Abstract: Examples herein propose operating redundant ML models which have been trained using a boosting technique that considers hardware faults. The embodiments herein describe performing an evaluation process where the performance of a first ML model is measured in the presence of a hardware fault. The errors introduced by the hardware fault can then be used to train a second ML model. In one embodiment, a second evaluation process is performed where the combined performance of both the first and second trained ML models is measured in the presence of a hardware fault. The resulting errors can then be used when training a third ML model. In this manner, the three trained ML models are trained to be error aware. As a result, during operation, if a hardware fault occurs, the three ML models have better performance relative to three ML models that where not trained to be error aware.

Abstract: An apparatus includes a data processing array having a plurality of array tiles. The plurality of array tiles include a plurality of compute tiles. The compute tiles include a core coupled to a random-access memory (RAM) in a same compute tile and to a RAM of at least one other compute tile. The data processing array is subdivided into a plurality of partitions. Each partition includes a plurality of array tiles including at least one of the plurality of compute tiles. The apparatus includes a plurality of clock gate circuits being programmable to selectively gate a clock signal provided to a respective one of the plurality of partitions.

Abstract: Methods and apparatus for time-to-digital conversion. An example apparatus includes a first input; a second input; a delay line coupled to the first input and comprising a plurality of first delay elements coupled in series, each of the plurality of first delay elements having a first delay time; a second delay element having an input coupled to the second input and having the first delay time; a third delay element having an input coupled to the second input and having a second delay time, the second delay time being smaller than the first delay time; a first set of arbiters having first inputs coupled to the delay line and having second inputs coupled to an output of the second delay element; and a second set of arbiters having first inputs coupled to the delay line and having second inputs coupled to an output of the third delay element.

Abstract: Embodiments herein describe correcting nonlinearity in a Digital-to-Time Converter (DTC) by relaxing a DTC linearity requirement, which results in the correction being co-adapted with a DTC gain calibration loop which can operate in parallel with a DTC integral nonlinearity (INL) correction loop. In one embodiment, the DTC gain calibration loop and the DTC INL correction loop are constrained when determining a nonlinearity correction code to improve the likelihood they converge. Once determined, the nonlinearity correction code can be combined with an digital code output by a time-to-digital converter (TDC) to generate a phase difference between a reference clock and a feedback clock.

Abstract: A network interface device comprises a plurality of components configured to process a flow of data one after another. A control component is configured to provide one or more control messages in said flow, said one or more control message being provided to said plurality of components one after another such that a configuration of one or more of said components is changed.

Abstract: Control of a reconfigurable platform can include determining, by a host computer, an interface universally unique identifier (UUID) of an interface of platform circuitry implemented on an accelerator, wherein the accelerator is communicatively linked to the host computer. An electronic request to run a partition design on the accelerator is received by the host computer. In response to the electronic request, the host computer determines an interface UUID for an interface of the partition design and determines compatibility of the partition design with the platform circuitry based on a comparison of the interface UUID of the partition design with the interface UUID of the platform circuitry. The partition design is implemented on the accelerator in response to determining that the partition design is compatible with the platform circuitry.

Abstract: An adder for fractional logarithmic number system (FLNS) format operands includes a compare-and-swap circuit that inputs first and second FLNS operands represented by fixed point values and provides a greater one as operand x and a lesser or equal one as operand y. Sign bits are sx and sy of x and y, respectively, qx and qy, are integer portions of x and y, respectively, fraction portions of x and y have integer values rx and ry, respectively. The compare-and-swap circuit is configured to provide sx as a sign bit, sz of a sum z=x(1+y/x) for x?0. A subtraction circuit subtracts (qy+ry/n)?(qx+rx/n) and outputs q? and r?, such that ?=y/x, where n=2wr and wr is a bit-width of rx and ry. An approximation circuit provides an approximation of (1+?) to a nearest FLNS value, ?, as fixed point value having an integer portion q? and a fraction portion that has an integer value r?.

Abstract: Instruction generation for a data processing array and microcontroller includes generating a tensor-level intermediate representation from a machine learning model using kernel expressions. Statements of the tensor-level intermediate representation are partitioned into a first set of statements and a second set of statements. From the first set of statements, kernel instructions are generated based on a reconfigurable neural engine model. The kernel instructions are executable by a compute tile of a data processing array to implement compute functions of the machine learning model. From the set of second statements, microcontroller instructions are generated based on a super-graph model. The microcontroller instructions are executable by a microcontroller of the data processing array to move data into and out from the data processing array.

Abstract: Examples describe memory refresh operations for memory subsystems. One example is a method for a memory controller, the method including entering a first state upon exiting self-refresh state, wherein the first state comprises activating a first timer. The method includes entering a second state from the first state upon detecting an end of an active period and detecting that the first timer has not expired. The method includes entering a third state from the second state upon detecting expiration of the second state, wherein the third state comprises re-entering the self-refresh state.

Abstract: Techniques and apparatus for dynamically modifying a kernel (and associated user-specified circuitry) for a dynamic region of a programmable integrated circuit (IC) without affecting (e.g., while allowing) operation of other kernels ((and other associated user-specified circuitry) in the programmable IC. Dynamically modifying a kernel may include, for example, unloading an existing kernel, loading a new kernel, or replacing a first kernel with a second kernel). In the case of networking (e.g., in a data center application) where the programmable IC may be part of a hardware acceleration card (e.g., a network interface card (NIC)), the kernel may be user code referred to as a “plugin.

Abstract: Circuits and methods for determining a maximum bias for computing softmax on a tensor include a processor circuit configured to transform in parallel, elements of each group of a plurality of groups of elements of a tensor X into respective power-of-two elements. The respective power-of-two element from element xt of the tensor is pt, pt=(xt*log2e), and pt has an integer part and a fraction part. A first comparison circuit (204) is configured to determine respective group-level biases for the groups. The group-level bias of groupm is dm, and dm is an integer part of a maximum of the power-of-two elements of groupm. A second comparison circuit is configured to determine a greatest one of the respective group-level biases to be a tensor-level bias, dmax.

Abstract: A universal interposer for an integrated circuit (IC) device has a body having a first surface and a second surface opposite the first surface. A first region is formed on a first side of the body along a first edge. The first region has first slots, each having an identical first bond pad layout. A second region is formed on the first side along a second edge, opposite the first edge. The second region has second slots having an identical second bond pad layout. A third region having third slots is formed on the first side between the first and second regions, each slot having an identical third bond pad layout. A pad density of the third bond pad layout is greater than the first bond pad layout. One of the third slots is coupled to contact pads disposed in a region not directly below any of the second slots.

Abstract: An integrated circuit (IC) device is disclosed which includes at least a first hybrid bond interface layer disposed between adjacent wafers of a wafer stack. Routing within the hybrid bond interface layer allows test pads exposed on a top wafer of the wafer stack to electrically couple test keys within the wafer stack. By utilizing the routing within the hybrid bond interface layer to index electrical connections between adjacent wafers, IC dies stacked on the wafers may be fabricated with less mask sets as compared to conventional designs.

Abstract: Controlling a data processing (DP) array includes creating a replica of a register address space of the DP array based on the design and the DP array. A sequence of instructions, including write instructions and read instructions, is received. The write instructions correspond to buffer descriptors specifying runtime data movements for a design for a DP array. The write instructions are converted into transaction instructions and the read instructions are converted into wait instructions based on the replica of the register address space. The transaction instructions and the wait instructions are included in an instruction buffer. The instruction buffer is provided to a microcontroller configured to execute the transaction instructions and the wait instructions to implement the runtime data movements for the design as implemented in the DP array. In another aspect, the instruction buffer is stored in a file for subsequent execution by the microcontroller.

Abstract: Providing first-in-first-out (FIFO) memory guidance for a multi-processor computing architecture includes compiling a design for a data processing array to generate a compiled design. The compiled design is mapped and routed to the data processing array. The compiled design is simulated using a modified device model of the data processing array. The modified device model uses infinite FIFO models. FIFO memory usage data is generated by tracking amounts of data stored in the infinite FIFO memory models during runtime of the simulation of the compiled design. FIFO memory requirements for one or more nets of the design are determined from the FIFO memory usage data and the compiled design.

Abstract: Methods, systems, and apparatus described herein make a multi-level PAM signal (PAM-N signal) at a transmitter using CMOS-based components. By forming the PAM-N signal at the transmitter, receivers do not have to recombine and/or realign multiple signals and only employs a single transmission line channel (or two transmission line channels in differential implementations) to convey the data stream to the receiver from the transmitter.

Abstract: Systems, methods, and apparatuses are described that enable IC architectures to enable a single anchor to connect to and accept a variety of chiplets at any port by way of a programming model that enables the anchor or chiplet to dynamically adapt to configurations, requirements, or aspects of any coupled component and provide an interface for the coupled components.

Abstract: Acceleration-ready program development includes providing a software library having a plurality of functions having compute identifiers. The software library is associated with a hardware library including one or more hardware accelerated functions. The hardware accelerated functions are associated with the compute identifiers. Each hardware accelerated function is a functional equivalent of a function of the software library having the same compute identifier. A hybrid executor layer is provided that, when executed by a data processing system with an acceleration-ready computer program built using the software library, is configured to initiate execution of a selected function of the acceleration-ready computer program using a processor of the data processing system or invoke a hardware accelerated function having a compute identifier matching the compute identifier of the selected function based on comparing acceleration criteria with acceleration rules.

Abstract: Circuit design development using block design containers can include opening, within a development environment generated by an Electronic Design Automation (EDA) system, a top-level block design specifying a circuit design and inserting, within the top-level block design using the EDA system, a block design container. The block design container specifies a source block design used as a sub-design within the top-level block design.

Abstract: An integrated circuit (IC) includes a video buffer memory and display driver circuitry. The video buffer memory includes a buffer memory map. The video buffer memory stores one or more raster lines of video data organized as tiled lines. Each of the tiled lines including two quartiles. The display driver circuitry is coupled to the video buffer memory. The display driver circuitry writes data associated with a portion of a first data line to a first one of the two quartiles of a first one of the tiled lines, and updates the buffer memory map. Further, the display driver determines a full display line being present within the video buffer memory based on the buffer memory map. The display driver further outputs the full display line to a display device.