Chip words
160 terms. The words of chip design and chip making, in plain language.
#
- 3D NAND
- NAND flash built upward in many stacked layers of cells, more than 200 in recent chips, instead of shrinking cells on a flat surface.
- 3D stacking
- Placing dies on top of each other in one package, connected vertically, to save space and shorten the distance data travels.
A
- Advanced packaging
- Packaging that joins several dies closely enough to work almost like one chip, side by side on an interposer (2.5D) or stacked (3D).
- AI accelerator
- A chip built to run neural networks fast and efficiently, such as Google's TPU, Nvidia's data centre GPUs or Amazon's Trainium.
- Analog and mixed-signal chip
- A chip that handles continuous real-world signals such as sound, voltage and radio, alone or together with digital logic.
- Arithmetic logic unit
- The part of a processor that does integer arithmetic (add, subtract) and logic (AND, OR, compare) on the values in its registers.
- Arm architecture
- The RISC instruction set first built by Acorn in 1985, licensed by Arm to hundreds of companies, and used in nearly every phone.
- ASIC
- A chip designed for one particular job, such as a Bitcoin miner, a video encoder or a company's own AI accelerator, rather than general use.
B
- Backside power delivery
- Moving a chip's power wiring underneath the transistors, to the back of the wafer, so the front is left free for signal wires.
- Ball grid array
- A package with a grid of tiny solder balls on its underside instead of pins, allowing hundreds or thousands of connections.
- Baseband processor
- The chip, or block of a chip, that handles a phone's radio communication: encoding, decoding and following the 4G or 5G rules.
- Binning
- Sorting chips after testing by how fast and how fully they work, then selling them as different models.
- Bipolar junction transistor
- The older transistor type, made of three doped layers, where a small base current controls a larger current. Still used in analog and RF chips.
- Branch prediction
- Guessing which way a program will go at an if-then decision, so the pipeline can keep working instead of waiting for the answer.
C
- Cache
- Small, fast memory on the chip that keeps copies of recently used data, so the processor rarely has to wait for slower main memory.
- Chemical mechanical planarisation
- Polishing the wafer flat between steps with a rotating pad and a chemical slurry, so each new layer starts on a level surface.
- Chip research conferences
- In chips, new circuits and transistors are usually first described in papers at a few conferences: ISSCC for circuits, IEDM and VLSI Symposium for devices.
- Chiplet
- A small die designed to be combined with others in one package, so a large processor is built from several pieces rather than one big die.
- CISC
- An instruction set with many rich, variable-length instructions, some doing several steps at once. x86 is the main example.
- Cleanroom
- A room where the air is filtered so that very few dust particles are present, because one speck can ruin a chip.
- Clock speed
- How many cycles per second a chip's internal clock ticks, from 740 kHz in the Intel 4004 to over 5 GHz in fast desktop chips.
- CMOS
- Complementary MOS: logic built from pairs of n-type and p-type transistors, so that almost no current flows except when a gate switches.
- Compute Express Link
- An open standard, first released in 2019, that runs over PCIe wiring and lets processors share memory with accelerators and memory expanders.
- CoWoS
- TSMC's Chip on Wafer on Substrate packaging, which puts processor dies and HBM stacks side by side on an interposer. Used by most AI GPUs.
- CPU
- The general-purpose processor that runs a computer's operating system and programs, one instruction after another.
D
- Dark silicon
- The part of a chip that must stay switched off at any moment because powering everything at once would exceed its power or heat limit.
- Data converter
- A circuit that turns an analog signal into numbers (ADC) or numbers back into an analog signal (DAC), linking chips to the physical world.
- DDR SDRAM
- Double data rate synchronous DRAM: the standard memory of PCs and servers, which moves data on both edges of each clock tick.
- Deep ultraviolet lithography
- Lithography with 248 nm or 193 nm ultraviolet light from excimer lasers. With water immersion and multiple patterning it reached 7 nm.
- Defect density
- The number of fatal flaws per unit of wafer area, often quoted per square centimetre. Together with die size it sets the yield.
- Dennard scaling
- The 1974 rule that as transistors shrink, their power density stays the same, so chips could get faster without getting hotter. It broke down around 2005.
- Deposition
- Adding thin films of metal, insulator or semiconductor to the wafer, sometimes just one layer of atoms at a time.
- Design for test
- Adding extra circuits to a chip so that manufacturing faults can be found quickly by test machines.
- Die
- One rectangle of silicon cut from a wafer, holding a complete circuit. It is what sits inside a chip package.
- Die shrink
- Remaking an existing chip design on a finer process node, so it comes out smaller, cheaper to make and often faster or cooler.
- Digital signal processor
- A processor tuned for maths on streams of signal data, such as audio, radio and images, with fast multiply-add hardware.
- Doping
- Adding a few atoms of another element to silicon so that it carries current by electrons (n-type) or by holes (p-type).
- DPU
- A data processing unit: a network card with its own processors that handles networking, storage and security work in data centres.
- DRAM
- Dynamic RAM: the main working memory of computers, storing each bit as charge in a tiny capacitor that must be refreshed many times a second.
- Dual in-line package
- A rectangular chip package with two parallel rows of pins that go through holes in a circuit board. The classic look of 1970s chips.
- Dynamic voltage and frequency scaling
- Raising or lowering a chip's voltage and clock speed on the fly to match the work, saving power when idle and boosting when busy.
E
- Electronic design automation
- The software used to design and check chips, from writing logic to laying out billions of transistors. Synopsys, Cadence and Siemens lead.
- EMIB and Foveros
- Intel's packaging methods: EMIB links dies side by side through a small silicon bridge in the substrate; Foveros stacks dies on top of each other.
- Etching
- Removing material from the wafer where the photoresist does not protect it, using chemicals or a plasma, to carve the chip's patterns.
- Extreme ultraviolet lithography
- Lithography using 13.5 nm light made by hitting tin droplets with a laser. ASML is the only supplier; it entered volume production in 2019.
F
- Fab
- A factory that makes chips on wafers. A leading-edge fab is among the most complex and expensive factories ever built.
- Fabless
- A chip company that designs chips but owns no fab, paying a foundry to make them. Nvidia, Qualcomm, AMD, MediaTek and Apple work this way.
- FinFET
- A transistor whose channel is a thin vertical fin of silicon wrapped by the gate on three sides, used from Intel's 22 nm in 2011.
- Flip chip
- Mounting a die face down so that small solder bumps across its surface connect directly to the package, instead of using wires.
- Floating-point unit
- Circuitry that does arithmetic on floating-point numbers, which store very large and very small values with a fixed number of significant digits.
- FLOPS
- Floating-point operations per second: the standard measure of computing speed for science and AI, quoted in teraflops, petaflops or exaflops.
- Foundry
- A company that manufactures chips designed by other companies. TSMC, founded in 1987, is the largest.
- FPGA
- A chip full of logic blocks and wiring that can be reconfigured after manufacture to become almost any digital circuit.
- Front end and back end of line
- The two halves of making a chip: the front end builds the transistors in the silicon, the back end builds the layers of wiring above them.
G
- Gate
- In a transistor, the electrode that switches it on and off. In logic design, a small circuit such as AND, OR or NOT built from transistors.
- Gate oxide and high-k metal gate
- The thin insulator under a transistor's gate. Silicon dioxide was replaced from 45 nm by hafnium-based high-k materials with metal gates.
- Gate-all-around transistor
- A transistor whose channel is a stack of thin horizontal sheets, each fully surrounded by the gate. The successor to the FinFET at 3 and 2 nm.
- GDDR
- Graphics DDR: DRAM built for graphics cards, trading power and capacity for very high data rates on each pin.
- GPU
- A processor with thousands of simple cores that work in parallel, built for graphics and now the main engine of AI training.
H
- Hardware description language
- A programming-like language for describing circuits, such as Verilog or VHDL, from which tools build the actual hardware.
- Heterogeneous cores
- Mixing fast, power-hungry cores with small, efficient ones on one chip, and moving work between them to save energy.
- High Bandwidth Memory
- Stacks of DRAM dies joined by through-silicon vias and placed beside a processor on an interposer, giving huge bandwidth for GPUs and AI chips.
- High-NA EUV
- The next generation of EUV scanner, whose optics gather light at a wider angle (0.55 numerical aperture) to print even finer lines.
- Hybrid bonding
- Joining two dies face to face so their copper pads fuse directly without solder, allowing connections only a few micrometres apart.
I
- Image sensor
- A chip covered in millions of light-sensitive pixels that turns an image into electrical signals: the eye of every digital camera and phone.
- Ingot
- A large single crystal of pure silicon, pulled slowly from molten silicon, that is sliced into wafers.
- Instruction pipeline
- Splitting the work of each instruction into stages, such as fetch, decode, execute and write back, so several instructions are in progress at once.
- Instruction set architecture
- The list of instructions a processor understands and how they are encoded. Software compiled for one ISA, such as x86 or Arm, runs on any chip that implements it.
- Instructions per cycle
- How many instructions a processor completes, on average, in each clock cycle. Performance is roughly IPC multiplied by clock speed.
- Integrated circuit
- A complete electronic circuit, with its transistors and wiring, made together on one piece of semiconductor. What people mean by "a chip".
- Integrated device manufacturer
- A chip company that both designs and manufactures its own chips, such as Intel, Samsung, Texas Instruments and Micron.
- Interconnect
- The stack of tiny metal wires above the transistors that connects them into circuits. A modern chip has more than a dozen layers.
- Interposer
- A thin slab of silicon or other material that sits under several dies and carries thousands of fine wires between them.
- Ion implantation
- Firing a beam of charged dopant atoms into the wafer to set exactly where silicon becomes n-type or p-type.
- IP core
- A ready-made, licensable design for one part of a chip, such as a CPU core, a memory controller or a USB interface.
L
- Leakage current
- Current that trickles through a transistor even when it is switched off, wasting power and producing heat while the chip is idle.
- Logic synthesis
- Software that turns a design written in Verilog or VHDL into a netlist: a list of real logic gates from a foundry's library and how they connect.
- Low-k dielectric
- An insulator with a lower dielectric constant than glass, placed between wires to cut the stray capacitance that slows signals and wastes power.
- LPDDR
- Low-power DDR: a version of DRAM designed for phones, tablets and thin laptops, using lower voltages and deep sleep states to save battery.
M
- Memory bandwidth
- How much data per second can move between a chip and its memory, measured in gigabytes or terabytes per second.
- Memory controller
- The circuit that manages reads and writes to DRAM. Once a separate chip, it now sits on the processor die.
- Memory management unit
- The hardware that translates the memory addresses programs use into real physical addresses, keeping programs separated and protected.
- Memory wall
- The growing gap between how fast processors compute and how fast memory can feed them data. Often the real limit on performance.
- MEMS
- Microelectromechanical systems: tiny moving parts such as springs, beams and membranes, carved from silicon with chip-making methods.
- Metrology and inspection
- Measuring and checking wafers during manufacture: line widths, layer thickness, alignment and defects, so problems are caught early.
- Microarchitecture
- How a processor is built inside to carry out its instruction set: its pipeline, caches, execution units and so on. Zen 4 and Golden Cove are both x86 microarchitectures.
- Microcode
- A layer of tiny internal instructions that some processors use to carry out their more complex instructions. It can often be patched after release.
- Microcontroller
- A small computer on one chip, with processor, memory and input-output built in, made to control a device such as a washing machine or a car part.
- Microprocessor
- A complete central processing unit on a single chip. The Intel 4004 of 1971 is generally counted as the first commercial one.
- Moore's law
- Gordon Moore's 1965 observation that the number of transistors on a chip doubles at a steady rate, later put at every two years.
- MOSFET
- The metal-oxide-semiconductor field-effect transistor: a voltage on the gate opens or closes a channel. The most manufactured device in history.
- Multi-core processor
- A chip with two or more complete processor cores, each able to run its own program. The standard design since the mid-2000s.
- Multi-project wafer
- A shared wafer run that carries designs from many customers at once, so each pays only for a small slice of the mask and wafer cost.
- Multiple patterning
- Printing one layer in two or more steps, or using deposited spacers to halve the spacing, so lines come out finer than the tool can print at once.
N
- NAND flash
- Non-volatile memory that keeps data with the power off, storing charge in cells strung together in series. It fills SSDs, phones and memory cards.
- NOR flash and EEPROM
- Non-volatile memories that can be read one byte at a time, used to hold the code of microcontrollers and the start-up firmware of devices.
- NPU
- A neural processing unit: a block inside a phone or laptop chip that runs AI tasks, such as photo processing or speech, using little power.
- Number formats
- How many bits a chip uses to store each number, from 64-bit floating point for science down to 8- and 4-bit formats for AI.
- NVLink
- Nvidia's high-speed link between its GPUs, and between GPUs and CPUs, much faster than PCIe, used to join many GPUs into one system.
O
- OSAT
- A company that packages and tests chips for others after the wafers leave the fab. ASE, Amkor and JCET are the largest.
- Out-of-order execution
- Letting a processor run instructions as soon as their inputs are ready, not strictly in program order, then putting the results back in order.
P
- p-n junction
- The boundary between p-type and n-type silicon, which lets current flow in one direction: the heart of diodes and transistors.
- Package substrate
- The small multilayer circuit board inside a package that carries signals and power from the die's bumps out to the balls or pins.
- Packaging
- The housing that protects a die and connects it to the circuit board, from simple plastic packages to stacks of many dies.
- PCI Express
- The standard high-speed link between a processor and graphics cards, SSDs and network cards. Each new version doubles the speed.
- Performance per watt
- How much useful work a chip does for each watt of power. In phones and data centres it often matters more than raw speed.
- Photolithography
- Printing a chip's patterns by shining light through a mask onto a light-sensitive coating on the wafer, layer by layer.
- Photomask
- A quartz plate carrying one layer of a chip's pattern in chrome or another absorber, used like a stencil in lithography.
- Photoresist
- A light-sensitive coating spread on a wafer. Light changes its chemistry so the exposed or unexposed parts can be washed away, leaving a pattern.
- Pitch
- The repeat distance between neighbouring lines on a chip, such as gate to gate or wire to wire. The real measure of how small a process prints.
- Place and route
- The step that decides where every cell sits on the die and draws the wires between them, producing the layout that becomes the masks.
- Planar transistor
- A flat transistor whose channel lies in the surface of the wafer, with the gate on top. The standard design until about 22 nm.
- Power management IC
- A chip that converts and distributes the right voltages to every part of a device, and manages battery charging and power saving.
- Process design kit
- The package a foundry gives designers for one process: transistor models, design rules, standard cells and the files that tie them to EDA tools.
- Process node
- A named generation of chip manufacturing, such as 7 nm or 3 nm. The number once measured a transistor feature but is now mostly a label.
Q
- Qubit
- The basic unit of a quantum computer: a bit that can be in a blend of 0 and 1 at once. Quantum processor chips hold dozens to hundreds of them.
R
- Radiation hardening
- Designing and making chips to survive radiation in space, aircraft and nuclear sites, where particles can flip bits or burn out circuits.
- Register
- A very small, very fast storage slot inside the processor that holds a value it is working on right now.
- Reticle limit
- The largest area a lithography scanner can print in one exposure, about 26 by 33 mm or roughly 858 mm². It caps the size of a single die.
- RISC
- A processor design idea from the 1980s: use a small set of simple, fixed-length instructions that can each run quickly in a pipeline.
- RISC-V
- An open, royalty-free RISC instruction set started at the University of California, Berkeley in 2010. Anyone can build a RISC-V chip.
- RTL
- Register-transfer level: describing a chip as registers that hold data and the logic that moves and changes it on each clock tick.
S
- Scanner and stepper
- The machine that exposes a wafer in lithography, moving it under the lens field by field. ASML, Nikon and Canon make them.
- Semiconductor
- A material such as silicon whose ability to carry current sits between a metal and an insulator, and can be controlled.
- SerDes
- A serialiser/deserialiser: the circuit that sends data at very high speed over a few wires, used in PCIe, Ethernet and chip-to-chip links.
- Shader core
- One of the many small processing units in a GPU. Nvidia calls them CUDA cores grouped in streaming multiprocessors; AMD calls the groups compute units.
- Silicon
- Element 14, the semiconductor nearly every chip is built on, refined to better than 99.9999999 per cent purity.
- Silicon on insulator
- Building transistors in a very thin layer of silicon that sits on a buried layer of oxide, which cuts leakage and stray capacitance.
- Silicon photonics
- Building optical parts, such as waveguides and modulators, on silicon chips so that data can travel as light instead of electrical signals.
- SIMD
- Single instruction, multiple data: one instruction that does the same operation on many numbers at once, used for graphics, media and AI.
- Simultaneous multithreading
- Letting one processor core run two (or more) programs' instructions at the same time, to keep its execution units busy.
- Speculative execution
- Running instructions before it is certain they are needed, and discarding the results if not. The source of the Spectre and Meltdown flaws.
- SRAM
- Static RAM: fast memory built from six transistors per bit that holds data as long as power is on, with no refresh. Used for caches and registers.
- Standard cell
- A ready-made building block, such as a logic gate or flip-flop, drawn to a fixed height so tools can line thousands of them up in rows.
- Stepping
- A revision of a chip after it enters production, to fix bugs or improve yield, named with codes such as A0, B1 or C0.
- Superscalar
- A processor that can start more than one instruction in the same clock cycle, using several execution units side by side.
- System on a chip
- A chip that holds most of a computer: CPU cores, graphics, memory controllers, an AI unit, a modem and more, all on one die.
- Systolic array
- A grid of simple multiply-add cells that pass data to their neighbours in a steady rhythm, ideal for matrix multiplication. The core of Google's TPU.
T
- Tape-out
- The moment a finished chip design is sent to the foundry to make masks. Named after the magnetic tapes once used to carry the files.
- TDP
- Thermal design power: roughly how much heat, in watts, a chip's cooling must remove under sustained load. Not always the same as its actual draw.
- Tensor core
- A unit inside a GPU or CPU that multiplies small matrices in one step, greatly speeding up neural network maths. Nvidia introduced them in 2017.
- Threshold voltage
- The gate voltage at which a transistor starts to conduct. Lower is faster but leaks more; higher leaks less but is slower.
- Through-silicon via
- A vertical copper connection that passes right through a silicon die, letting stacked dies talk to each other directly.
- Tick-tock
- Intel's plan from 2007 to 2016 of alternating a new process node (tick) with a new microarchitecture (tock), roughly one each year.
- Timing closure
- Making sure every signal on a chip arrives within its clock cycle, under all conditions, before the design can be sent to the fab.
- TOPS
- Tera operations per second: trillions of operations a second, the usual measure of AI chip and NPU speed, normally for 8-bit integer maths.
- Transistor
- A tiny electrically controlled switch or amplifier. Modern chips hold billions of them, wired together to do logic and store data.
- Transistor count
- How many transistors a chip contains: about 2,300 in the Intel 4004 of 1971, and over 100 billion in the largest chips of the 2020s.
- Transistor density
- How many transistors fit in a square millimetre of silicon, usually quoted in millions per mm². A fairer way to compare process nodes than their names.
U
- UCIe
- An open standard, launched in 2022, for the links between chiplets in one package, so dies from different makers can work together.
V
- Verification
- Proving a chip design does what it should before it is made. It often takes more engineering effort than writing the design itself.
- VLIW
- A design where each instruction bundles several operations that run in parallel, leaving the compiler, not the hardware, to schedule them.
W
- Wafer
- A thin, polished disc of crystal silicon on which hundreds of chips are made at once. The standard size since the 2000s is 300 mm across.
- Wafer test
- Testing every die on the wafer with fine probes before it is cut up, so that only working dies are packaged.
- Wafer-scale integration
- Using a whole wafer as one giant chip instead of cutting it into dies, as Cerebras does for its AI processors.
- Wide-bandgap semiconductors
- Materials such as gallium nitride and silicon carbide that handle higher voltages, temperatures and frequencies than silicon. Used in chargers and electric cars.
- Wire bonding
- Connecting a die to its package with fine gold, copper or aluminium wires, each thinner than a human hair, welded at both ends.
- Word size
- How many bits a processor naturally handles at once in its registers and arithmetic, from 4 bits in the 4004 to 64 bits today.
X
- x86
- The instruction set that began with Intel's 8086 in 1978 and still runs most PCs and servers, extended to 64 bits by AMD in 2003.
Y
- Yield
- The share of chips on a wafer that work. High yield is what makes a process profitable, and it falls as dies get bigger.