Appearance
Topic Drill 1.1.5: Hardware for Machine Learning Systems
About this drill
LO AI-1.1.5 (K2): Compare the choices available for hardware to implement machine learning systems.
Lesson: Ch 1, section 1.1.5
This drill covers every point in section 1.1.5:
- Different hardware for training and for inference (the smartphone speech recognition example)
- The three capabilities ML benefits from: large data structures, massive parallelism, low-precision arithmetic (quantization) and its benefits
- CPU vs. GPU: precision, cores, clock speed; GPUs best for small-scale ML work
- AI-specific hardware (ASIC, System-on-a-Chip): features, edge use, cloud training
- Neuromorphic processors: not von Neumann, brain-inspired
Topic Drill: 1.1.5
Q1. A speech recognition model runs on low-end smartphones, but the team trained it on a cloud platform. Which statement from the syllabus does this illustrate?
ASmartphones are the best hardware for training ML models
BDifferent types of hardware may be used for training and for inference
CCloud hardware is only used for inference
DSpeech recognition requires neuromorphic processors
Q2. A small start-up has a modest ML project and budget for one hardware upgrade. According to the syllabus, which option is generally the best for small-scale ML work?
AGPUs
BCPUs with the highest available clock speed
CNeuromorphic processors
DCustom ASICs
Q3. A security camera must detect objects in real time on the device itself, with low power use. Where does the syllabus suggest the model should be trained, and which kind of hardware suits the camera?
ATrain on the camera's CPU; run on a cloud GPU
BTrain on a neuromorphic processor; run on a CPU
CTrain and run on the same smartphone-class chip
DTrain in the cloud on specialized hardware; run on AI-specific hardware such as an ASIC or System-on-a-Chip at the edge
Q4. Which statement about neuromorphic processors is CORRECT?
AThey are optimized to run on the von Neumann architecture
BThey are general-purpose CPUs with extra cores
CThey use brain-inspired designs that mimic neuronal structures instead of the von Neumann architecture
DThey are the standard hardware for training LLMs today
Q5. Which TWO capabilities does the syllabus say ML typically benefits from in hardware?
Select 2 options.
AHigh clock speeds
BThe ability to work with large data structures
CMassively parallel (concurrent) processing, e.g., for matrix multiplication
DHigh-precision arithmetic for complex operations
EA small number of very powerful cores
Q6. A team stores and computes a model's values with 4 bits instead of 32 bits. Which outcome does the syllabus associate with this low-precision arithmetic (quantization)?
AMore accurate predictions, because fewer rounding steps are needed
BFaster processing, lower power consumption, smaller and more cost-effective chips, and reduced bandwidth requirements
CThe model can only run on CPUs
DTraining is no longer needed
Q7. Why are general-purpose CPUs less efficient than GPUs for training and running ML models?
AThey run at lower clock speeds than GPUs
BThey cannot perform matrix multiplication
CThey are designed only for graphics processing
DThey provide complex, high-precision operations that ML does not typically need, and only a few cores
Q8. Which features does the syllabus attribute to AI-specific hardware such as purpose-built ASICs and System-on-a-Chip devices?
AA single high-speed core and high-precision arithmetic
BUnlimited memory and no power constraints
CMultiple cores, specialized data management and the capability to perform in-memory processing
DCompatibility with all ML development frameworks without conversion
0 / 8 answered
← 1.1.4 Generative AI · Next drill: 1.1.6 Development and Hosting →