2.3. Memory Technologies

Unit 1 part 2: We ended part 1 of unit 1 by saying that we have different memory technologies which vary in terms of their access times and cost per GB. For example, we said that SRAMs are very fast and its speed is about one 0.5 to 2.5 nanoseconds, its access time; that means, it is on an average about one-tenth as fast as the processor ok. However, the cost per GB of this type of memories is also very huge. The cost per GB is about 2000 dollars to 5000 dollars.

This chunk introduces the concept of different memory technologies available in computer systems, primarily focusing on speed and cost. SRAM (Static Random-Access Memory) is highlighted as a very fast type of memory, with access times between 0.5 to 2.5 nanoseconds. This means it can quickly provide data to the processor, enabling high performance. However, the downside is the high cost, ranging from $2000 to$5000 per gigabyte, making it less viable for large capacities.

Think of SRAM like a luxury sports car—it's incredibly fast but also very expensive, making it suitable for specific high-performance applications where speed is crucial, but not practical for everyday use.

Then we have DRAMs which are about 150 to 100 times slower than SRAMs; that means, to bring a certain amount of data a data unit a word from DRAM the processor will require about hundreds of processor cycles to do so. The speed of a DRAM is typically in the range of 50 to 70 nanoseconds; that is the access time is in the range of 50 to 70 nanoseconds. But, it is also about hundred times cheaper than SDRAMs. So, the typical cost of DRAM units range in between 20 dollars to 75 dollars per GB.

This chunk discusses DRAM (Dynamic Random-Access Memory), which, while significantly slower than SRAM—by factors of 150 to 100—is more affordable. DRAM's access time is between 50 to 70 nanoseconds. This means retrieving data takes longer, but because DRAM costs between $20 to$75 per gigabyte, it's a more practical option for larger memory needs. It's often used as main system memory in computers.

Imagine DRAM as a family sedan. It's not as fast as the sports car, but it's much more economical and spacious, making it suitable for day-to-day driving—just like DRAM is suitable for everyday computing tasks.

Magnetic disks or hard disks are far more cheaper; about 1000 times cheaper than DRAMs being only about 0.2 to 2 dollars per GB. However, it is also about 1000 times slower than DRAM units. Its access times ranges in between 5 to 20 milliseconds. So, to bring a data word from the hard disk, the processor required tens of thousands of processor cycles.

This segment explains the cost-effectiveness and slower access of magnetic disks (hard drives). They are an order of magnitude cheaper (about $0.2 to$2 per GB) compared to DRAM. However, the trade-off is that accessing data can take 5 to 20 milliseconds, meaning the processor may have to wait significantly longer, sometimes requiring tens of thousands of cycles to retrieve data, which affects performance.

Think of magnetic disks as a public library. They hold a vast amount of information at a low cost, but finding the specific book you want can take time, especially during busy hours, illustrating the slower access relative to DRAM.

To achieve the best performance what would we desire? We would desire a very large capacity memory which can hold all our programs and data and which works at the pace of the processor. That means, if a processor requires a memory word in one cycle it is available in the processor from memory in the next cycle itself. However, in practice we saw the cost and performance parameters and it is difficult to achieve. So, to achieve the greatest performance memory should be able to keep pace with the processor.

This chunk emphasizes the importance of having a memory system that keeps up with the processor's speed to avoid performance bottlenecks. Ideally, the memory would not only be large enough to store all necessary data but would also deliver that data at the speed needed—for instance, quickly supplying memory words as the processor requests them. This ideal situation is challenging due to cost and technological limits.

Imagine a chef in a restaurant who needs ingredients quickly while preparing multiple dishes. If the pantry (memory) is well-stocked and organized (ideally fast), the chef works efficiently. But if it often takes time to retrieve items (slower memory), it slows down meal preparation (system performance).

Therefore, we have a design trade off. So, although the faster memories—SRAMs are very fast in terms of access time, they are also very costly. The solution is to have memory hierarchy where smaller more expensive and faster memories are supplemented by larger, cheaper and slower memories.

In this chunk, the necessity of a memory hierarchy is discussed as a solution to achieving balanced performance. The hierarchy consists of a combination of different types of memory—expensive and fast (like SRAM) for immediate access, supported by larger but slower types (like DRAM and magnetic disks) for bulk storage. This structure allows for optimizing costs without severely impacting performance.

Think of a Swiss Army knife which has different tools for varying tasks—some are quick to use (like the knife) but others take longer to set up (like the screwdriver). Using a combination effectively saves both time and effort, just as a memory hierarchy optimizes both speed and cost in computing.

--