of the speech recognition systems he identifies in the
`
`background of his declaration. Mr. Schmandt further admitted he has never
`
`supervised anyone in the process of mapping speech recognition software to a
`
`hardware architecture like Chen’s.
`
`As Dr. Anderson explains in detail, the Petition’s combination would have
`
`raised a number of complications and conflicts that the POSA would not have been
`
`qualified to address or resolve. Indeed, Jiang and Chen are inherently ill-suited for
`
`combination, since Jiang’s technique requires extensive communication between
`
`computational components whereas Chen does not allow communication between
`
`some of its processors and memories.
`
`
`
`
`
`
`
`
`
`
`
`- 1 -
`
`
`
`
`
`
`
` IPR2023-00037
`PATENT OWNER’S RESPONSE
`
`Moreover, real-world evidence from contemporaneous references teaches
`
`that the combination would not be expected to speed up Jiang’s recognition
`
`process, and that the POSA would not have been motivated by such an expectation.
`
`To the contrary, Mathew I teaches that a five-processor parallel processing system
`
`slowed down the speech recognition process compared to a two-processor system
`
`due to the associated synchronization overhead. Mr. Schmandt’s other purported
`
`motivations—alleged cost benefits and “flexibility and scalability”—are likewise
`
`baseless, as Mr. Schmandt’s deposition revealed. By contrast, Dr. Anderson
`
`explains in detail why the POSA would not have been motivated to combine Jiang
`
`with Chen in view of the known challenges and disadvantages of such a
`
`combination balanced against the low likelihood of benefits.
`
`Finally, the Petition further fails because it provides no theory as to how the
`
`combination of Jiang and Chen would operate to meet limitations 1(d) and 1(e). As
`
`Mr. Schmandt admitted, his theory requires that Jiang modified to use Chen’s
`
`hardware would practice both limitations, but his declaration does not even
`
`mention Chen with respect to limitations 1(d) and 1(e), much less explain how the
`
`combination of Jiang and Chen would meet them. And although Mr. Schmandt
`
`attempted to fill the void with a new, ad hoc theory presented for the first time at
`
`his deposition, the evidence demonstrates that theory would not have been enabled.
`
`
`
`
`
`
`
`
`
`
`
`- 2 -
`
`
`
`
`
`
`
` IPR2023-00037
`PATENT OWNER’S RESPONSE
`
`Mr. Schmandt’s failure to present a workable explanation as to how the POSA
`
`could have arrived at the invention of the challenged claims by combining Jiang
`
`and Chen is definitive evidence that the claims are far from obvious.
`
`Accordingly, the Board should not find any challenged claim unpatentable.
`
`II. Background of the ’140 Patent
`
`U.S. Patent 10,971,140, titled “Speech recognition using parallel
`
`processors,” is directed to an improved speech recognition circuit that “uses
`
`parallel processors for processing the input speech data in parallel.” Ex. 1001,
`
`1:18-20. The ’140 patent teaches multiple processors “arranged in groups or
`
`clusters,” with each group or cluster of processors connected to one of several
`
`“partial lexical memories” that “contains part of the lexical data.” Ex. 1001, 3:13-
`
`18. “Each lexical tree processor is operative to process the speech parameters using
`
`a partial lexical memory and the controller controls each lexical tree processor to
`
`process a lexical tree corresponding to partial lexical data in a corresponding
`
`partial lexical memory.” Ex. 1001, 3:19-24. The ’140 patent further teaches that
`
`the invention “provides a circuit in which speech recognition processing is
`
`performed in parallel by groups of processors operating in parallel in which each
`
`group accesses a common memory of lexical data.” Ex. 1001, 3:62-66.
`
`
`
`
`
`
`
`
`
`
`
`- 3 -
`
`
`
`
`
`
`
` IPR2023-00037
`PATENT OWNER’S RESPONSE
`
`The specification of the ’140 patent thus provides: “[T]he present invention
`
`provides a circuit in which speech recognition processing is performed in parallel
`
`by groups of processors operating in parallel in which each group accesses a
`
`common memory of lexical data. . . . Each processor within a group can access the
`
`same lexical data as any other processor in the group. The controller can thus
`
`control the parallel processing of input speech parameters in a more flexible
`
`manner. For example, it allows more than one processor to process input speech
`
`parameters using the same lexical data in a lexical memory. This is because the
`
`lexical data is segmented into domains which are accessible by multiple
`
`processors.” Ex. 1001, 3:62-4:18.
`
`Figure 2 of the patent, annotated below, further illustrates that architecture
`
`by showing two groups of lexical tree processors, with each group containing
`
`multiple processors 1-k, and each group of processors connected to a dedicated
`
`“acoustic model memory,” such that there are least two acoustic model memories
`
`for at least two groups of processors. Ex. 1001, Fig. 2 (annotated).
`
`
`
`
`
`
`
`
`
`
`
`- 4 -
`
`
`
`
`
`
`
` IPR2023-00037
`PATENT OWNER’S RESPONSE
`
`
`
`
`
`Moreover, the ’140 patent expressly distinguishes its novel architecture from
`
`two prior known alternative designs, which the patent describes as less
`
`advantageous. In particular, the ’140 patent teaches that: “By providing a plurality
`
`of processors in a group with a common memory, flexibility in the processing is
`
`provided without being bandwidth limited by the interface to the memory that
`
`would occur if only a single memory were used for all processors. The
`
`arrangement is more flexible than the parallel processing arrangement in which
`
`each processor only has access to its own local memory and requires fewer
`
`memory interfaces (i.e. chip pins).” Ex. 1001, 4:1-9. The patent thus distinguishes
`
`
`
`
`
`
`
`
`
`
`
`- 5 -
`
`
`
`
`
`
`
` IPR2023-00037
`PATENT OWNER’S RESPONSE
`
`its design from (1) a one-memory-to-all-processors design, which it describes as
`
`bandwidth limited in the processor to memory interface; and (2) a one-memory-to-
`
`one-processor design, which would require more memory interfaces and would be
`
`less flexible.
`
`III. Person of ordinary skill in the art
`
`This Response addresses the Petition’s deficiencies by assuming the Petition’s
`
`proffered definition of the person of ordinary skill in the art.
`
`As Dr. Anderson explains, however, the field of electrical engineering
`
`encompasses many specialties, including fiber optics, analog circuits, digital
`
`circuits, very large-scale integration (VLSI), systems and controls, digital signal
`
`processing, bio-medical sensors and systems, electrical energy, electromagnetics,
`
`nanotechnology, telecommunications, computer systems and software, and others.
`
`Ex. 2020 ¶ 17.
`
`Thus, if a student desires to study speech recognition, they would specialize
`
`in the area of digital signal processing (DSP) and then pursue studies specific to the
`
`advanced subspecialty of DSP in their graduate studies. On the other hand, a student
`
`who desires to learn computer architecture would take a different set of core courses
`
`in one of the computer engineering specialties. Ex. 2020 ¶ 17.
`
`
`
`
`
`
`
`
`
`
`
`- 6 -
`
`
`
`
`
`
`
` IPR2023-00037
`PATENT OWNER’S RESPONSE
`
`Mr. Schmandt’s proposed person of ordinary skill in the art is a specialist in
`
`speech recognition. Any such person of ordinary skill would not also be expected to
`
`have specialized in parallel processing architectures and methods or high
`
`performance computing in addition to speech recognition. Ex. 2020 ¶ 18. Indeed,
`
`Dr. Anderson has never taught a course that would have equipped a masters-level
`
`student in speech recognition to apply a parallel processing architecture like Chen’s
`
`to known speech recognition techniques. Ex. 2020 ¶ 18. While a person with a
`
`master’s degree in one of the fields identified in Mr. Schmandt’s POSA definition
`
`could be a person of ordinary skill in speech recognition or a person of ordinary skill
`
`in high performance computing and parallel processing, that person would not be
`
`both. Ex. 2020 ¶ 19.
`
`IV. The POSA would not have had a reasonable expectation of success with
`respect to the Petition’s combination of Jiang and Chen
`
`The Petition’s theory as to all challenged claims relies on at least a
`
`combination of Jiang and Chen. In particular, the Petition and Mr. Schmandt
`
`propose that the ordinary artisan would have found it obvious “to configure a
`
`computing platform comprising clusters of processors as taught by Chen to
`
`perform the speech recognition techniques of Jiang.” Ex. 1003 ¶ 68.
`
`
`
`
`
`
`
`
`
`
`
`- 7 -
`
`
`
`
`
`
`
` IPR2023-00037
`PATENT OWNER’S RESPONSE
`
`“To render a claim obvious, the prior art, taken as a whole, must enable a
`
`skilled artisan to make and use the claimed invention.” Raytheon Techs. Corp. v.
`
`GE, 993 F.3d 1374, 1380 (Fed. Cir. 2021) (emphasis added). “[E]ven though a
`
`non-enabling reference can play a role in an obviousness analysis, the evidence of
`
`record must still establish that a skilled artisan could have made the claimed
`
`invention.” Id. at 1381.
`
`Here, Jiang does not enable performing its speech recognition techniques
`
`using the particular clustered processor embodiment the Petition has selected from
`
`Chen, and Chen does not enable using that clustered processor embodiment to
`
`perform Jiang’s speech recognition techniques. Ex. 2020 ¶ 24. Mr. Schmandt
`
`admitted that reality at his deposition.
`
`Ex. 2017 at 86:9-17.
`
`
`
`
`
`
`
`
`
`
`
`- 8 -
`
`
`
`
`
`
`
`
`
` IPR2023-00037
`PATENT OWNER’S RESPONSE
`
`
`
`Accordingly, the evidence of record beyond Jiang and Chen must establish
`
`that the ordinary artisan could have “configure[d] a computing platform
`
`comprising clusters of processors as taught by Chen to perform the speech
`
`recognition techniques of Jiang,” as the Petition’s theory requires. Ex. 1003 ¶ 68.
`
`Mr. Schmandt and the Petition, however, do not provide that evidence.
`
`Instead, Mr. Schmandt’s declaration simply states the sweeping and conclusory
`
`generalization that the “modifications necessary, such as configuring Chen’s
`
`circuitry to specifically recognize speech pursuant to Jiang would have required
`
`only software programming and well-known computing techniques and
`
`structures, and thus would have led a POSITA to a reasonable expectation of
`
`success.” Ex. 1003 ¶ 68 (emphasis added).
`
`As Mr. Schmandt admitted under cross-examination, his declaration is thus
`
`silent as to any underlying details to support his conclusion as to the POSA’s
`
`capability; instead, his declaration merely states the conclusion without any
`
`underlying explanation.
`
`
`
`
`
`
`
`
`
`
`
`- 9 -
`
`
`
`
`
`
`
` IPR2023-00037
`PATENT OWNER’S RESPONSE
`
`
`
`Ex. 2017 at 48:13-21.
`
`
`
`An expert’s “ipse dixit declaration,” however, cannot prove obviousness. TQ
`
`Delta, LLC v. Cisco Sys., 942 F.3d 1352, 1362 (Fed. Cir. 2019). Mr. Schmandt’s
`
`“conclusory statements and unspecific expert testimony are [ ] thus inadequate to
`
`support the Board’s finding” of obviousness. Id. at 1363.
`
`In addition to being devoid of evidentiary value, Mr. Schmandt’s conclusory
`
`assurances as to the POSA’s capabilities are also not credible in view of his
`
`deposition testimony.
`
`Although Mr. Schmandt’s declaration claims he has “built” a number of
`
`speech recognition systems, see Ex. 1003 ¶¶ 4, 6, Mr. Schmandt admitted under
`
`cross-examination that he has never built the processor to memory architecture of
`
`any speech recognition system identified in his declaration.
`
`
`
`
`
`
`
`
`
`
`
`- 10 -
`
`
`
`
`
`
`
` IPR2023-00037
`PATENT OWNER’S RESPONSE
`
`
`
`
`
`
`
`
`
`
`Ex. 2017 at 34:24-35:5.
`
`
`Ex. 2017 at 32:3-9.
`
`Moreover, Mr. Schmandt has never supervised anyone involved in the
`
`process of mapping a speech recognition model to a clustered processor and
`
`memory architecture like Chen’s.
`
`
`
`
`
`
`
`
`
`
`
`- 11 -
`
`
`
`
`
` IPR2023-00037
`PATENT OWNER’S RESPONSE
`
`
`
`Ex. 2017 at 146:20-24.
`
`
`
`Mr. Schmandt thus lacks an adequate basis from his personal knowledge and
`
`experience to opine that the combination of Jiang and Chen was within the skill
`
`level of ordinary artisan in the field of speech recognition.
`
`By contrast, as Dr. David Anderson explains, implementing Jiang’s speech
`
`recognition techniques on Chen’s clustered processor and memory architecture
`
`would have been highly complex, and far outside the skill level of the ordinary
`
`artisan in the field of speech recognition as defined by the Petition, without the
`
`benefit of the ’140 Patent. Ex. 2020 ¶ 27. In particular, moving Jiang’s speech
`
`recognition techniques to Chen’s clustered, parallel processors and memories
`
`would have required coordinating multiple caches, avoiding memory conflicts,
`
`controlling task sharing in an efficient manner, resolving synchronous bottlenecks,
`
`and addressing communication bandwidth and latency issues among the various
`
`hardware components, and developing a messaging strategy to coordinate
`
`
`
`
`
`
`
`
`
`
`
`- 12 -
`
`
`
`
`
`information sharing between and within clusters, among other challenges. Ex. 2020
`
`
`
` IPR2023-00037
`PATENT OWNER’S RESPONSE
`
`¶ 27.
`
`Such tasks were not within the skill level of the ordinary artisan in the field
`
`of speech recognition as defined by the Petition prior to the ’140 Patent. Ex. 2020 ¶
`
`27. As Dr. Anderson explains, the field of electrical engineering encompasses
`
`many specialties, including fiber optics, analog circuits, digital circuits, very large-
`
`scale integration (VLSI), systems and controls, digital signal processing, bio-
`
`medical sensors and systems, electrical energy, electromagnetics, nanotechnology,
`
`telecommunications, computer systems and software, and others. Ex. 2020 ¶¶ 17-
`
`19, 28 (pp. 7-9, 17). Students of electrical or computer engineering specialize in
`
`one of these areas after taking some general courses. Id. The ordinary artisan in the
`
`field of speech recognition would have specialized in the area of digital signal
`
`processing (DSP), and would have further pursued studies specific to speech
`
`recognition, an advanced subspecialty of DSP, in their graduate studies. Id. The
`
`POSA in the field of speech recognition would not have been capable of
`
`addressing the complex challenges of combining Chen with Jiang in the manner
`
`the Petition has proposed any more than a dermatologist could be expected to also
`
`perform open heart surgery. Ex. 2020 ¶ 28 (p. 17).
`
`
`
`
`
`
`
`
`
`
`
`- 13 -
`
`
`
`
`
`
`
` IPR2023-00037
`PATENT OWNER’S RESPONSE
`
`Moreover, speech recognition requires evaluating the comparative
`
`likelihoods of many possible outcomes for each incoming frame of samples. Ex.
`
`2020 ¶ 29 (p. 17). Jiang teaches the use of the Viterbi decoding algorithm. Ex.
`
`1004 at 1:19-39, 2:1-46. The Viterbi algorithm compares the likelihoods of all
`
`possible sequences for a given input frame, given the most likely outcomes from
`
`the previous frame. Id.; Ex. 2020 ¶ 29 (p. 18). To perform these comparisons
`
`requires extensive communication between computational components that can
`
`have a drastic impact on the system design and performance. Ex. 2020 ¶ 29 (p. 18).
`
`By contrast, Chen teaches that the memory associated with one cluster is not
`
`directly or adjacently accessible by the processors and memories in each other
`
`cluster. See Ex. 1003 ¶ 95; Ex. 1005 at 9:10-39; Ex. 2020 ¶ 29 (p. 18). As a result,
`
`performing Viterbi decoding when the information needed by each node may be in
`
`a cluster that is not directly accessible, as in Chen, is not likely to be successful in a
`
`practical speech recognition system. Ex. 2020 ¶ 29 (p. 18).
`
`As explained by Hennesy & Patterson and Dr. Anderson, communication
`
`between parallel processing nodes is one of several considerations that can
`
`determine the feasibility of a particular parallel architecture for use with a
`
`particular problem. Ex. 2019 at 534-535; Ex. 2020 ¶ 30 (p. 18). Hennesy &
`
`Patterson also points out that the private memory model in which memory is
`
`
`
`
`
`
`
`
`
`
`
`- 14 -
`
`
`
`
`
`
`
` IPR2023-00037
`PATENT OWNER’S RESPONSE
`
`connected only within a cluster may be suitable “for applications that require little
`
`or no communication.” Ex. 2019 at 533; Ex. 2020 ¶ 30 (p. 18). Chen implements
`
`such a private memory architecture but the speech recognition search stage is an
`
`application that requires significant communication. Ex. 2020 ¶ 30 (p. 18).
`
`Therefore, based on teachings at the time of the ’140 Patent, the POSA would not
`
`expect the proposed combination of Chen with Jiang to be successful. Ex. 2020 ¶
`
`30 (pp. 18-19).
`
`Dr. Anderson’s own experience in supervising a team that attempted to
`
`implement an application within a parallel processing environment is illustrative.
`
`Ex. 2020 ¶ 31. In addition to support from Dr. Anderson, this team of engineers
`
`worked in close collaboration with a DSP application expert at Texas Instruments
`
`for six months in order to implement a custom computer vision application on a
`
`multi-processor system supplied by Texas Instruments (Texas Instrument’s multi-
`
`processor daVinci platform). Id. Even with training, expert oversight, and
`
`extensive efforts on behalf of the team, the inter-processor communication details
`
`prevented the successful implementation of their system. Id. Rather, the final
`
`product was only able to do a simple tracking of a black ball on a white
`
`background—something that could have more easily been performed on a single
`
`processor. Id. While this system was not a speech recognition system, it was a
`
`
`
`
`
`
`
`
`
`
`
`- 15 -
`
`
`
`
`
`
`
` IPR2023-00037
`PATENT OWNER’S RESPONSE
`
`signal processing system that was simpler than a speech recognition system in
`
`terms of complexity (only a few hundreds of lines of code). Id. The failure of this
`
`design example merely demonstrates that transitioning a signal processing system
`
`to any particular multi-processing architecture is not necessarily trivial or practical,
`
`and that even those beyond the level of ordinary skill in the field of speech
`
`recognition could not necessarily expect success with respect to far simpler
`
`combinations than the one the Petition has proposed. Id.; Kloster Speedsteel AB v.
`
`Crucible, Inc., 793 F.2d 1565, 1574 (Fed. Cir. 1986) (“If the invention here would
`
`not have been obvious to one of extraordinary skill, it follows that in this case it
`
`would not be obvious to one with lesser skills.”).
`
`Thus, even accepting arguendo that Jiang and Chen both contemplate the
`
`use of a plurality of processors, a POSA would not have “been capable of simply
`
`substituting” Chen’s clustered processors to implement Jiang’s speech recognition
`
`techniques, as Mr. Schmandt’s declaration simplistically presumes. Compare Ex.
`
`1003 ¶¶ 67, 74 with Ex. 2020 ¶¶ 27-32 (pp. 16-20).
`
`The evidence of record thus does not demonstrate that an ordinary artisan in
`
`the field of speech recognition could have configured Chen’s clustered processor
`
`and memory architecture to perform Jiang’s speech recognition techniques with a
`
`reasonable expectation of success.
`
`
`
`
`
`
`
`
`
`
`
`- 16 -
`
`
`
`
`
`
`
` IPR2023-00037
`PATENT OWNER’S RESPONSE
`
`V. The Petition and Mr. Schmandt fail to demonstrate a motivation to
`combine Jiang with Chen
`
`Apple and Mr. Schmandt allege that the POSA would have been motivated
`
`to combine Jiang with Chen due to the alleged benefits of improved speed and
`
`power, a relaxed pruning threshold, and reduced cost. Ex. 1003 ¶¶ 67, 74.
`
`Separately, Mr. Schmandt contends that “the flexibility and scalability afforded to
`
`the MARS machine by its clustered architecture [ ] would have motivated a
`
`POSITA to implement clustered architecture in other speech recognition circuits.”
`
`Ex. 1003 ¶ 71. The evidence, however, does not support any such benefits, or the
`
`existence of a motivation to combine.
`
`Mr. Schmandt’s declaration assumes that the clustered processor and
`
`memory arrangement of Chen would have necessarily improved the speed and
`
`power of Jiang’s speech recognition, and thus allowed for relaxed pruning
`
`threshold. Ex. 1003 ¶¶ 67, 74. Notably, Mr. Schmandt’s declaration does not
`
`undertake any quantitative or other particularized analysis to demonstrate why
`
`Chen would allegedly speed up Jiang and relax its pruning threshold, or by how
`
`much. Ex. 2020 ¶ 35. Rather, Mr. Schmandt seems to simply assume the
`
`conclusion that more processors would have been better. Id.
`
`
`
`
`
`
`
`
`
`
`
`- 17 -
`
`
`
`
`
`
`
` IPR2023-00037
`PATENT OWNER’S RESPONSE
`
`Contemporaneous real-world evidence, however, refutes Mr. Schmandt’s
`
`assumptions. Ex. 2020 ¶ 36. Mathew I provides a detailed report on computer
`
`architecture modifications intended to improve the speed of a speech recognizer,
`
`and in particular through the addition more processors. Ex. 2021 at 11. Mathew I
`
`explains that it developed “a parallel version” of the Sphinx speech recognition
`
`system in which the distance calculation (GAU, or Gaussian probability
`
`estimation) was run on a separate processor from the search stage (HMM
`
`processing) with parallelization. Id. This two-processor parallel processing design
`
`was only 1.67x faster than the “original sequential version,” not 1.97x faster as the
`
`authors had predicted based on Amdahl’s law. Ex. 2021 at 11; Ex. 2020 ¶ 36. The
`
`authors then further modified the search stage (HMM phase) “to use 4 processors
`
`instead of 1,” thus moving from a two-processor system (one for distance
`
`calculation and one for search) to a five-processor system (one processor for the
`
`calculation stage, and four for search). Ex. 2021 at 11; Ex. 2020 ¶ 36. As Mathew I
`
`reports, “the resulting 5 processor version was slower than the 2 processor version
`
`due to the high synchronization overhead.” Ex. 2021 at 11 (emphasis added); Ex.
`
`2020 ¶ 36.
`
`As Dr. Anderson explains, Mathew I demonstrates that the ordinary artisan
`
`could not have assumed that Chen’s complex, clustered parallel processing
`
`
`
`
`
`
`
`
`
`
`
`- 18 -
`
`
`
`
`
`
`
` IPR2023-00037
`PATENT OWNER’S RESPONSE
`
`architecture with numerous processors in each cluster and multiple distributed
`
`memories would have improved the speed of Jiang’s speech recognition system or
`
`relaxed its pruning threshold in view of the high synchronization overhead and
`
`other complications that such a system would have implicated. Ex. 2020 ¶ 37.
`
`Mr. Schmandt’s declaration testimony as to the alleged cost-based
`
`motivations of the combination are likewise entirely baseless. As Mr. Schmandt
`
`admitted at his deposition, his declaration opines that the combination of Jiang and
`
`Chen would have been less expensive, but never states what the combination
`
`would have been less expensive than, or on what basis.
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`- 19 -
`
`
`
`
`
`
`
` IPR2023-00037
`PATENT OWNER’S RESPONSE
`
`Ex. 2017 at 61:14-62:9. Mr. Schmandt also made clear that his declaration
`
`testimony was not based on a cost comparison between Chen’s clustered
`
`processing architecture and any other allegedly alternative clustered processing
`
`
`
`architectures.
`
`Ex. 2017 at 63:8-15. And Mr. Schmandt likewise admitted that his declaration
`
`provides no quantitative cost comparison between Jiang’s unmodified architecture
`
`
`
`
`
`
`
`
`
`
`
`
`
`- 20 -
`
`
`
`
`
`and his proposed modified architecture in combination with Chen. Ex. 2017 at
`
`
`
` IPR2023-00037
`PATENT OWNER’S RESPONSE
`
`40:23-41:4.
`
`
`
`In sum, there is no basis for Mr. Schmandt’s declaration testimony that the
`
`POSA would have been motivated to combine Jiang with Chen based on a
`
`purported cost or expense benefit from the combination. Ex. 2020 ¶ 39.
`
`
`
`Moreover, Mr. Schmandt testified that a motivation to modify would require
`
`the ordinary artisan to first assess the cost implications of a potential modification
`
`against the imputed or actual value of the modification, considering both any
`
`specific performance improvements and the value of those improvements to end
`
`users. Ex. 2017 at 41:18-42:13. According to Mr. Schmandt: “Cost would be part
`
`of a design tradeoff and also looking at your customer base.” Ex. 2017 at 42:12-
`
`42:13. Yet Mr. Schmandt readily admitted he did not undertake any qualitative or
`
`quantitative cost-benefit analysis in reaching his opinions regarding a motivation to
`
`combine.
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`- 21 -
`
`
`
`
`
` IPR2023-00037
`PATENT OWNER’S RESPONSE
`
`Ex. 2017 at 42:14-19.
`
`
`
`Mr. Schmandt’s declaration testimony as to a cost-based motivation thus
`
`does not meet his own criteria for what a POSA would need to consider before
`
`forming a motivation.
`
`Nor does it meet the Federal Circuit’s requirement as to proving a
`
`motivation to combine. Winner Int’l Royalty Corp. v. Wang, 202 F.3d 1340, 1349
`
`(Fed. Cir. 2000) (holding that “[m]otivation to combine requires” evidence the
`
`alleged benefits of the combination, in view of the associated tradeoffs and costs,
`
`would have been “on balance, desirable.”).
`
`
`
`Equally without merit is Mr. Schmandt’s “flexibility and scalability”
`
`motivation. Ex. 1003 ¶¶ 71, 72. Mr. Schmandt’s declaration cites Hon as
`
`purportedly teaching that clustered processing architectures in general present the
`
`benefits of flexibility and scalability, and that this would have motivated the
`
`combination of Jiang and Chen. Id. But Mr. Schmandt’s deposition testimony
`
`clarifies that Hon’s cited passages instead teach that MARS was flexible because
`
`its pipeline was programmable, and that it was scalable because more processors
`
`could be added to it. Ex. 2017 at 105:8-106:4; Ex. 2020 ¶ 40. In both instances,
`
`Hon was comparing the MARS system to another system, SRI-Berkeley, which did
`
`not have a programmable pipeline or extendable processor architecture. Ex. 2017
`
`
`
`
`
`
`
`
`
`
`
`- 22 -
`
`
`
`
`
`
`
` IPR2023-00037
`PATENT OWNER’S RESPONSE
`
`at 105:8-106:4; Ex. 2020 ¶ 40; Ex. 1006 at 24. Hon’s discussion of MARS is thus
`
`specific to two particular advantages of MARS over SRI-Berkeley, and Hon thus
`
`does not provide a generalized “flexibility and scalability” motivation for
`
`combining Jiang with a clustered processing architecture, much less Chen’s in
`
`particular. Ex. 2020 ¶ 40. Indeed, Mr. Schmandt has made no showing that Jiang
`
`itself, without modification, was not already both flexible and scalable in the same
`
`manner that Hon describes with respect to MARS. Ex. 2020 ¶ 41.
`
`
`
`Ultimately, Mr. Schmandt testified that it is never desirable to add another
`
`processor to a speech recognition system when that system can already perform
`
`adequately.
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`- 23 -
`
`
`
`
`
` IPR2023-00037
`PATENT OWNER’S RESPONSE
`
`
`
`Ex. 2017 at 163:4-12 (emphasis added). Yet Mr. Schmandt and Apple have made
`
`no showing that Jiang’s system could not perform adequately without being
`
`combined with Chen. Accordingly, there is no motivation to combine here under
`
`Mr. Schmandt’s own criteria.
`
`
`
`As Dr. Anderson further explains, the POSA would not have been motivated
`
`to implement Jiang’s speech recognition techniques in Chen’s clustered processing
`
`architecture. Ex. 2020 ¶¶ 42-45. Jiang’s existing architecture relies on a shared
`
`memory structure in which all components have access to the same memory. Ex.
`
`1004, Fig. 1; Ex. 2020 ¶ 42. The relevant embodiment the Petition selects from
`
`Chen, however, relies on a distributed memory consisting of multiple private
`
`memories, with one memory per cluster, in which access to those private memories
`
`is extended also to directly adjacent clusters. Ex. 2020 ¶ 42. As Chen itself states,
`
`“[t]o effectively port an existing software program from a shared memory model to
`
`a private memory model can require significant reprogramming effort.” Ex. 1005
`
`
`
`
`
`
`
`
`
`
`
`- 24 -
`
`
`
`
`
`
`
` IPR2023-00037
`PATENT OWNER’S RESPONSE
`
`at 2:6-9 (emphasis added); Ex. 2020 ¶ 42. That is because “a distributed memory
`
`model necessarily requires that the processes and data which comprise a parallel
`
`programming problem be partitioned into separate tasks that are each small enough
`
`to fit into the private local memory of a processor and at the same time are large
`
`enough to recover the overhead required to establish a parallel task in a message
`
`passing environment.” Ex. 1005 at 2:9-17 (emphasis added); Ex. 2020 ¶ 42. The
`
`POSA would have been aware of Chen’s warning that porting Jiang’s speech
`
`recognition techniques to Chen’s distributed memory architecture would require
`
`“significant reprogramming effort,” including the complex task of finding the
`
`optimal approach to partitioning the processes and data of Jiang’s speech
`
`recognition techniques to balance the competing problems of memory size and
`
`overhead. Ex. 2020 ¶ 43. In addition, the overhead required to perform message
`
`passing (communications for information sharing) between private-memory
`
`clusters can be onerous. Ex. 2020 ¶ 43. One example of this problem was
`
`demonstrated in Mathew I and cited previously. Id. Since speech recognition
`
`search requires information to be shared among multiple recognition paths
`
`(candidate HMM state sequences) during the Viterbi decoding, the design of the
`
`communication and memory architecture would be critical. Id. Chen presents a
`
`particularly poor choice for the speech recognition search stage because it
`
`
`
`
`
`
`
`
`
`
`
`- 25 -
`
`
`
`
`
`
`
` IPR2023-00037
`PATENT OWNER’S RESPONSE
`
`explicitly does not have direct communication between all the clusters. Ex. 1003 ¶
`
`95; Ex. 1005, 9:10-39; Ex. 2020 ¶ 43. The POSA would not have been motivated
`
`to combine Jiang with Chen in the manner proposed in view of those obstacles. Ex.
`
`2020 ¶ 44. Indeed, the POSA would not have expected to successfully manage the
`
`overhead implications of such a combination in order to reasonably expect
`
`benefits. Ex. 2020 ¶ 44. As already explained, Mathew I teaches that even a five-
`
`processor parallel processing system ended up performing more slowly than a two-
`
`processor system “due to the high synchronization overhead.” Ex. 2021 at 11; Ex.
`
`2020 ¶ 44. The Petition’s combination, by contrast, requires at least eight
`
`processors and four distributed memories running in parallel across four clusters.
`
`Ex. 1003 ¶ 73.
`
`
`
`
`
`
`
`
`
`
`
`- 26 -
`
`
`
`
`
`
`
` IPR2023-00037
`PATENT OWNER’S RESPONSE
`
`The ordinary artisan would have known that mapping Jiang’s speech recognition
`
`techniques to such a complex architecture would have not, on balance, have been
`
`desirable in view of the high downsides and the low likelihood of benefits. Ex.
`
`
`
`2020 ¶ 45.
`
`
`
`
`
`
`
`
`
`
`
`- 27 -
`
`
`
`
`
`Accordingly, the POSA would not have been motivated to combine Jiang
`
`
`
` IPR2023-000