Computer Memory vol 2

Created by Sie Lecken

Sense amplifier role
In DRAM, sense amplifiers detect the tiny voltage change on bitlines and amplify it to readable logic levels (0 or 1).

1/44

TermDefinition
Sense amplifier roleIn DRAM, sense amplifiers detect the tiny voltage change on bitlines and amplify it to readable logic levels (0 or 1).
Precharge phaseStep in DRAM where bitlines are reset to a neutral voltage before accessing a new row.
Activate commandTells DRAM to open a specific row by connecting it to the sense amplifiers via the wordline.
Read commandTells DRAM to output the data from the activated row to the memory bus.
Write commandTells DRAM to store new data into the currently activated row.
Row bufferTemporary storage inside a DRAM bank that holds data from the last activated row.
BankIndependent section of a DRAM chip that can operate in parallel with others for higher performance.
Bank interleavingTechnique of accessing multiple banks alternately to improve throughput and hide latency.
tRCD (Row to Column Delay)Time delay between activating a row and accessing a column within that row.
tRP (Row Precharge Time)Time required to close a currently open row before opening another.
tRAS (Row Active Time)Minimum time a row must remain open for reliable data operations.
CAS latency (CL)Number of clock cycles between sending a READ command and receiving the data.
Command queueIn memory controllers, the list of pending read/write operations waiting to be executed.
Memory schedulingOrder in which the memory controller issues commands to optimize performance and reduce conflicts.
Dual-channel memoryConfiguration where two memory channels operate simultaneously for doubled bandwidth.
Quad-channel memoryConfiguration using four channels for even greater bandwidth.
Memory timing exampleExample: DDR5-6000 CL30 means 6000 MT/s speed and 30 cycles CAS latency.
Burst lengthNumber of consecutive data words transferred per memory access.
PrefetchMechanism that fetches more data than immediately needed to improve efficiency.
Refresh intervalTypical DRAM cells must be refreshed every ~64 milliseconds to retain data.
Auto-refreshAutomatic process where the memory controller refreshes DRAM rows periodically.
Retention timeHow long a DRAM cell can hold charge before data begins to degrade.
Why DRAM needs refreshingBecause capacitors leak charge over time, even if power is maintained.
Power-down modeLow-power state of DRAM to save energy when idle.
Self-refresh modeMode where DRAM refreshes itself internally without CPU involvement.
On-die ECC (Error Correction Code)Feature in modern DDR5 that corrects small errors inside the chip before data leaves it.
External ECCUsed in servers to detect and correct multi-bit memory errors for high reliability.
Row hammer effectPhenomenon where repeatedly accessing one row causes electrical interference in adjacent rows, flipping bits.
Row hammer mitigationDRAM manufacturers add refresh logic or error correction to prevent this effect.
Burst transferHow DRAM sends multiple consecutive data words per request to increase efficiency.
Address mappingTranslation of CPU addresses into DRAM row, column, bank, and channel addresses.
SubarraySmaller section within a DRAM bank that improves internal parallelism.
Rank interleavingAlternating between ranks to improve memory throughput.
DRAM scaling challengeAs transistors and capacitors shrink, it becomes harder to store enough charge and prevent leakage.
Why DRAM is expensiveRequires precise manufacturing and high density of small capacitors.
Why SSDs are cheaper per GBUse NAND flash which packs more bits per cell and is non-volatile.
LPDDR (Low Power DDR)Mobile DRAM designed for phones and laptops; optimized for lower power use.
GDDR (Graphics DDR)High-bandwidth DRAM type used in GPUs for intensive data throughput.
HBM (High Bandwidth Memory)Stacked DRAM connected to GPU/CPU with very wide bus for massive bandwidth.
Memory hierarchy exampleCPU cache (nanoseconds) → DRAM (tens of ns) → SSD (microseconds) → HDD (milliseconds).
CPU–DRAM communicationThe CPU requests data via the memory controller over high-speed buses.
NUMA (Non-Uniform Memory Access)In multi-CPU systems, memory is divided so each CPU has faster access to its local memory.
Memory bottleneckThe situation where CPU performance is limited by slower memory access.
Future of DRAMResearch exploring new materials and 3D structures to increase density and reduce energy.