Lu, Yukuan

Sources:

• Download file from Github with CLI tools
• Download file from Google Drive with CLI tools

Sources:

• Installing Jupyter

Sources:

• Building makemore Part 5: Building a WaveNet
• WaveNet 2016 from DeepMind

See my Github repo for the full implementation.

Sources:

1. Remzi H. Arpaci-Dusseau & Andrea C. Arpaci-Dusseau. (2014). Virtualization. Operating Systems: Three Easy Pieces (Version 0.8.). Arpaci-Dusseau Books, Inc..

Sources:

• TPU v1: In-Datacenter Performance Analysis of a Tensor Processing Unit. 2017.
• TPU v2, v3: A Domain Specific Supercomputer for Training Deep Neural Networks.
• AI Chips: Google TPU
• HotChips 2019 Tutorial Cloud TPU: Codesigning Architecture and Infrastructure

Many architects believe that major improvements in cost-energy-performance must now come from domain-specific hardware.

Sources:

• Adjrej Karpathy's video Building makemore Part 4: Becoming a Backprop Ninja.
• Paper: Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift
• PyTorch: BatchNorm1d

Sources:

• Adjrej Karpathy's video Building makemore Part 4: Becoming a Backprop Ninja.

Sources:

• UWashington: CSE378, Lecture11
• UWashington: CSE378, Lecture12

Note: the assembly code in this article can be MIPS or RISCV. This shouldn't be consufusing since the only big difference between them is that MIPS add a $ before the name of each register:

1
2
3
4
5

# MIPS:
add $t0, $t1, $t2  # add values in $t1 and $t2, the result is stored in $t0

# RISC-V
add t0, t1, t2 # add values in t1 and t2, the result is stored in t0

Sources:

• UWashington: CSE378, Lecture12

Note: the assembly code in this article can be MIPS or RISCV. This shouldn't be consufusing since the only big difference between them is that MIPS add a $ before the name of each register:

1
2
3
4
5

# MIPS:
add $t0, $t1, $t2  # add values in $t1 and $t2, the result is stored in $t0

# RISC-V
add t0, t1, t2 # add values in t1 and t2, the result is stored in t0

Papers:

1. RSSM

2. DreamerV1

3. DreamerV2

4. DreamerV3