The CPU executes instructions, RAM holds the data and code currently in use, and the operating system decides who gets access to both. Every action on a computer is some version of this three-way conversation.
Fig. 01 — The CPU, RAM and OS work together
Swipe sideways to see the whole diagram →
Step by step
You launch a program
The OS finds the executable on storage and creates a process for it.
Code is loaded into RAM
The CPU cannot execute directly from disk, so instructions and data are copied into memory first.
Fetch
The CPU reads the next instruction from memory, using the program counter to know where it is.
Decode
The control unit works out what the instruction means and which units are needed.
Execute
The ALU performs the calculation, or data is moved between registers and memory.
Cache does the heavy lifting
Recently used data is kept in L1, L2 and L3 cache, because reaching RAM every time would leave the CPU waiting for most of its cycles.
The OS arbitrates
It schedules processes onto cores, maps virtual addresses to physical RAM, and swaps out pages when memory runs short.
Real-world example
Closing a document without saving loses it because the edits lived only in RAM, which is volatile and clears on power loss. Saving writes them to an SSD, which keeps data without power. That single distinction explains most 'I lost my work' moments.
Where people get this wrong
RAM is thought of as where programs run and storage as where they sit. The detail that matters is that RAM is volatile — the reason unsaved work vanishes is not a bug, it is the physics of the component.
Common interview questions
What is the fetch-decode-execute cycle?
The basic loop of a CPU: read an instruction, work out what it means, carry it out, then move to the next.
Why is cache memory needed if we already have RAM?
RAM is roughly a hundred times slower than the CPU. Cache holds frequently used data close to the core so the CPU is not left idle.
What is the difference between volatile and non-volatile memory?
Volatile memory such as RAM loses its contents without power. Non-volatile storage such as an SSD retains them.
What is virtual memory?
An abstraction that gives each process its own address space, with rarely used pages moved out to disk so programs can exceed physical RAM.
Inside the Druvexaa simulation
The simulation follows a single instruction from storage into RAM, into cache, into a register and through the ALU, with the timing cost of each step shown to scale.
Manufacturers count storage in decimal gigabytes (1 GB = 1,000,000,000 bytes) while operating systems display binary gibibytes (1 GiB = 1,073,741,824 bytes). The hardware is not lying — the two sides are using different units.
Fig. 02 — Why does a 32 GB pen drive show only 29.5 GB
Swipe sideways to see the whole diagram →
Step by step
Two different units
Storage makers use powers of ten; operating systems use powers of two. The gap is about 7% at gigabyte scale and grows with size.
Do the arithmetic
32,000,000,000 bytes divided by 1,073,741,824 gives roughly 29.8 GiB. That accounts for almost the entire difference.
Formatting overhead
Creating a filesystem consumes space for the boot sector, allocation tables and journal.
Reserved blocks
Flash drives and SSDs hold back spare blocks to replace worn cells. This over-provisioning is invisible to you but real.
Cluster size
Files occupy whole clusters. Thousands of small files leave slack space that is allocated but unused.
The gap grows with capacity
A 1 TB drive shows around 931 GiB. Same maths, larger absolute difference.
Real-world example
A 1 TB SSD showing 931 GB in Windows is behaving exactly as designed. macOS reports it closer to 1 TB because it displays decimal units, matching the label — same drive, different display convention.
Where people get this wrong
This is read as a marketing trick. It is a unit mismatch, and the same drive shows a different number on macOS than on Windows for exactly that reason. Nobody is taking your storage.
Common interview questions
What is the difference between GB and GiB?
A GB is 10⁹ bytes; a GiB is 2³⁰ bytes, about 7.4% larger.
Is a manufacturer overstating capacity?
No. They use the SI decimal definition consistently. The mismatch is in how the operating system labels its binary units.
Why does a formatted drive show slightly less than the pure conversion?
Filesystem metadata — allocation tables, journals and reserved sectors — occupies real space.
What is over-provisioning in SSDs?
Reserved flash capacity used for wear levelling and to replace failed blocks, which extends the drive's life.
Inside the Druvexaa simulation
The simulation counts out the actual bytes on the drive and divides them both ways in front of you, then subtracts the filesystem structures so the final number matches what your file manager reports.
A pen drive's real speed depends on the USB standard, the flash memory inside it, and the controller managing that flash. The port's rated speed is a ceiling, not a promise.
Fig. 03 — Why do USB drives transfer at different speeds
Swipe sideways to see the whole diagram →
Step by step
The port sets the ceiling
A USB 3.0 drive in a USB 2.0 port is capped at USB 2.0 speeds. The slowest link in the chain wins.
Flash type matters
Cheap TLC or QLC flash writes far more slowly than the MLC used in higher-end drives, especially once any fast cache is exhausted.
The controller matters
A better controller handles more parallel channels and manages wear more efficiently, which raises sustained speed considerably.
Reads beat writes
Writing to flash requires erasing blocks first, so write speeds are typically a fraction of read speeds.
File size matters
One 4 GB file copies far faster than 40,000 small files, because each file carries filesystem overhead.
Sustained speed drops
Many drives write fast into a small cache and then slow down sharply once it fills — this is why a large copy starts quick and then crawls.
Real-world example
A large video copies to a pen drive at 90 MB/s for the first few seconds, then falls to 12 MB/s. The fast cache filled up and you are now seeing the true speed of the underlying flash.
Where people get this wrong
The number printed on the packaging is read as the speed you will get. It is the signalling rate of the interface. The flash inside is usually the real limit, which is why two drives on the same standard can differ by a factor of ten.
Common interview questions
Why is real USB speed always below the rated speed?
The rating is raw signalling in bits per second. Protocol overhead, encoding and the drive's own flash limits reduce usable throughput considerably.
Does a USB 3.0 drive work in a USB 2.0 port?
Yes, USB is backwards compatible, but it runs at the older port's speed.
Why are writes slower than reads on flash?
Flash must erase a whole block before rewriting it, and erase operations are much slower than reads.
Why do many small files copy so slowly?
Each file needs its own metadata operations, so overhead dominates and the drive never reaches its sequential speed.
Inside the Druvexaa simulation
Copy a large file in the simulation and watch the write cache fill and the transfer rate collapse to the true sustained speed of the flash behind it.
Analog-to-digital conversion turns a continuous real-world signal, such as sound pressure on a microphone, into numbers a computer can store and process. It works by measuring the signal at regular intervals and rounding each measurement to a fixed scale.
Fig. 04 — Analog-to-Digital Conversion
Swipe sideways to see the whole diagram →
Step by step
The signal arrives
A microphone produces a continuously varying voltage that mirrors the sound pressure hitting it.
Sampling
The ADC measures that voltage thousands of times per second. CD audio uses 44,100 samples per second.
The Nyquist rule
To capture a frequency accurately you must sample at more than twice that frequency. Human hearing tops out near 20 kHz, which is why 44.1 kHz was chosen.
Quantisation
Each sample is rounded to the nearest available level. 16-bit depth gives 65,536 levels; 24-bit gives over 16 million.
Quantisation error
The tiny rounding difference is heard as noise. More bits means finer steps and less audible noise.
Encoding
The levels are written as binary numbers, ready for storage, compression or transmission.
Real-world example
Your phone's voice notes app runs an ADC continuously while recording. Playback runs the reverse — a DAC turns those numbers back into a voltage that moves the speaker cone.
Where people get this wrong
Sampling is pictured as recording. It is measuring — the continuous signal between two samples is gone forever, and everything you hear on playback is reconstructed from the points that were kept.
Common interview questions
What is the Nyquist theorem?
To reconstruct a signal faithfully you must sample at more than twice its highest frequency component.
What is the difference between sample rate and bit depth?
Sample rate is how often you measure, which sets frequency range. Bit depth is how precisely you measure, which sets dynamic range and noise floor.
What is aliasing?
Frequencies above half the sample rate folding back and appearing as false lower frequencies. An anti-aliasing filter removes them before sampling.
Why is 44.1 kHz the standard for CD audio?
It is just over twice the roughly 20 kHz limit of human hearing, leaving headroom for the anti-aliasing filter.
Inside the Druvexaa simulation
Speak into the simulation and watch your own voice being sampled live, then lower the sample rate and bit depth and hear the reconstruction degrade in real time.
Cache is a small, very fast memory that sits between the CPU and RAM, holding data the processor is likely to need next. It exists because RAM is roughly a hundred times slower than the CPU that depends on it.
Fig. 05 — Cache Memory
Swipe sideways to see the whole diagram →
Step by step
The speed gap
A modern core executes several instructions per nanosecond. Waiting 80 ns for RAM wastes hundreds of cycles.
Multiple levels
L1 is tiny and immediate. L2 is larger and slightly slower. L3 is shared between cores and larger still.
Cache hit and miss
A hit serves data immediately. A miss fetches from the next level down and costs the difference in latency.
Temporal locality
Recently used data is likely to be used again soon, so it is worth keeping close.
Spatial locality
Data near recently used data is likely to be needed, so caches fetch a whole line — typically 64 bytes — not a single value.
Write policies
Write-through updates RAM immediately and is simpler. Write-back updates RAM later and is faster but must track modified lines.
Real-world example
Looping over an array in order is far faster than jumping randomly through it, even with identical instruction counts. Sequential access uses every value in each fetched cache line; random access wastes most of every fetch.
Where people get this wrong
Cache is treated as automatic and therefore not the programmer's concern. Access pattern is often the single largest performance factor in data-heavy code — the same algorithm can run several times faster purely by traversing memory in the order it is laid out. Cache size on a spec sheet matters far less than whether the code respects locality.
Common interview questions
What is the difference between temporal and spatial locality?
Temporal means recently accessed data is likely to be accessed again. Spatial means data near it is likely to be accessed next.
Why is L1 cache so small?
Speed requires physical proximity to the core and simple lookup logic. Making it larger would make it slower, defeating its purpose.
What is a cache line?
The fixed block, usually 64 bytes, transferred between cache levels. Fetching whole lines is what exploits spatial locality.
What is the difference between write-through and write-back?
Write-through updates main memory on every write. Write-back defers the update, which is faster but requires tracking dirty lines.
Inside the Druvexaa simulation
The CPU and memory simulation shows a request being checked against L1, then L2, then L3, then RAM, with the cost of each level drawn to scale so a cache miss is visibly expensive rather than an abstract statistic.
An HDD stores data magnetically on spinning platters read by a moving mechanical head. An SSD stores data in flash memory chips with no moving parts, which is why it is faster, quieter and more resistant to physical shock.
Fig. 06 — Difference between an SSD and an HDD
Swipe sideways to see the whole diagram →
Step by step
An HDD physically moves
The head must travel to the right track and wait for the platter to rotate the data underneath it. This seek time dominates random access.
An SSD addresses electrically
Any block is reachable in roughly the same time, so random reads cost almost the same as sequential ones.
Interface matters
A SATA SSD is capped near 550 MB/s by the connection itself. An NVMe SSD over PCIe removes that ceiling entirely.
Flash must erase before writing
Cells are erased in large blocks, which is why sustained write speed is lower than read speed and why controllers work hard to manage it.
Wear levelling
Each cell tolerates a limited number of write cycles, so the controller spreads writes evenly across the drive to extend its life.
TRIM
The operating system tells the drive which blocks are no longer in use, letting the controller erase them in advance instead of during a write.
Real-world example
Booting from an SSD is dramatically faster than from an HDD not because the total data is large, but because startup reads thousands of small files scattered across the drive — exactly the pattern where mechanical seek time destroys an HDD.
Where people get this wrong
The two are compared on sequential transfer speed, which is the least representative measurement. Everyday responsiveness is dominated by small random reads, where an SSD is not twice as fast but roughly a hundred times faster. A benchmark showing an HDD at half the sequential speed of a SATA SSD badly understates the difference you actually feel.
Common interview questions
Why is an SSD so much faster for random access?
There is no mechanical movement. Any address is reachable electrically in roughly constant time, while an HDD must physically seek.
What is wear levelling?
Distributing writes evenly across flash cells so no single region wears out early, extending the usable life of the drive.
What does TRIM do?
It informs the SSD which blocks the filesystem no longer needs, allowing the controller to erase them ahead of time and keep write performance high.
Why is NVMe faster than SATA?
NVMe connects over PCIe with a protocol designed for flash and deep parallel queues, removing the bandwidth and command-queue limits SATA inherited from mechanical drives.
Inside the Druvexaa simulation
The storage simulation runs the same workload against a modelled platter drive and a flash drive, showing the mechanical head physically travelling for each random read while the flash drive answers immediately — the gap made visible rather than quoted.
A core is an independent processing unit. A thread is a sequence of instructions a core executes. Clock speed is how many cycles a core runs per second — and none of the three tells you performance on its own.
Fig. 07 — CPU Cores, Threads and Clock Speed
Swipe sideways to see the whole diagram →
Step by step
Cores do real parallel work
Four cores can genuinely execute four things at once, provided the software is written to use them.
Threads share a core
Simultaneous multithreading lets one core keep working on a second thread while the first waits on memory. It adds perhaps 15 to 30 percent, not another full core.
Clock speed is cycles, not work
5 GHz means five billion cycles per second. How much is achieved per cycle depends entirely on the design.
IPC is the missing number
Instructions per cycle varies hugely between architectures. A newer 3.5 GHz chip routinely beats an older 4.5 GHz one.
Turbo and thermals
Advertised peak clocks apply to one or two cores for short periods. Sustained all-core speed is lower and depends on cooling.
Match the chip to the work
Games and most everyday software depend on single-thread speed. Rendering, compiling and video encoding scale with core count.
Real-world example
A 16-core workstation can feel slower than an 8-core laptop for everyday use. Most desktop software cannot spread across sixteen cores, so what you actually experience is the speed of one core.
Where people get this wrong
Cores and gigahertz are compared across generations as if they were the same unit. They are not — architecture changes how much each cycle accomplishes, so specification numbers only compare meaningfully within the same family. Benchmarks for the work you actually do are the only reliable comparison.
Common interview questions
What is the difference between a core and a thread?
A core is physical hardware that executes instructions. A thread is a sequence of instructions; multithreading lets one core interleave two.
Why can a lower-clocked CPU outperform a higher-clocked one?
Higher instructions per cycle. A newer architecture accomplishes more work in each cycle, so raw frequency does not decide the outcome.
Does doubling cores double performance?
Only for work that parallelises well. Amdahl's law means the sequential portion limits the overall gain.
Inside the Druvexaa simulation
The CPU, RAM and OS simulation follows instructions through the fetch, decode and execute cycle, making it clear why a cycle and a completed instruction are not the same thing.
These topics are written to the same depth. A dedicated Druvexaa simulation is planned for each — until then, every one links to the closest existing simulation and quiz.
A GPU contains thousands of simple cores designed to perform the same operation across huge amounts of data at once. A CPU has a few complex cores designed to handle varied, branching work quickly.
Fig. 08 — GPU and how is it different from a CPU
Swipe sideways to see the whole diagram →
Step by step
Different design goals
A CPU minimises the time for any one task. A GPU maximises total work completed across many identical tasks.
SIMD execution
The same instruction runs across many data elements simultaneously, which suits pixels, vertices and matrix elements perfectly.
Memory bandwidth
GPUs use very wide, fast memory because thousands of cores starve instantly without it.
Why AI moved to GPUs
Neural network training is dominated by matrix multiplication, which is exactly the workload GPUs were already built for.
Where GPUs are poor
Branching, sequential logic and tasks with unpredictable control flow. The operating system will never run on one.
VRAM is the hard limit
A model or scene that does not fit in the GPU's memory either fails or falls back to far slower system memory.
Real-world example
Rendering a frame means computing millions of pixels with the same shader. A CPU would work through them largely in sequence; a GPU processes thousands simultaneously, which is the whole reason real-time graphics is possible.
Where people get this wrong
A GPU is assumed to be simply a faster processor. It is faster only for work that is wide and uniform. Give it branching, sequential logic and it is dramatically slower than a CPU — which is why the two coexist rather than one replacing the other.
Common interview questions
Why are GPUs used for machine learning?
Training is dominated by large matrix operations that parallelise across thousands of cores, which is precisely the GPU execution model.
What is the difference between latency and throughput optimisation?
A CPU minimises time for a single task. A GPU maximises total completed work, accepting higher latency per individual item.
Why does VRAM capacity matter so much?
The working set must fit in GPU memory. Exceeding it forces slow transfers over the system bus or prevents the workload running at all.
Inside the Druvexaa simulation
No dedicated simulation yet. The CPU, RAM and OS simulation shows sequential instruction execution, which is the contrast that makes the GPU's parallel model easy to understand.
RAM comes in two fundamental kinds. DRAM stores each bit in a capacitor that must be refreshed constantly — it is cheap and dense, and it is what your memory sticks are. SRAM holds each bit in a latch, is much faster, and is used for CPU cache.
Fig. 09 — Types of RAM (DDR, SRAM, DRAM)
Swipe sideways to see the whole diagram →
Step by step
DRAM needs refreshing
Each capacitor leaks charge and must be rewritten thousands of times per second. That refresh cycle is part of why DRAM is slower.
SRAM holds its state
A latch keeps the value as long as power is present, so there is no refresh delay — but it needs roughly six transistors per bit.
DDR means double data rate
Data transfers on both the rising and falling edge of the clock, doubling throughput without doubling clock speed.
Generations raise bandwidth
DDR5 offers substantially more bandwidth than DDR4 and is not physically or electrically compatible with it.
Timings and latency
CAS latency measures cycles of delay before data appears. Higher-frequency memory often has higher timings, so the two partly cancel out.
Dual channel
Two matched modules give the memory controller two paths at once. Two 8 GB sticks noticeably outperform one 16 GB stick.
Real-world example
Adding a second matching memory module often improves performance more than upgrading to a faster single module, because dual channel doubles the available paths to memory rather than just raising the clock.
Where people get this wrong
Memory is judged by capacity and frequency alone. Once you have enough capacity for your workload, more brings no benefit at all — an idle 32 GB in a machine that never exceeds 12 GB does nothing. Channel configuration and having enough for the actual working set matter far more than the number on the box.
Common interview questions
What is the difference between SRAM and DRAM?
SRAM uses latches, needs no refresh and is fast but expensive, so it is used for cache. DRAM uses capacitors, requires refresh, and is dense and cheap enough for main memory.
What does the double in DDR refer to?
Data is transferred on both the rising and falling edges of the clock signal, doubling transfer rate for a given clock.
Why is dual channel memory faster?
It gives the memory controller two independent paths, roughly doubling bandwidth compared with a single module.
Inside the Druvexaa simulation
No dedicated simulation yet. The CPU, RAM and OS simulation shows the latency gap between cache and main memory to scale, which is the practical difference between SRAM and DRAM.
The motherboard is the board every other component connects to. It carries the CPU socket, memory slots, expansion slots and the chipset that routes data between all of them.
Fig. 10 — Motherboard
Swipe sideways to see the whole diagram →
Step by step
The socket sets the CPU
Each socket type accepts one family of processors. This is the single largest constraint on any upgrade.
The chipset sets the features
It determines how many PCIe lanes, USB ports and storage devices are available, and whether overclocking is permitted.
PCIe lanes are finite
A graphics card, an NVMe drive and a capture card all draw from a limited pool. Populating certain slots disables others.
Form factor sets the size
ATX, Micro-ATX and Mini-ITX trade expansion slots for physical size and must match the case.
Power delivery matters
The VRM converts and regulates power to the CPU. Weak power delivery causes a high-end chip to throttle under sustained load.
Firmware ties it together
UEFI initialises everything at power-on. A firmware update is often what allows a newer CPU to work in an existing board.
Real-world example
Buying a faster processor for an existing machine frequently fails not because the CPU is wrong but because the socket differs — which turns a single-part upgrade into a motherboard, memory and CPU replacement.
Where people get this wrong
The motherboard is treated as a passive tray that parts plug into. It decides which processors you can ever fit, how many fast storage devices you can run at once, and whether a demanding CPU can sustain its clocks. It is the component that quietly determines every future upgrade you are allowed to make.
Common interview questions
What does the chipset determine?
Available PCIe lanes, USB and storage connectivity, supported memory speeds, and features such as overclocking support.
Why can't every CPU fit every motherboard?
The socket is physically and electrically specific to a processor family, and the chipset and firmware must support that model.
What is a VRM and why does it matter?
The voltage regulator module supplying clean, stable power to the CPU. Inadequate VRMs cause throttling under sustained load.
Inside the Druvexaa simulation
No dedicated simulation yet. The system startup simulation covers the firmware stage that lives on this board and initialises every component described here.