Accelerators for the Architect · arch

Computer Architecture · All levels

6topics
6lessons

An accelerator wins when the workload has steady parallel work, the dataflow keeps operands local, and the platform cost of launching and feeding it stays below the saved compute.

Pattern	Assumes	Breaks when
Systolic / MAC array	Dense regular math with reuse	Shapes mismatch the array, sparsity grows
SIMD / vector	Wide regular loops	Lanes diverge, masks go sparse
Reduction tree	Many-to-few combines	Reduction depth exceeds accumulator residency
Streaming pipeline	Producer-consumer flow	FIFOs underrun, backpressure stalls

Launch, movement, compute and drain path.

Work flows from launch to movement to resident compute to drain; each stage can cap the whole.

Lesson	What it covers
When an accelerator pays off	Amdahl bounds plus launch and transfer costs, with kill criteria.
Dataflow patterns as reuse contracts	Weight, output and input stationary styles, with traffic counted by hand.
Systolic execution and utilization cliffs	Fill, drain and padding losses, proved by a runnable MAC pipeline.
Locality, scratchpads and tile sizing	Bytes per operation, scratchpads and tile fit against SRAM capacity.
DMA and double buffering	Overlapping transfers with compute, proved by a runnable overlap model.
Integration, launch cost and signoff	Coherency choice, launch budgets, tail latency and release gates.

Check yourself Peak array throughput doubles but end-to-end time barely moves. Name two non-compute suspects.

Transfer time and launch overhead. If tiles arrive late or each launch pays a fixed cost, the array waits and the gain stays unrealized.

Up nextWhen an Accelerator Pays Off

Practice this lesson Report an error

Pattern

Assumes

Breaks when

Systolic / MAC array

Dense regular math with reuse

Shapes mismatch the array, sparsity grows

SIMD / vector

Wide regular loops

Lanes diverge, masks go sparse

Reduction tree

Many-to-few combines

Reduction depth exceeds accumulator residency

Streaming pipeline

Producer-consumer flow

FIFOs underrun, backpressure stalls

Lesson

What it covers

When an accelerator pays off

Amdahl bounds plus launch and transfer costs, with kill criteria.

Dataflow patterns as reuse contracts

Weight, output and input stationary styles, with traffic counted by hand.

Systolic execution and utilization cliffs

Fill, drain and padding losses, proved by a runnable MAC pipeline.

Locality, scratchpads and tile sizing

Bytes per operation, scratchpads and tile fit against SRAM capacity.

DMA and double buffering

Overlapping transfers with compute, proved by a runnable overlap model.

Integration, launch cost and signoff

Coherency choice, launch budgets, tail latency and release gates.