Pipeline structure of memory for high-fast row-cycle

Details Pipeline structure of memory for high-fast row-cycle Pipeline structure of memory for high-fast row-cycle

2001-04-06

2003-04-08

Lebentritt, Michael S. (Department: 2824)

Static information storage and retrieval

Addressing

Multiple port access

C365S230060, C365S230080

Reexamination Certificate

active

06545936

ABSTRACT:

BACKGROUND OF THE INVENTION
1. Field of the Invention
The present invention relates to a DRAM structure for reducing row latency for an irregular row access and for improving the effective bandwidth by varying a DRAM cell core structure. Specifically, the invention relates to a pipeline structure of a memory for a fast row cycle, which is different from a structure used in a conventional fast cycle RAM (FCRAM) and is established by modifying a cell core access in the channel structure of a virtual channel memory (VCM) that has been known to have its performance is improved when applied to a conventional structure and by introducing a row buffer and a latch to a decoder, to secure a sufficient bandwidth for sequential accesses and, simultaneously, to reduce a period of time required for a row path through application of a new technique to obtain short latency, thereby realizing a fast random row cycle.
2. Description of the Related Art
In general, the performance of a DRAM can be represented by a bandwidth and latency. The bandwidth means the amount of information that can be transmitted at a time. The bandwidth is proportioned to the frequency of asignal transmitted and the number of signal lines and it is determined by the column path of the DRAM. Accordingly, it is preferable that the DRAM has as high bandwidth as possible. For this, there have been proposed a variety of methods of making the DRAM column path short. Meanwhile, the latency means a period of time starting from input of a specific address for accessing the DRAM to output of data from the DRAM. The latency is determined mostly by the row path of the DRAM. The DRAM has better performance when the latency is smaller because it is preferable that the data is output fast.
When the DRAM used in a system is sequentially accessed, the maximum bandwidth of the DRAM can be sufficiently used in terms of the structural characteristic of the DRAM. In recent computer systems, however, there has been proposed is an unified memory architecture (UMA) which uses a system memory as a frame buffer for decreasing costs and there exist many masters directly accessing a main memory in a single system, such as cache controller, PCI controller, graphic controller, etc.
Furthermore, the form of memory access becomes irregular with an increase in the number of software developed in a language having a modularized form like C++. In this case, when the latency of the DRAM is longer, a period of time during which data is output for one access becomes longer, resulting in a remarkable reduction in the effective bandwidth. This means that the latency is very important element for telling the performance of the DRAM together with the bandwidth.
The latency of the DRAM is determined by the row path as described above. The row path is restricted by physical factors such as RC time constant of line, compared to the column path affecting the bandwidth so that it is difficult to reduce. Accordingly, there have been proposed various techniques for preventing the latency from being increased or for reducing the latency. These techniques include a multi-bank and access sequence controlling method, a method of using a temporary buffer, an address non-multiplexing and row path pipeline technique and a method of integrating an SRAM in a DRAM.
In the multi-bank and access sequence controlling method, when there are continuous accesses between different banks among multiple banks, page miss penalty is reduced through bank interleaving and the DRAM access sequence is controlled to improve the effective bandwidth. However, this method does not decrease the latency by reducing the time in the row path but puts many banks inside the DRAM to overlap DRAM access operations to thereby obtain an effect equivalent to a reduction in the latency. Accordingly, though the latency is decreased when accesses occur for different banks, the conventional latency is reflected on the output when sequential accesses are required for the same bank. In addition, the multiple inner banks deteriorate noise characteristic.
Furthermore, the aforementioned method of using a temporary buffer is to change the structure of a cell core to directly reduce the time required for the row path, to thereby decrease the latency. This method employs a temporary data buffer to a sense amplifier. Specifically, this technique uses the temporary data buffer to minimize the signal amplitude of a bit line to thereby reduce a pre-charge time and data detection time, and introduces a pipeline concept to the row path to realize the cycle time of 10 nsec for an irregular row access. However, the method of using the temporary buffer is still a conceptual idea so that techniques for operating it are not prepared yet. In addition, this method cannot obtain fast row latency when applied to an actual system because it has problems of restoring data stored in the temporary buffer to the cell core and complexity of a controller for controlling the restoration.
The address non-multiplexing and row path pipeline method is to introduce the pipeline concept to the row path and adopts address non-multiplexing to realize the row cycle of 20 ns. A fast cycle RAM (FCRAM) employs this method. The address non-multiplexing and row path pipeline method selects a sub-wordline structure, reduces the sub-block size of the cell core to decrease the load of a driver driving the cell core and shortens the period of time required for the row path by using a direct sensing method. However, this technique makes the sub-block size very small in order to reduce the cell core access time and adds many additional circuits, resulting in an increase in the area by 30-40%, compared to other DRAM structures having the same integrity. Furthermore, it uses the non-multiplexing method as an address input method so that it cannot be compatible with existing systems in terms of interface, for example, the number of pins for address input. Moreover, since its data outputting method is different from the conventional DRAM structure, an additional interface circuit is required to the data output port in order to apply this method to a system currently being used.
Finally, the method of integrating an SRAM in a DRAM, which is different from the above-mentioned three methods, is to integrate the SRAM in the DRAM to secure a sufficient bandwidth even for an irregular row access. This basically uses temporal and spatial locality of a memory access pattern. Specifically, the SRAM integrated together with the DRAM is used as a data buffer and the operations of the cell cores of the SRAM and DRAM are separated from each other to be capable of operating simultaneously, thereby reducing the latency due to page miss and improving the effective bandwidth. An enhanced synchronous DRAM (ESDRAM), cache DRAM (CDRAM), wide CDRAM and virtual channel memory (VCM) employ this method. There is described below the technical background and features of the VCM structure that has been known to have its performance improved when applied to the conventional structure for the purpose of securing a sufficient bandwidth for sequential accesses and reducing the time required for the row path through application of a new technique to obtain a short latency.
The VCM, proposed by NEC Co. of Japan, is constructed in a manner that an SRAM buffer is integrated in a DRAM, similar to the ESDRAM and CDRAM structures, to increase the effective bandwidth using the SRAM buffer. The integrated SRAM buffer is called “channel”, and the operation of the DRAM cell core and the operation of the SRAM buffer are separated from each other through this channel to conceal the pre-charge time due to page miss as in the ESDRAM and CDRAM. However, the VCM controls data transmission between the SRAM and DRAM through an external control while the ESDRAM and CDRAM perform a control for data transmission and consistency maintenance between the SRAM and DRAM cores using a controller integrated together with the SRAM in the DRAM. The VCM has relatively simple structure because a logic part for the

Affiliated with

Also associated with

Rating

Pipeline structure of memory for high-fast row-cycle does not yet have a rating. At this time, there are no reviews or comments for this patent.

If you have personal experience with Pipeline structure of memory for high-fast row-cycle, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Pipeline structure of memory for high-fast row-cycle will most certainly appreciate the feedback.

Rate now

Comments { 0 }

Profile ID: LFUS-PAI-O-3049681