# US 2024/0004792 A1 · Data l2 cache with split access

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2024/0004792 A1
- Document ID: US20240004792A1
- Status: Published application
- Kind: A1
- Application number: 17/809,886
- Filed: 2022-06-29
- Published: 2024-01-04
- Earliest claimed: 2022-06-29
- Assignees: Ampere Computing LLC
- Applicants: Ampere Computing LLC
- Inventors: Rahul NADKARNI
- Web page: https://patentgrep.com/patent/US20240004792A1

## CPC

- G06F 12/0811: with multilevel cache hierarchies
- G06F 12/0895: of parts of caches, e.g. directory or tag array
- G06F 12/0855: Overlapped cache accessing, e.g. pipeline (G06F12/0846 takes precedence)
- G06F 12/0864: using pseudo-associative means, e.g. set-associative or hashing
- G06F 2212/1008: Correctness of operation, e.g. memory ordering
- G06F 2212/1024: Latency reduction
- G06F 2212/507: using speculative control
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)

## Abstract

A memory with data array (e.g., L2 cache) addressable in rows and columns and techniques to access data therein are proposed. Unlike conventional data arrays, the proposed memory allows data access to be initiated based on a row (or set) address even though the column (or way) address is not yet available. When the column address is determined, it can be used to select the correct data. Since the data access is started prior to determining the column address, memory access latency is reduced.

## Claims

### Claim 1

A memory, comprising:
a data array comprising a plurality of data blocks addressable in rows and columns; and
an access circuit configured to receive as inputs clock (CLK) signals and a read/write operation (RD_WRB) signal,
wherein when the RD_WRB signal indicates a read operation, the access circuit is configured to:
initiate a data read from the data array during a first cycle based on a row address provided to the memory prior to the first cycle, the first cycle being a period of a first CLK signal, and
output a read data block at least by an end of a second cycle, the read data block being a data block of the plurality of data blocks corresponding to the row address and a column address, the second cycle being a period of a second CLK signal and subsequent to the first cycle, and
wherein the column address is not available prior to the first cycle, and
wherein the data array is a single memory instance.

### Claim 2 (depends on 1)

The memory of claim 1,
wherein the data array is N-way associative such that there are M×N data blocks, M being a number of row addresses and N being a number of column addresses, and
wherein the row address is m, m=0 . . . M−1, and
wherein during the first cycle, the access circuit is configured to read N data blocks corresponding to the row address m.

### Claim 3 (depends on 2)

The memory of claim 2, wherein during the first and second cycles, the access circuit is configured to:
compare one or more of N tags associated with the N data blocks with an input tag provided to the memory,
determine, when the comparison indicates that an nth tag of the N tags matches the input tag, an address n of the matching tag as the column address, n=0 . . . N−1, and
output data block n of the N data blocks as the read data block.

### Claim 4 (depends on 3)

The memory of claim 3, wherein the access circuit is configured to output the read data block during the second cycle.

### Claim 5 (depends on 3)

The memory of claim 3, wherein the access circuit is configured to multiplex the data block n of the N data blocks.

### Claim 6 (depends on 3)

The memory of claim 3, wherein in comparing the one or more N tags, the access circuit is configured to write to a hit/miss word comprising N bits, each bit of the hit/miss word corresponding to one of the N tags, each bit of the hit/miss word indicating whether the corresponding tag matches or does not match the input tag.

### Claim 7 (depends on 6)

The memory of claim 6, wherein the access circuit is configured to multiplex the data block n of the N data blocks directly from the hit/miss word without further encoding and decoding of the hit/miss word.

### Claim 8 (depends on 1)

The memory of claim 1, wherein when the RD_WRB signal indicates a write operation, the access circuit is configured to:
initiate and complete a data write to the data array during one cycle based on a row address provided to the memory prior to the one cycle, the one cycle being a period of a CLK signal,
wherein the data write is a write of a data block corresponding to the row address and a column address, the data block being provided to the memory prior to the one cycle.

### Claim 9 (depends on 1)

The memory of claim 1, wherein the memory is a level-2 (L2) cache.

### Claim 10 (depends on 1)

The memory of claim 1, wherein the first and second cycles are successive cycles.

### Claim 11

A method of accessing a memory comprising a data array comprising a plurality of data blocks addressable in rows and columns, the method comprising:
receiving clock (CLK) signals and a read/write operation (RD_WRB) signal; and
performing a memory read when the RW_WRB signal indicates a read operation, the memory read comprising:
initiating a data read from the data array during a first cycle based on a row address provided to the memory prior to the first cycle, the first cycle being a period of a first CLK signal; and
outputting a read data block at least by an end of a second cycle, the read data block being a data block of the plurality of data blocks corresponding to the row address and a column address, the second cycle being a period of a second CLK signal and subsequent to the first cycle,
wherein the column address is not available prior to the first cycle, and
wherein the data array is a single memory instance.

### Claim 12 (depends on 11)

The method of claim 11,
wherein the data array is N-way associative such that there are M×N data blocks, M being a number of row addresses and N being a number of column addresses, and
wherein the row address is m, m=0 . . . M−1, and
wherein initiating the data read comprises reading N data blocks corresponding to the row address m during the first cycle.

### Claim 13 (depends on 12)

The method of claim 12, wherein outputting the read data block comprises:
comparing one or more of N tags associated with the N data blocks with an input tag provided to the memory;
determining, when the comparison indicates that an nth tag of the N tags matches the input tag, an address n of the matching tag as the column address, n=0 . . . N−1; and
outputting data block n of the N data blocks as the read data block.

### Claim 14 (depends on 13)

The method of claim 13, wherein the read data block is output during the second cycle.

### Claim 15 (depends on 13)

The method of claim 13, wherein outputting the read data block comprises multiplexing the data block n of the N data blocks.

### Claim 16 (depends on 13)

The method of claim 13, wherein in determining the address n of the matching tag comprises:
writing to a hit/miss word comprising N bits, each bit of the hit/miss word corresponding to one of the N tags, each bit of the hit/miss word indicating whether the corresponding tag matches or does not match the input tag.

### Claim 17 (depends on 16)

The method of claim 16, wherein the outputting the data block n comprises multiplexing the data block n of the N data blocks directly from the hit/miss word without further encoding and decoding of the hit/miss word.

### Claim 18 (depends on 11)

The method of claim 11, further comprising:
performing a memory write when the RW_WRB signal indicates a write operation, the memory write comprising:
initiating and completing a data write to the data array during one cycle based on a row address provided to the memory prior to the one cycle, the one cycle being a period of a CLK signal,
wherein the data write is a write of a data block corresponding to the row address and a column address, the data block being provided to the memory prior to the one cycle.

### Claim 19 (depends on 11)

The method of claim 11, wherein the memory is a level-2 (L2) cache.

### Claim 20 (depends on 11)

The method of claim 11, wherein the first and second cycles are successive cycles.
