FPGAs Flashcards
optional (22 cards)
What is an FPGA?
A Field-Programmable Gate Array is a reconfigurable integrated circuit that can be programmed for specific tasks.
Why are FPGAs useful for machine learning?
They provide parallelism, hardware-level customization
What are common ML applications for FPGAs?
Speech recognition, self-driving cars
What is a neural processing unit (NPU)?
A specialized processor designed to accelerate neural network computations.
What operations do NPUs typically perform?
Memory transfers, dot products
What is a multi-layer deep neural network?
A neural network with multiple layers of neurons that processes inputs through weights and activations.
What hardware features are beneficial for ML?
Parallel multipliers, accumulators
What are the limitations of microprocessors in ML?
Limited parallelism fixed datapath width
How do GPUs differ from FPGAs for ML?
GPUs are easier to program and good for matrix ops but use more power and lack task-specific efficiency.
What is a tensor processing unit (TPU)?
A processor specialized for tensor operations used in deep learning offering high performance and efficiency.
What is reconfigurable computing?
Using programmable hardware like FPGAs to adapt circuits for specific computational tasks.
What is a look-up table (LUT) in an FPGA?
A small memory used in FPGA logic elements to implement logic functions.
What are DSP blocks in FPGAs?
Fixed-function blocks for multiply-accumulate operations, critical in ML computations.
What is dynamic reconfiguration in FPGAs?
The ability to change part of the FPGA design while it’s running, like loading new weights.
What are the advantages of FPGAs in ML?
High throughput, reconfigurability
What is the Microsoft Project Brainwave?
An FPGA-based architecture for low-latency deep learning in Azure data centers.
What is a systolic array?
A hardware architecture where data moves rhythmically through processing elements for efficient computation.
What is a tensor tile?
An AI ASIC component integrated with FPGAs to accelerate tensor operations.
What is the significance of memory bandwidth in DL hardware?
It limits how fast data can be supplied to processing units, often becoming a bottleneck.
What is the roofline model?
A performance model illustrating limits imposed by computation speed and memory bandwidth.
Why are custom hardware architectures important for deep learning?
They allow optimizations tailored to ML tasks, improving speed
What is the benefit of using low-precision arithmetic in ML hardware?
It reduces memory and computation requirements, improving speed and efficiency.
