Reading extended data burst from memory

Abstract

The invention relates to a method for reading data with a data burst length (BL) greater than two from a semiconductor memory apparatus, comprising the following steps:

Description

DESCRIPTION

[0001] The invention relates to a method for reading data with a data burst length BL>2 from a semiconductor memory apparatus in accordance with claim 1.

[0002] Read latency is increasingly a problem for semiconductor memories operated at high frequency. Particularly for fast network memories intended to act, by way of example, for the purpose of temporary buffer-storage of a 10 Gbit/s data stream over a few milliseconds, semiconductor memories in which data are output as closely as possible to the time of the read command are required. The read latency, which represents the period of time which elapses between the read command and the first data bit on the data bus, is comparatively high, particularly in the case of DRAM memory architectures. Buffer-stores used for a network processor unit (NPU) are therefore often SRAM memory chips, which have relatively low read latency. However, such SRAM memory chips have the drawback that they cannot be produced using very large scale integration and are therefore comparatively expensive.

[0003] In view of these problems, it is an object of the invention to specify a method for reading data from a semiconductor memory apparatus which allows the read latency to be reduced and therefore relaxed.

[0004] This object is achieved by a method having the steps specified in claim 1. Preferred embodiments are covered by the dependent claims.

[0005] In line with the invention, a method for reading data with a data burst length BL>2 from a semiconductor memory apparatus comprises the following steps:

[0006] a first address part for memory cells which are to be addressed in the semiconductor memory apparatus is received from an address bus in the semiconductor memory apparatus at a time t0;

[0007] a read command is transferred to a command bus in the semiconductor memory apparatus in order to initiate read access to the memory cells at a time t1, which is later in time than the time t0;

[0008] a second address part for the memory cells which are to be addressed is received from the address bus at the time t1; and

[0009] data associated with the addressed memory cells are transferred to a data bus in the semiconductor memory apparatus at a time t2, which is later than the time t1.

[0010] The inventive method allows the read latency to be relaxed for semiconductor memory apparatuses which are read with a data burst length BL>2. In this context, a burst length of >2 means that data are output during at least two clock periods of a clock signal in the semiconductor memory apparatus. A fundamental feature of the inventive method is that some of the address of memory cells which are to be addressed is received in the semiconductor memory apparatus from the address bus for the purpose of decoding the address before the actual read command which initiates read access. In this regard, the first address part is made available on the address bus early enough before the read command for the first address part to be able to be received from the address bus before the actual read command. In particular, row and column decoding devices in the semiconductor memory apparatus may already have obtained the first address part at a time at which there is still no associated read command on the command bus. At the time t0, the first address part is thus “fetched” or received from the address bus while the second address part is “fetched” or received at the later time t1. The cell array in the semiconductor memory apparatus can thus actually be accessed one clock cycle before the read command, for example.

[0011] Conventionally, methods for reading data from semiconductor memory apparatuses with burst lengths BL>2 in the address bus (and command bus) involve an “NOP” (no operation) being executed between two successive access operations when full use is to be made of the data bus's time. The address bus is thus not in optimum use between successive read access operations to the cell array in the semiconductor memory apparatus.

[0012] In line with the invention, at least some of the full addresses of memory cells which are to be addressed are received before the actual read command. This allows the semiconductor memory apparatus's read latency to be shortened, which results in novel areas of use for such a semiconductor memory apparatus, particularly as a buffer-store for an NPU.

[0013] Preferably, the semiconductor memory apparatus is a DRAM, particularly an “RLDRAM” (reduced latency DRM). The data are preferably provided on the data bus at double data rate, which means that a “DDR” memory chip is involved.

[0014] The method preferably comprises the step of generating a cyclic clock signal, and the address parts and the read command are transferred and received at times of edges of the clock signal. Preferably, the address parts and the read command are transferred and received at times of rising edges of the clock signal. The address information is preferably available on the address bus during one clock cycle. By way of example, the address parts are provided on the address bus for one clock period from the time of a falling edge of the clock signal onward and are received in the semiconductor memory apparatus from the address bus upon the next rising edge of the clock signal.

[0015] Preferably, the time t0 is one clock cycle before the time t1. This means that the first address part is actually received one cycle before the actual read command. The second address part is preferably received from the address bus at the time of the read command, i.e. at the time t1. Hence, “prefetch addressing” of the memory cells which are to be addressed takes place before the actual read command.

[0016] The first address part preferably comprises row addresses. On account of the “rectangular” cell array architectures often used with DRAM memory concepts, activating (“opening”) rows in the memory cell array often requires more time than activating (“opening”) columns. Hence, at least some of the row addresses in the first address part, which is transferred before the actual read command, are preferably transferred, which means that more time is available to activate or open them. In this case, the column addresses are preferably transferred with the second address part, since activating these takes less time.

[0017] Preferably, the first address part comprises bank addresses and at least some block addresses. By way of example, one preferred memory architecture comprises 8 banks with 16 blocks each. Preferably, the bank addresses (for example 3 bit addresses for 8 banks) and at least some of the block addresses (for example 2 bits of 4 bit addresses) are actually transferred with the first address part. In this case, the rest of the block addresses are transferred with the second address part.

[0018] Preferably, the second address part comprises the column addresses of the memory cells which are to be addressed. Since the time needed for activating or opening rows is longer in typical DRAM memory architectures (rectangular cell array) than the corresponding activation time for columns, it is also possible to transfer the column addresses in the second address part, which is later in time, without increasing the read latency.

[0019] Preferably, the second address part comprises at least some block addresses. As already illustrated above, provision can be preferably be made for some of the block addresses actually to be transferred with the first address part. The rest of the block addresses (for example 2 bits of 4 bit block addresses) are transferred with the second address part.

[0020] The invention is described below by way of example with reference to preferred embodiments and accompanying drawings, in which:

[0021] FIG. 1 shows a known method for reading data from a semiconductor memory apparatus with a burst length BL=2;

[0022] FIG. 2 shows a schematic timing diagram for a known method for reading data from a semiconductor memory apparatus with a burst length BL=4; and

[0023] FIG. 3 shows a timing diagram for a preferred inventive method for reading data from a semiconductor memory apparatus with a burst length BL=4.

[0024] FIG. 1 shows a schematic timing diagram for a conventional method for reading data from a semiconductor memory apparatus. “CLK” denotes a clock signal, “CMD” denotes the command bus, “ADR” denotes the address bus and “data” denotes the data bus. To simplify understanding, FIG. 1 (and the rest of the figures) shows only the command, address and data signals for two read commands (R1, R2).

[0025] A read command R1 is shown by an arrow on the command bus CMD. The read command is executed simultaneously with a rising edge of the clock signal CLK. During a clock cycle from one falling edge to the next falling edge of the clock signal, address information AR1 is available on the address bus ADR and is received from the address bus ADR at the time of the read command R1 for the purpose of address decoding. The first data bit is provided on the data bus after a read latency RL of two clock periods, i.e. after a delay time of two clock periods following the read command R1. The data associated with the first read command R1 are provided on the data bus at double data rate (DDR). The data DR10 and DR11 are thus transmitted on the rising and falling edges within a clock signal period.

[0026] The timing diagram in FIG. 1 shows that, for a burst length BL=2, when the data bus is in full use (information written to the data bus all the time), the address and command buses are also in full use. Address information AR2 for a subsequent read command R2 is thus transferred directly after the addresses AR1. In other words, the addresses AR2 for the second read command R2 are provided on the address bus ADR during a clock cycle which directly follows the clock period in which the address information AR1 for the first read command R1 was provided.

[0027] This situation takes a different turn with a known method for reading data from a semiconductor memory apparatus in which a burst length BL=4 is used. If full use is to be made of the data bus in this case, i.e. the data DR1[0:3] and the data DR2[0:3] are intended to follow one another directly in time, then full use can no longer be made of the address bus ADR's time. Instead, a “dead time” results on the address bus ADR between the transferred addresses AR1, which are associated with the first read command R1, and the addresses AR2, which are associated with the second read command R2. During this dead time (NOP), no address information is provided on the address bus.

[0028] A preferred inventive method for reading data from a semiconductor memory apparatus is shown in FIG. 3. In the case of this embodiment, data are read with a burst length of BL=4. The data are provided on the data bus at double data rate (DDR), which means that full use is made of the data bus's time, i.e. information is written to it all the time. Unlike the method described in connection with FIG. 2, however, the method illustrated in FIG. 3 involves full use being made of the address bus ADR's time. In the preferred method, a first address part AR10, which is associated with the first read command R1, is received in the semiconductor memory apparatus from the address bus ADR at a time t0 before the read command R1 is applied to the command bus CMD. Hence, the first address part AR10 is actually received one clock cycle before the read command R1, which is executed at a time t1. At the time t1, which is one clock period after the time t0, the second address part AR11 for the first read command R1 is received. The first and second address parts AR10 and AR11 are respectively provided on the address bus ADR for one clock period half a clock period before the reception times t0 and t1.

[0029] The read latency RL, denoting the period of time between read access and the first data bit on the data bus, can be reduced by one cycle as a result of this “prefetch addressing”. Accordingly, the read latency RL is 2 in the inventive embodiment illustrated in FIG. 3, whereas RL=3 in the case of the conventional reading method in FIG. 2. The data DR1[0:3] associated with the first read command R1 are provided on the data bus at double data rate from the time t2 onward. Full use of the address bus ADR's time allows a reduction in the read latency to be achieved as a result of reception of a first address part AR10 (or generally ARX0) before the actual associated read command R1 (or generally RX).

[0030] A description will now be given of an adaptation of the inventive reading method to an 8×2×16 Mbit DRAM memory array. Row addressing (WL selection) requires, by way of example, 3 bits for bank addresses, 4 bits for block addresses and 9 bits for row addresses. If there are 13 signals available, 3 bits+2 bits+9 bits thereof can actually be transferred one clock cycle earlier. Hence, by way of example, all the bank addresses BANK<2:0> and all the row addresses AROW<8:0> can actually be transferred with the first address part. However, the blocks BLOCK<3:0> are addressed by the first address part only incompletely, since just 2 bits of 4 bits of the block addresses are transferred. The result of this is that 4 rows (WS) need to be opened in one bank, with the word line which is to be opened not being able to be selected until the last 2 bits of the block addresses have been transferred. Alternatively, the word lines can also be opened first, with the rest of the addresses (2 bits) being taken into account only when the MDQS switch opens (i.e. when the columns are read). The latter addresses are thus converted into one column address as it were.

[0031] The advantages of the inventive method are particularly a reduction in the read latency by one clock cycle, which is of particular significance for low frequencies. Even if a full address is not available for opening a WL, then at least some of the address decoding can be brought forward. In addition, the number of address pins is reduced by a factor of two.

Claims

1. A method for reading data with a data burst length (BL) greater than two from a semiconductor memory apparatus, comprising the following steps: a first address part (AR10) for memory cells which are to be addressed in the semiconductor memory apparatus is received from an address bus (ADR) in the semiconductor memory apparatus at a time t0; a read command (R1) is transferred to a command bus (CMD) in the semiconductor memory apparatus in order to initiate read access to the memory cells at a time t1, which is later in time than the time t0; a second address part (AR11) for the memory cells which are to be addressed is received from the address bus (ADR) at the time t1; and data associated with the addressed memory cells are transferred to a data bus (Data) in the semiconductor memory apparatus at a time t2, which is later than the time t1.

2. The method as claimed in claim 1, where the semiconductor memory apparatus is an RLDRAM.

3. The method as claimed in claim 1 or 2, where the method comprises the step of producing a cyclic clock signal (CLK), and the address parts (AR10, AR11) and the read command (R1) are transferred and received at times of edges of the clock signal (CLK).

4. The method as claimed in claim 3, where the time t0 is one clock cycle before the time t1.

5. The method as claimed in one of the preceding claims, where the first address part (AR10) comprises row addresses.

6. The method as claimed in one of the preceding claims, where the first address part (AR10) comprises bank addresses and at least some block addresses.

7. The method as claimed in one of the preceding claims, where the second address part (AR11) comprises column addresses.

8. The method as claimed in one of the preceding claims, where the second address part (AR11) comprises at least some block addresses.