Abstract: A disclosed technique includes triggering a change for a first set of one or more functional elements and for a second set of one or more functional elements from a high-power state to a low-power state; saving first state of the first set of one or more functional elements via a first set of one or more save-state elements; saving second state of the second set of one or more functional elements via a second set of one or more save-state elements; powering down the first set of one or more functional elements and the second set of one or more functional elements; and transmitting the first state and the second state to a memory.

Abstract: Systems, apparatuses, and methods for implementing a depth buffer pre-pass are disclosed. A rendering application uses a binning approach to render primitives of a virtual scene on a tile-by-tile basis, with each tile corresponding to a portion of the screen. The application causes a depth buffer pre-pass to be performed for the primitives of the tile before a pixel shader is invoked. During the depth buffer pre-pass, only the depth part of the virtual scene is rendered to determine which pixel samples are visible and which pixel samples are hidden. Then, the scene is redrawn, but the pixel samples that are hidden are not sent to the pixel shader. In cases where a relatively large percentage of primitives overlap, this technique increases the efficiency of the rendering application since pixel shading can be avoided for the pixel samples that are hidden.

Abstract: A method of training a neural network includes, at a local computing node, receiving remote parameters from a set of one or more remote computing nodes, initiating execution of a forward pass in a local neural network in the local computing node to determine a final output based on the remote parameters, initiating execution of a backward pass in the local neural network to determine updated parameters for the local neural network, and prior to completion of the backward pass, transmitting a subset of the updated parameters to the set of remote computing nodes.

Abstract: Various semiconductor chip packages are disclosed. In one aspect, a semiconductor chip package is provided that includes a package substrate that has a first edge and a second edge opposite to the first edge. A semiconductor chip is mounted on the package substrate. A thermal interface material is positioned on the semiconductor chip. A lid is positioned over the thermal interface material. A spring biasing mechanism is included that is operable to bias the lid away from the package substrate so that the lid, when subjected to a compressive force, can translate toward the package substrate and impart a compressive force on the thermal interface material.

Abstract: A processor includes a plurality of vector sub-processors (VSPs) and a plurality of memory banks dedicated to respective VSPs. A first memory bank corresponding to a first VSP includes a first plurality of high vector general purpose register (VGPR) banks and a first plurality of low VGPR banks corresponding to the first plurality of high VGPR banks. The first memory bank further includes a plurality of operand gathering components that store operands from respective high VGPR banks and low VGPR banks. The operand gathering components are assigned to individual threads while the threads are executed by the first VSP.

Abstract: A method and system for recording and logging errors in a computer system includes reading first error handling information with respect to a transaction. The first error handling information is stored in a first component, and based upon a condition of the storage in the first component, an oldest error information is evicted from the first component.

Abstract: A memory controller includes a command queue and an arbiter for selecting entries from the command queue for transmission to a DRAM. The arbiter transacts streaks of consecutive read commands and streaks of consecutive write commands. The arbiter has a current mode indicating the type of commands currently being transacted, and a cross mode indicating the other type. The arbiter is operable to monitor commands in the command queue for the current mode and the cross mode, and in response to designated conditions, send at least one cross-mode command to the memory interface queue while continuing to operate in the current mode. In response to an end streak condition, the arbiter swaps the current mode and the cross mode, and transacts the cross-mode command.

Abstract: Systems, apparatuses, and methods for implementing a graphics processing unit (GPU) coprocessor are disclosed. The GPU coprocessor includes a SIMD unit with the ability to self-schedule sub-wave procedures based on input data flow events. A host processor sends messages targeting the GPU coprocessor to a queue. In response to detecting a first message in the queue, the GPU coprocessor schedules a first sub-task for execution. The GPU coprocessor includes an inter-lane crossbar and intra-lane biased indexing mechanism for a vector general purpose register (VGPR) file. The VGPR file is split into two files. The first VGPR file is a larger register file with one read port and one write port. The second VGPR file is a smaller register file with multiple read ports and one write port. The second VGPR introduces the ability to co-issue more than one instruction per clock cycle.

Abstract: Systems, apparatuses, and methods for conveying and receiving information as electrical signals in a computing system are disclosed. A computing system includes multiple transmitters sending singled-ended data signals to multiple receivers. A termination voltage is generated and sent to the multiple receivers. The termination voltage is coupled to each of signal termination circuitry and signal sampling circuitry within each of the multiple receivers. Any change in the termination voltage affects the termination circuitry and affects comparisons performed by the sampling circuitry. Received signals are reconstructed at the receivers using the received signals, the signal termination circuitry and the signal sampling circuitry.

Abstract: Preserving memory ordering between offloaded instructions and non-offloaded instructions is disclosed. An offload instruction for an operation to be offloaded is processed and a lock is placed on a memory address associated with the offload instruction. In response to completing a cache operation targeting the memory address, the lock on the memory address is removed. For multithreaded applications, upon determining that a plurality of processor cores have each begun executing a sequence of offload instructions, the execution of non-offload instructions that are younger than any of the offload instructions is restricted. In response to determining that each processor core has completed executing its sequence of offload instructions, the restriction is removed. The remote device may be, for example, a processing-in-memory device or an accelerator coupled to a memory.

Abstract: A parallel processing (PP) level coherence directory, also referred to as a Processing In-Memory Probe Filter (PimPF), is added to a coherence directory controller. When the coherence directory controller receives a broadcast PIM command from a host, or a PIM command that is directed to multiple memory banks in parallel, the PimPF accelerates processing of the PIM command by maintaining a directory for cache coherence that is separate from existing system level directories in the coherence directory controller. The PimPF maintains a directory according to address signatures that define the memory addresses affected by a broadcast PIM command. Two implementations are described: a lightweight implementation that accelerates PIM loads into registers, and a heavyweight implementation that accelerates both PIM loads into registers and PIM stores into memory.

Abstract: Techniques for performing cache operations are provided. The techniques include tracking performance events for a plurality of test sets of a cache, detecting a replacement policy change trigger event associated with a test set of the plurality of test sets, and in response to the replacement policy change trigger event, operating non-test sets of the cache according to a replacement policy associated with the test set.

Abstract: Methods, devices, and systems for retrieving information based on cache miss prediction. A prediction that a cache lookup for the information will miss a cache is made based on a history table. The cache lookup for the information is performed based on the request. A main memory fetch for the information is begun before the cache lookup completes, based on the prediction that the cache lookup for the information will miss the cache. In some implementations, the prediction includes comparing a first set of bits stored in the history table with a second set of bits stored in the history table. In some implementations, the prediction includes comparing at least a portion of an address of the request for the information with a set of bits in the history table.

Abstract: A cache includes an upstream port, a downstream port, a cache memory, and a control circuit. The control circuit temporarily stores memory access requests received from the upstream port, and checks for dependencies for a new memory access request with older memory access requests temporarily stored therein. If one of the older memory access requests creates a false dependency with the new memory access request, the control circuit drops an allocation of a cache line to the cache memory for the older memory access request while continuing to process the new memory access request.

Abstract: A heterogeneous processing system includes at least one central processing unit (CPU) core and at least one graphics processing unit (GPU) core. The heterogeneous processing system is configured to compute an activation for each one of a plurality of neurons for a first network layer of a neural network. The heterogeneous processing system randomly drops a first subset of the plurality of neurons for the first network layer and keeps a second subset of the plurality of neurons for the first network layer. Activation for each one of the second subset of the plurality of neurons is forwarded to the CPU core and coalesced to generate a set of coalesced activation sub-matrices.

Abstract: Accesses to a mipmap by a shader in a graphics pipeline are monitored. The mipmap is stored in a memory or cache associated with the shader and the mipmap represents a texture at a hierarchy of levels of detail. A footprint in the mipmap of the texture is marked based on the monitored accesses. The footprint indicates, on a per-tile, per-level-of-detail (LOD) basis, tiles of the mipmap that are expected to be accessed in subsequent shader operations. In some cases, the footprint is defined by a plurality of footprint indicators that indicate whether the tiles of the mipmap are expected to be accessed in subsequent shader operations. In that case, the plurality of footprint indicators are set to a first value to indicate that the tile was not access during the first frame or a second value to indicate that the tile was accessed during the first frame.

Abstract: Techniques for controlling prefetching of instructions into an instruction cache are provided. The techniques include tracking either or both of branch target buffer misses and instruction cache misses, modifying a throttle toggle based on the tracking, and adjusting prefetch activity based on the throttle toggle.

Abstract: A system and method for efficient data transfer in a computing system are described. A computing system includes multiple nodes that receive tasks to process. A bridge interconnect transfers data between two processing nodes without the aid of a system bus on the motherboard. One of the multiple bridge interconnects of the computing system is an optical bridge interconnect that transmits optical information across the optical bridge interconnect between two nodes. The receiving node uses photonic integrated circuits to translate the optical information into electrical information for processing by electrical integrated circuits. One or more nodes switch between using an optical bridge interconnect and a non-optical bridge interconnect based on one or more factors such as measured power consumption and measured data transmission error rates.

Abstract: An integrated circuit includes a plurality of tiles receiving a power supply voltage, each having a corresponding analog circuit and operates in response to a first voltage, and a hardware controller receiving a voltage identification code and provides the first voltage to each of the plurality of tiles in response thereto. The hardware controller comprises a test time controller determining coefficients of a waveform that describes an average correspondence between the power supply voltage and the first voltage for the plurality of tiles, and a boot time controller determining a respective error signal indicating an error between the waveform and a respective actual waveform for each of the plurality of tiles, and providing the respective error signal to the corresponding analog circuit of each of the plurality of tiles. The corresponding analog circuit of each of the plurality of tiles adjusts the first voltage according to the respective error signal.

Abstract: A method and apparatus for managing a controller includes indicating, by a processor of a first device, to the controller of a second device to enter a second power state from a first power state. The controller of the second device responds to the processor of the first device with a confirmation. The processor of the first device transmits a signal to the controller of the second device to enter the second power state. Upon receiving a wake event, the controller of the second device exits the second device from the second power state to the first power state.