W9D1 & W12D2 & W16D1: FanQijia

W12D2/menu.md

@@ -0,0 +1 @@
+ [范棋珈](FanQijia\11.30 Multi-Core & Static pipe. Intro..pdf) 

W16D1/menu.md

@@ -0,0 +1 @@
+[范棋珈](FanQijia/12.26 Review IV.pdf)

W9D1/FanQijia/W9D1-FanQijia-Mem_Tech.md

@@ -0,0 +1,89 @@
+# W9D1: Memory Technology
+
+> by Qijia Fan
+
+General structure:
+
+![img1](img1.jpg)
+
+The main memory is only controlled by CPU (or by I/O with DMA).
+
+## 1. SRAM
+
+> CAAQA5 P97
+
+- Static Random Access Memory
+- Implement by transistors (晶体管) and typically use 6 transistors to store 1 bit.
+- The structure is a **combinational logic circuit**, similar to the latch.
+- No need to refresh.
+- Usage: Cache
+
+## 2. DRAM
+
+> CAAQA5 P98
+
+- Dynamic Random Access Memory
+- Use one capacitor (电容) and transistor to store 1 bit.
+- **Need to be refreshed** periodically to prevent losing information (leakage current of capacitors).
+- Usage: Main Memory
+
+### Some techniques to improve proformance
+
+- FP-DRAM (Fast Page ...)
+
+- EDO-DRAM (Extended Data Output ...)
+
+- SDRAM (Synchronous  ...)
+
+  > CAAQA5 P100
+  >
+  > 2 **banks** on chip, a clock signal to DRAM, transfer synchronous (同步) to system clock
+
+- DDR DRAM (Double Data Rate  ...)
+
+  > CAAQA5 P100
+  >
+  > transfer data on both rising and falling edge 
+
+## 3. The organization of DRAM
+
+- Every bank of DRAM is a 2D matrix and consists of several **rows**.
+
+- Every DRAM access contains two stages (divide the address into 2 parts):
+
+  ![image-20221107184311803](img2.png)
+
+  1. RAS (Row Access Strobe, 行选通)
+  2. CAS (Column Access Strobe, 列选通)
+
+  > WE_L: write enable low
+  >
+  > OE_L: output enable low
+
+- **Multi banks (组):** 
+
+  ![image-20221107184311803](img3.jpg)
+
+  - Break the memory into 2-8 blocks (块) that can operate independently.
+  - Interleaved (交错) Memory
+
+## 4. The performance of the main memory
+
+- $AMAT_\text{mem} = T_\text{address} + T_\text{access} + T_\text{transfer and send data}$
+
+- Assume that we want to send 4 words and $T_\text{addr.}=1, T_\text{access} = 6,T_\text{trans} = 1$, then
+
+  - Simple Memory Performance: $=4\times(1+6+1)=32$
+  - Wide Memory Performance: $=1 + 6 + 1=8$
+  - Interleaved Memory Performance: $=1+6+4\times 1=11$
+
+  > Interleaved memory performs better than the others when the word count increase and $ T_\text{access}$ is close to $\text{BlockNumber}\times T_\text{trans}$.
+
+### Avoiding Bank Conflicts
+
+- S/W: loop interchange or declaring array not power of 2
+- H/W: Prime number of banks (in the following part?).
+
+### Fast Bank Number and CRT
+
+> will be introduced in the next class

W9D1/menu.md

@@ -0,0 +1 @@
+[范棋珈](FanQijia/W9D1-FanQijia-Mem_Tech.md)