Skip to content

Glossary and Further Reading

Glossary

ACK / NACK
In I²C, the acknowledgement (or its absence) that a receiver signals after each byte by pulling SDA low (ACK) or leaving it high (NACK).
AMAT (average memory access time)
Hit time plus miss rate times miss penalty; with several levels, the miss penalty of one level is the AMAT of the next.
Amdahl's law
If a fraction \(p\) of the work is sped up by a factor \(s\), the overall speed-up is \(1/((1-p) + p/s)\), which can never exceed \(1/(1-p)\).
AXI
Arm's Advanced eXtensible Interface: five independent VALID/READY channels (AW, W, B, AR, R) with bursts and outstanding, reorderable transactions.
Back-pressure
A receiver slowing a sender by keeping READY low; the sender holds its data and waits.
Baud rate
The number of bits per second on a serial line; both ends of a UART must be set to the same value.
BOOM
The Berkeley Out-of-Order Machine: a parameterisable superscalar out-of-order RISC-V core written in Chisel.
Branch target buffer (BTB)
A small table, indexed by a branch's address, that remembers where the branch jumped last time, so fetch can redirect immediately.
Bubble
An empty slot that travels down a pipeline in place of an instruction that was stalled.
Burst
A run of data beats described by a single address beat.
Cache
A small, fast memory that keeps copies of recently used data close to the processor.
Cache line (block)
The unit of data a cache stores and moves; typically 64 bytes.
Chipyard
Berkeley's framework that composes Rocket, BOOM, caches, TileLink and peripherals into complete SoCs.
Common data bus (CDB)
In Tomasulo's algorithm, the bus on which a finished unit broadcasts (tag, value) so that all waiting instructions can capture it.
Constant-time code
Code whose running time, memory addresses and control flow do not depend on secret data.
Coherence
The guarantee that all caches agree about the value of each individual memory location.
Consistency model
The rules for the order in which accesses to different locations may be observed by other cores.
CPI / IPC
Cycles per instruction, and its reciprocal, instructions per cycle.
CPOL / CPHA
The two SPI settings that fix the idle level of the clock and the edge on which data is sampled.
Crossbar
A full switch connecting every manager to every subordinate, so that non-conflicting transfers proceed in parallel.
Dennard scaling
The long-lived trend in which smaller transistors also needed less voltage, keeping power per unit area constant; it ended around 2005.
DMA (direct memory access)
A hardware engine that copies blocks of data between memory and devices without CPU involvement for each word.
Directory
A record, per cache block, of which cores hold copies; used for coherence at larger scale.
DLP / ILP / TLP
Data-level, instruction-level and thread-level parallelism.
Flit
A flow-control unit: the fixed-size piece of a packet that moves through a network-on-chip.
Forwarding (bypassing)
Wires that carry a result straight from a pipeline register into a later instruction's ALU input, skipping the register file.
False sharing
Slowdown caused by unrelated variables that happen to share a cache block, so that writes by different cores invalidate each other's copies.
Fence (memory barrier)
An instruction that forces earlier memory operations to complete before later ones begin.
GPIO
A general-purpose input/output pin that software can drive or read.
Hazard
Anything that stops the next instruction from proceeding in its planned cycle: structural, data or control.
Hysteresis
Resistance to change: in a 2-bit branch predictor, the need for two successive surprises before the prediction flips.
I²C
A two-wire (SDA and SCL), open-drain, multi-device serial bus with addresses and per-byte acknowledgement.
Instruction window
The pool of decoded instructions from which an out-of-order core may choose what to execute.
Interconnect
The fabric of buses, crossbars or networks that lets the blocks of a chip communicate.
ISA (instruction set architecture)
The contract between software and hardware: registers, instructions, memory model, exceptions.
Iron Law
CPU time = instruction count × CPI × clock period.
Latency / throughput
The time for one operation to complete, and the number of operations completed per unit time.
Load-use hazard
A load followed immediately by an instruction that uses the loaded value; costs one stall even with forwarding.
Load/store queue (LSQ)
The structure that holds in-flight memory operations in order, performing store-to-load forwarding and memory disambiguation.
Manager / subordinate
The block that starts a bus transaction, and the block that answers it (formerly master and slave).
Memory-mapped I/O (MMIO)
Device registers placed in the address space so that ordinary loads and stores reach them.
Memory wall
The growing gap between processor speed and DRAM latency.
MESI
The four-state invalidation coherence protocol: Modified, Exclusive, Shared, Invalid.
Microarchitecture
The implementation of an ISA: pipeline, caches, predictors, execution units.
MIPS
Millions of instructions per second; a misleading performance metric.
MSHR
Miss status handling register: tracks one outstanding cache miss in a non-blocking cache.
Network-on-chip (NoC)
A packet-switched network of routers and links inside a chip.
NUMA
Non-uniform memory access: memory whose access time depends on which node owns it.
Open-drain
An output that can only pull a line low; a pull-up resistor supplies the high level.
Out-of-order execution
Starting instructions in any order that respects their dependences, while committing them in program order.
Pipeline
A datapath divided into stages so that several instructions are in progress at once.
Precise exception
An exception for which every earlier instruction has completed and no later instruction has changed any state.
Prefetching
Fetching data into the cache before it is demanded.
RAW / WAR / WAW
Read after write (a true dependence), write after read and write after write (false, name dependences).
Register renaming
Mapping architectural registers onto a larger physical register file so that every write gets a fresh register.
Reorder buffer (ROB)
A queue of all in-flight instructions in program order; instructions commit from its head.
Reservation station
A buffer in front of a functional unit in which an instruction waits for its operands.
Roofline model
A plot of attainable performance against arithmetic intensity, bounded by a bandwidth roof and a compute roof.
RV64GC
The RISC-V base 64-bit integer ISA with the M, A, F, D and C extensions: the usual baseline for a Linux-class core.
Scoreboard
The central bookkeeping structure of the CDC 6600 that tracked units and registers for out-of-order execution.
Side channel
A leak of secret information through unintended signals such as time, power or cache state.
SMT (simultaneous multithreading)
Issuing instructions from several threads into the same execution units in the same cycle.
SPI
Serial peripheral interface: a four-wire, clocked, full-duplex bus with a chip-select line per device.
Speculation
Executing instructions down a predicted path before the branch or fault that governs them is resolved.
Superscalar
A core that fetches, issues and commits several instructions per cycle.
TileLink
An open interconnect protocol from SiFive that scales from simple registers to full cache coherence.
TLB (translation lookaside buffer)
A small cache of recent virtual-to-physical page translations.
Tomasulo's algorithm
Dynamic scheduling with reservation stations, tag-based renaming and a common data bus.
Transient execution
Execution of wrong-path instructions that is later squashed; its architectural effects are undone but its cache footprint remains.
UART
Universal asynchronous receiver/transmitter: a two-wire serial link with no shared clock.
VALID/READY
The handshake in which a transfer happens only when the sender's VALID and the receiver's READY are both high.
VIPT
Virtually indexed, physically tagged cache: the index comes from untranslated address bits, the tag from the translated address.
VLIW
Very long instruction word: an architecture in which the compiler packs independent operations into one wide instruction.

Further reading

The following are standard references for the subjects of this book, and the places where the tools it mentions are documented. Check publication details and current versions before citing.

Textbooks

  • John L. Hennessy and David A. Patterson, Computer Architecture: A Quantitative Approach. The standard graduate text on pipelining, instruction-level parallelism, memory hierarchy and multiprocessors.
  • David A. Patterson and John L. Hennessy, Computer Organization and Design: The Hardware/Software Interface, RISC-V edition. A gentler introduction that builds a RISC-V processor, pipeline and memory system.

Original sources named in the text

  • R. M. Tomasulo, "An efficient algorithm for exploiting multiple arithmetic units", IBM Journal of Research and Development, 1967, the paper behind reservation stations and the common data bus.
  • J. E. Smith and A. R. Pleszkun, work on implementing precise interrupts in pipelined processors (1988), the origin of the reorder buffer.
  • The Spectre and Meltdown papers of 2018, by their respective research teams, which introduced the transient-execution attacks of Chapter 10.

Specifications and projects

  • The RISC-V unprivileged and privileged specifications, published by RISC-V International.
  • Arm's AMBA specifications (APB, AHB, AXI, CHI) and the TileLink specification from SiFive.
  • The open-source projects of Chapters 8 and 11: Ripes (github.com/mortbopet/Ripes), gem5 (github.com/gem5/gem5), Chipyard (github.com/ucb-bar/chipyard), rocket-chip, BOOM, Verilator, GTKWave, Spike (github.com/riscv-software-src/riscv-isa-sim), PicoRV32 (github.com/YosysHQ/picorv32) and EEMBC's CoreMark.
  • The SPEC CPU 2017 documentation, the Embench suite, and the STREAM, LINPACK and MLPerf pages, for the rules and reporting formats of the benchmarks of Chapter 9.