Abstract: Integrated circuit layouts are disclosed that include metal layers with metal tracks having separate metal sections along the metal tracks. The separate metal sections along a single track may be electrically isolated from each other. The separate metal sections may then be electrically connected to different voltage tracks in metal layers above and/or below the metal layer with the separate metal sections. One or more of the metal layers in the integrated circuit layouts may also include metal tracks at different voltages (e.g., power and ground) that are adjacent to each other within a power grid layout. The metal tracks may be separated by electrically insulating material. The metal tracks and the electrically insulating material between the tracks may create capacitance in the power grid layout.

Abstract: A technique for recovering from a hang in a virtualized accelerated processing device (“APD”) is provided. In the virtualization scheme, different virtual machines are assigned different “time-slices” in which to use the APD. When a time-slice expires, the APD stops operations for a current VM and starts operations for another VM. To stop operations on the APD, a virtualization scheduler sends a request to idle the APD. The APD responds by completing work and idling. If one or more portions of the APD do not complete this idling process before a timeout expires, then a hang occurs. In response to the hang, the virtualization scheduler informs the hypervisor that a hang has occurred. The hypervisor performs a function level reset on the APD and informs the VM that the hang has occurred. The VM responds by stopping command issue to the APD and re-initializing the APD for the function.

Abstract: A system for performing a scan test of a processor core includes a scan test module and a processor including a processor core and an input/output die, where the input/output die is coupled to the processor core. The scan test module transmits, in parallel to the input/output die, scan test input data. A serializer/deserializer module of the input/output die receives the input data, serializes the input data, and transmits the serialized input data to the processor core. A serializer/deserializer module of the processor core receives the serialized scan test input data, deserializes the input data, receives result data generated in dependence upon the input data, serializes the result data, and transmits the serialized result data to the input/output die. The input/output die serializer/deserializer module receives the result data, deserializes the result data, and provides the result data to the scan test module. Error detection can be carried out through redundancy.

Abstract: Techniques are disclosed for processing cache operations. The techniques include determining a set of cache lines that include data for a vector memory access request; determining bank allocation priorities for the set of cache lines, wherein the bank allocation priorities are chosen to result in the set of cache lines being evenly distributed among the banks; determining actual banks for the set of cache lines; and accessing the cache lines in one or more access iterations, wherein at least one of the one or more access iterations includes accessing multiple cache lines in different banks at the same time.

Abstract: A processor applies a software hint policy to a portion of a cache based on access metrics for different test regions of the cache, wherein each test region applies a different software hint policy for data associated with cache entries in each region of the cache. One test region applies a software hint policy under which software hints are followed. The other test region applies a software hint policy under which software hints are ignored. One of the software hint policies is selected for application to a non-test region of the cache.

Abstract: A memory controller an arbiter which causes streaks of read commands and streaks of write commands over the memory channel. During a streak, the arbiter monitors an indicator of data bus efficiency of the memory channel. Responsive to the indicator showing that data bus efficiency is less than a designated threshold, the arbiter stops the current streak and start a streak of the other type.

Abstract: A method and apparatus for processing color data includes storing fragment pointer and color data together in a color buffer. A delta color compression (DCC) key indicating the color data to fetch for processing is stored, and the fragment pointer and color data is fetched based upon the read DCC key for decompression.

Abstract: A memory controller interfaces with a dynamic random access memory (DRAM) over a memory channel. A refresh control circuit monitors an activate counter which counts a rolling number of activate commands sent over the memory channel to a memory region of the DRAM. In response to the activate counter being above an intermediate management threshold value, the refresh control circuit only issue a refresh management (RFM) command if there is no REF command currently held at the refresh command circuit for the memory region.

Abstract: A system and method for a lightweight fence is described. In particular, micro-operations including a fencing micro-operation are dispatched to a load queue. The fencing micro-operation allows micro-operations younger than the fencing micro-operation to execute, where the micro-operations are related to a type of fencing micro-operation. The fencing micro-operation is executed if the fencing micro-operation is the oldest memory access micro-operation, where the oldest memory access micro-operation is related to the type of fencing micro-operation. The fencing micro-operation determines whether micro-operations younger than the fencing micro-operation have load ordering violations and if load ordering violations are detected, the fencing micro-operation signals the retire queue that instructions younger than the fencing micro-operation should be flushed. The instructions to be flushed should include all micro-operations with load ordering violations.

Abstract: Methods for reducing boot time of a system-on-a-chip (SOC) by reducing double data rate (DDR) memory training and memory context restore. Dynamic random access memory (DRAM) controller and DDR physical interface (PHY) settings are stored into a non-volatile memory and the DRAM controller and DDR PHY are powered down. On system resume, a basic input/output system restores the DRAM controller and DDR PHY settings from non-volatile memory, and finalizes the DRAM controller and DDR PHY settings for operation with the SOC. Reducing the boot time of the SOC by reducing DDR training includes setting DRAMs into self-refresh mode, and programing a self-refresh state machine memory operation (MOP) array to exit self-refresh mode and update any DRAM device state for the target power management state. The DRAM device is reset, and the self-refresh state machine MOP array reinitializes the DRAM device state for the target power management state.

Abstract: A graphics processing unit (GPU) schedules recurrent matrix multiplication operations at different subsets of CUs of the GPU. The GPU includes a scheduler that receives sets of recurrent matrix multiplication operations, such as multiplication operations associated with a recurrent neural network (RNN). The multiple operations associated with, for example, an RNN layer are fused into a single kernel, which is scheduled by the scheduler such that one work group is assigned per compute unit, thus assigning different ones of the recurrent matrix multiplication operations to different subsets of the CUs of the GPU. In addition, via software synchronization of the different workgroups, the GPU pipelines the assigned matrix multiplication operations so that each subset of CUs provides corresponding multiplication results to a different subset, and so that each subset of CUs executes at least a portion of the multiplication operations concurrently.

Abstract: Platform power management includes boosting performance in a platform power boost mode or restricting performance to keep a power or temperature under a desired threshold in a platform power cap mode. Platform power management exploits the mutually exclusive nature of activities and the associated headroom created in a temperature and/or power budget of a server platform to boost performance of a particular component while also keeping temperature and/or power below a threshold or budget.

Abstract: A method of context bouncing includes receiving, at a command processor of a graphics processing unit (GPU), a conditional execute packet providing a hash identifier corresponding to an encapsulated state. The encapsulated state includes one or more context state packets following the conditional execute packet. A command packet following the encapsulated state is executed based at least in part on determining whether the hash identifier of the encapsulated state matches one of a plurality of hash identifiers of active context states currently stored at the GPU.

Abstract: Systems, apparatuses, and methods for arbitrating threads in a computing system are disclosed. A computing system includes a processor with multiple cores, each capable of simultaneously processing instructions of multiple threads. When a thread throttling unit receives an indication that a shared cache has resource contention, the throttling unit sets a threshold number of cache misses for the cache. If the number of cache misses exceeds this threshold, then the throttling unit notifies a particular upstream computation unit to throttle the processing of instructions for the thread. After a time period elapses, if the cache continues to exceed the threshold, then the throttling unit notifies the upstream computation unit to more restrictively throttle the thread by performing one or more of reducing the selection rate and increasing the time period. Otherwise, the unit notifies the upstream computation unit to less restrictively throttle the thread.

Abstract: Systems, apparatuses, and methods for enabling indirect chaining of command buffers are disclosed. A system includes at least first and second processors and a memory. The first processor generates a plurality of command buffers and stores the plurality of command buffers in the memory. The first processor also generates and stores, in the memory, a table with entries specifying addresses of the plurality of command buffers and an order in which to process the command buffers. The first processor conveys an indirect buffer packet to the second processor, where the indirect buffer packet specifies a location and a size of the table in the memory. The second processor retrieves an initial entry from the table, processes a first command buffer at the address specified in the initial entry, and then returns to the table for the next entry upon completing processing of the first command buffer.

Abstract: A technique for performing a ray intersection test, the method comprising: receiving a request for an early termination ray intersection test for a ray; testing the ray against one or more early termination box nodes and one or more normal box nodes of a bounding volume hierarchy; and based on the test of the ray against the one or more early termination box nodes, determining whether to end traversal of the bounding volume hierarchy and determine whether the ray intersects geometry for the purpose of the ray intersection test.

Abstract: A method and apparatus for predicting and managing a device failure includes responsive to a predicted failure of a memory device, the predicted failure based on sensor data associated with the memory device, determining a further action for the memory device.

Abstract: Systems, apparatuses, and methods for employing system probe filter aware last level cache insertion bypassing policies are disclosed. A system includes a plurality of processing nodes, a probe filter, and a shared cache. The probe filter monitors a rate of recall probes that are generated, and if the rate is greater than a first threshold, then the system initiates a cache partitioning and monitoring phase for the shared cache. Accordingly, the cache is partitioned into two portions. If the hit rate of a first portion is greater than a second threshold, then a second portion will have a non-bypass insertion policy since the cache is relatively useful in this scenario. However, if the hit rate of the first portion is less than or equal to the second threshold, then the second portion will have a bypass insertion policy since the cache is less useful in this case.

Abstract: A controller assigns variable length addresses to addressable elements that are connected to a network. The variable length addresses are determined based on probabilities that packets are addressed to the corresponding addressable element. The controller transmits, to the addressable elements via the network, a routing table indicating the variable length addresses assigned to the addressable elements. Routers or addressable elements receive the routing table and route one or more packets over the network to an addressable element using variable length addresses included in a header of the one or more packets.

Abstract: An integrated circuit includes a plurality of functional blocks. Utilization information for the various functional blocks is generated. Based on that information, the power consumption and thus the performance levels of the functional blocks can be tuned. Thus, when a functional block is heavily loaded by an application, the performance level and thus power consumption of that particular functional block is increased. At the same time, other functional blocks that are not being heavily utilized and thus have lower performance requirements can be kept at a relatively low power consumption level. Thus, power consumption can be reduced overall without unduly impacting performance.