Nvidia Corporation

Colyn Scott Case

Sean Jeffrey Treichler

Dmitry Vyshetsky

William Tsu

Donald A Bittel

Manas Mandal

Peter Tong

Sharon Rose Clay

Gary D Lorensen

David Brian Glasco

Steven E Molnar

Emmett Michael Kilgariff

John H Edmondson

James Leroy Deming

David G Reed

Anthony Michael Tamasi

Dane Mrazek

Anders M Kugler

William Craig McKnight

Mark Allen Mosley

Sonny S Yeoh

Paul MacDougal

Ashish Kaul

David Alan Jarosh

Emmett M Kilgrariff

Yin Fung (David) Tang

Circuits, methods, and apparatus for modifying the data rate of a data bus. In a circuit having two processors coupled by a data bus, the processors each learn that the other is capable of operating at a modified data rate. The data rate is then changed to the modified rate. Each processor may learn of the other's capability by reading a vendor identification, for example from a vendor defined message stored on the other processor. Alternately, each processor may provide an instruction to the other to operate at the modified rate, for example by writing to the other processor's extended capability registers. In another circuit having two processors communicating over a bus, it is determined that both are capable of transmitting and receiving data at a modified data rate. An instruction is provided to one or both of the processors to transmit at the modified rate.

PCIE clock rate stepping for graphics and platform processors

One embodiment of the present invention sets forth a crossbar unit that is coupled to a plurality of client subsystems. The crossbar unit is configured to transmit data packets between the client subsystems and includes a high-bandwidth channel and a narrow-bandwidth channel. The high-bandwidth channel is used for transmitting large data packets, while the narrow-bandwidth is used for transmitting smaller data packets. The transmission of data packets may be prioritized based on the source and destination clients as well as the type of data being transmitted. Further, the crossbar unit includes a buffer mechanism for buffering data packets received from source clients until those data packets can be received by the destination clients.

Consolidated crossbar that supports a multitude of traffic types

One embodiment of the present invention sets forth a system and method for supporting high-throughput virtual to physical address translation using compressed TLB cache lines with variable address range coverage. The amount of memory covered by a TLB cache line depends on the page size and page table entry (PTE) compression level. When a TLB miss occurs, a cache line is allocated with an assumed address range that may be larger or smaller than the address range of the PTE data actually returned. Subsequent requests that hit a cache line with a fill pending are queued until the fill completes. When the fill completes, the cache line's address range is set to the address range of the PTE data returned. Queued requests are replayed and any that fall outside the actual address range are reissued, potentially generating additional misses and fills.

Efficient memory translator with variable size cache line coverage

Devices and methods provide data from multiple storage locations to a processor. A data block containing data required by a processor is stored in two or more locations, e.g., in a local memory and a system memory, both of which are accessible to the processor's memory interface. The memory interface directs each read request for mirrored data to one or another of the mirror locations. Selection of a mirror location to be read is based on substantially real-time information about which mirror location is best able to handle the request. For instance, the selection of a mirror location to access can be based at least in part on information about the level of activity on various buses that connect the processor to the mirror locations.

Memory interface with dynamic selection among mirrored storage locations

A raster operations (ROP) unit interleaves read and write requests for efficiently communicating with a frame buffer via a PCI Express (PCI E) link or other system bus that provides separate upstream and downstream data transfer paths. One example of a ROP unit processes pixels in groups, performing read modify writeback sequences for each group. The read requests associated with pixels in a second group are advantageously interleaved with the writeback requests for pixels in the first group prior to sending the requests on the system bus.

Raster operations unit with interleaving of read and write requests using PCI express

In a virtual memory system, address translation information is provided using a cluster that is associated with some range of virtual addresses and that can be used to translate any virtual address in its range to a physical address, where the sizes of the ranges mapped by different clusters may be different. Clusters are stored in an address translation table that is indexed by virtual address so that, starting from any valid virtual address, the appropriate cluster for translating that address can be retrieved from the translation table. The clusters are dynamically created from a fragmented pool of physical addresses as new virtual address mappings are requested by consumers of the virtual memory space.

Memory management for virtual address space with translation units of variable range size

A bus interface permits an upstream bandwidth and a downstream bandwidth to be separately selected. In one implementation a link control module forms a bidirectional link with another bus interface by separately configuring link widths of an upstream unidirectional sub-link and a downstream unidirectional sub-link.

Asymmetrical bus for bus link width optimization of a graphics system

In some applications, such as video motion compression processing for example, a request pattern or 'stream' of requests for accesses to memory (e.g., DRAM) may have, over a large number of requests, a relatively small number of requests to the same page. Due to the small number of requests to the same page, conventionally sorting to aggregate page hits may not be very effective. Reordering the stream can be used to 'bury' or &#x201c;hide&#x201d; much of the necessary precharge/activate time, which can have a highly positive impact on overall throughput. For example, separating accesses to different rows of the same bank by at least a predetermined number of clocks can effectively hide the overhead involved in precharging/activating the rows.

Page stream sorter for poor locality access patterns

Optimization logic that optimizes a stream of requests being transmitted onto a link by a link interface unit can be enabled or disabled based on a performance metric that represents a measure of the degree to which a response to a request is likely to be slowed due to congestion, propagation delays, or other bottlenecks in the system. For example, the performance metric can be based on a measured level of link activity due to requests from the transmitting device and/or a prediction as to behavior (e.g., access time) of the target device that receives the stream of requests. The control logic advantageously does not require extra signals to be carried on the bus.

Control device for data stream optimizations in a link interface

A virtual address translation table and an on-chip address cache are usable for translating virtual addresses to physical addresses. Address translation information is provided using a cluster that is associated with some range of virtual addresses and that can be used to translate any virtual address in its range to a physical address, where the sizes of the ranges mapped by different clusters may be different. Clusters are stored in an address translation table that is indexed by virtual address so that, starting from any valid virtual address, the appropriate cluster for translating that address can be retrieved from the translation table. Recently retrieved clusters are stored in an on-chip cache, and a cached cluster can be used to translate any virtual address in its range without accessing the address translation table again.

Virtual address translation system with caching of variable-range translation clusters

Multiple data transfer requests can be merged and transmitted as a single packet on a packetized bus such as a PCI Express (PCI-E) bus. In one embodiment, requests are combined if they are directed to contiguous address ranges in the same target device. An opportunistic merging procedure is advantageously used that merges a first request with a later request if the first request and the later request are mergeable and are received within a holdoff period that is dynamically determined based on a level of bus activity; otherwise, requests can be transmitted without merging.

Packet combiner for a packetized bus with dynamic holdoff time

Asymmetrical bus

A cache shared by multiple clients implements a client specific policy for replacing entries in the event of a cache miss. A request from any client can hit any entry in the cache. For purposes of replacing entries, at least of the clients is restricted, and when a cache miss results from a request by the restricted client, the entry to be replaced is selected from a fixed subset of the cache entries. When a cache misses results from a request by any client other than the restricted client, any cache entry, including a restricted entry, can be selected to be replaced.

Growing community of inventors

Hyde Park, VT, United States of America

Colyn Scott Case