Split View: Google TPU 완전 해부: Systolic Array가 행렬 곱셈을 어떻게 완벽히 해결하는가

Systolic Array: 4x4 grid of MAC (Multiply-Accumulate) units
Each cell: receives inputs from left (A row) and top (B column),
           multiplies them, adds to its accumulator

Time step 0: Setup
     B[0,0]  B[1,0]  B[2,0]  B[3,0]   <- B matrix columns flow DOWN
       |       |       |       |
A[0,0]->[MAC][MAC][MAC][MAC]
A[1,0]->[MAC][MAC][MAC][MAC]
A[2,0]->[MAC][MAC][MAC][MAC]
A[3,0]->[MAC][MAC][MAC][MAC]
         |     |     |     |
       C col  C col  ...   ...   <- results flow OUT

Time step 1: A[0,0] reaches MAC[0,0]: computes A[0,0]*B[0,0]
Time step 2: A[0,0] passes to MAC[0,1], A[0,1] enters MAC[0,0]
             Each cell accumulates partial products over time

KEY INSIGHT:
- Every MAC unit is busy EVERY SINGLE CLOCK CYCLE
- No memory reads during computation (data is already "in flight")
- TPU v1: 256x256 = 65,536 MACs running simultaneously!
- Perfect data reuse: each value read from memory contributes to 256 operations

# Naive matrix multiplication — the memory access problem
def matmul_naive(A, B, N):
    C = [[0] * N for _ in range(N)]
    for i in range(N):
        for j in range(N):
            for k in range(N):
                # READ A[i][k] FROM MEMORY
                # READ B[k][j] FROM MEMORY
                C[i][j] += A[i][k] * B[k][j]
    return C

# Memory accesses: O(N^3) — catastrophic at large N
# For N=256: 16,777,216 memory reads!
# Cache miss penalty: ~100-300 cycles each

# Systolic Array solution:
# A values: read once, flow through entire row (256 ops per read)
# B values: read once, flow through entire column (256 ops per read)
# Total memory accesses: O(N^2) instead of O(N^3)
# For N=256: 65,536 reads vs 16M reads — 256x fewer!

TPU v1 Hardware Specifications:
+-----------------------------------------------------+
| Systolic Array:   256 x 256 = 65,536 MAC units      |
| Data types:       INT8 weights, INT32 accumulator    |
| Clock speed:      700 MHz                            |
| On-chip memory:   28 MB (Weight FIFO + Unified Buf)  |
| Memory bandwidth: 30 GB/s (DDR3)                     |
| Power draw:       40W                                |
| Process node:     28nm CMOS, PCIe card form factor   |
+-----------------------------------------------------+

Performance calculation:
65,536 MACs x 700 MHz x 2 (one multiply + one add) = 92 TOPS (INT8)

Competitive comparison:
- TPU v1:  92 TOPS @ 40W  = 2.30 TOPS/W
- K80 GPU: 8.7 TOPS @ 300W = 0.029 TOPS/W

Result:
- 79x better energy efficiency than K80
- 10.6x higher raw throughput

GPU Philosophy: "I can do everything"
+---------------------------------------------+
| Thousands of general-purpose CUDA cores     |
| Flexible CUDA programming model             |
| Complex branching and control flow support  |
| Arbitrary memory access patterns            |
| Graphics, simulation, AI — all supported    |
| Overhead: scheduler, register file, caches  |
+---------------------------------------------+

TPU Philosophy: "I only do matrix multiplication — perfectly"
+---------------------------------------------+
| Systolic Array (dedicated to GEMM)          |
| Deterministic dataflow (compiler decides)   |
| Limited operation set (MAC-centric)         |
| No complex control logic needed             |
| 100% hardware utilization, no waste         |
| Maximized energy efficiency                 |
+---------------------------------------------+

Why specialization wins:
Transformer operations: ~95% are GEMM (matrix multiply)
-> Hardware specialized for GEMM dominates everything else

# FLOPS breakdown for a GPT-2-scale transformer (124M params)
layer_config = {
    'd_model': 768,
    'n_heads': 12,
    'n_layers': 12,
    'seq_len': 1024
}

d = layer_config['d_model']     # 768
n = layer_config['seq_len']     # 1024
L = layer_config['n_layers']    # 12

# Q, K, V projections: 3 separate [n x d] x [d x d] GEMMs per layer
qkv_flops = 3 * 2 * n * d * d * L

# Attention: Q x K^T (batched GEMM) + weighted sum by V
attn_flops = (2 * n * n * d + 2 * n * n * d) * L

# FFN: two GEMMs [d -> 4d -> d]
ffn_flops = (2 * n * d * (4*d) + 2 * n * (4*d) * d) * L

# Output projection per layer
out_flops = 2 * n * d * d * L

total = qkv_flops + attn_flops + ffn_flops + out_flops
print(f"QKV projections: {qkv_flops/total*100:.1f}%")   # ~38%
print(f"Attention GEMM:  {attn_flops/total*100:.1f}%")  # ~12%
print(f"FFN:             {ffn_flops/total*100:.1f}%")   # ~41%
print(f"Output proj:     {out_flops/total*100:.1f}%")   # ~9%
# Total GEMM: ~100% (attention is also batched GEMM)

TPU v4 Memory Hierarchy:
+-----------------------------------------------------------+
|     Weight FIFO (on-chip, feeds directly to Systolic)     |
|     Capacity: 32 MB / Bandwidth: 2.7 TB/s                 |
|     Purpose: stream weights into array without stalls      |
+-----------------------------------------------------------+
|     Unified Buffer (on-chip, holds activations)           |
|     Capacity: 256 MB / Bandwidth: 900 GB/s                |
|     Purpose: inter-layer intermediate activation storage  |
+-----------------------------------------------------------+
|     High Bandwidth Memory (HBM, off-chip)                 |
|     Capacity: 32 GB / Bandwidth: 1.2 TB/s                 |
|     Purpose: store the full model weights                 |
+-----------------------------------------------------------+

Data flow:
HBM -> Weight FIFO -> Systolic Array  (weight stream)
HBM -> Unified Buffer -> Systolic Array  (activation stream)
Systolic Array -> Unified Buffer -> next layer

Weight Stationary:
- Pre-load weights (W) into each MAC unit
- Stream input activations (X) through the array
- Compute: W x X
- Advantage: maximizes weight reuse (great for large batches)
- Used by: TPU v1 for inference

Output Stationary:
- Each MAC accumulates one output element C[i][j]
- Stream both input matrices through
- Advantage: minimizes output writes (great for small batches)
- Better for single-sample inference

When to use which:
- Batch inference (batch_size > 32): Weight Stationary
- Single-sample inference: Output Stationary
- Training (large batches): Weight Stationary

Floating-point format comparison:

FP32:     [1 sign][8 exponent][23 mantissa] = 32 bits
          Range: +/- 3.4 x 10^38
          Precision: ~7 decimal digits

FP16:     [1 sign][5 exponent][10 mantissa] = 16 bits
          Range: +/- 6.5 x 10^4 (dangerously narrow!)
          Precision: ~3-4 decimal digits
          Problem: gradients vanish/explode during training

bfloat16: [1 sign][8 exponent][7 mantissa]  = 16 bits (Google's design)
          Range: +/- 3.4 x 10^38 (SAME as FP32!)
          Precision: ~2-3 decimal digits
          Key trick: just drop the last 16 bits of FP32!

Why bfloat16 wins for deep learning:
1. Same exponent range as FP32 -> no overflow/underflow in gradients
2. Less mantissa -> DL needs range more than precision
3. FP32 -> bfloat16 conversion: just truncate last 16 bits (free!)
4. Mixed precision: FP32 master weights + bfloat16 compute

import jax
import jax.numpy as jnp
import numpy as np

# Demonstrating bfloat16 properties
x_fp32 = np.float32(1.0 / 3.0)
x_bf16 = jnp.bfloat16(1.0 / 3.0)
x_fp16 = np.float16(1.0 / 3.0)

print(f"FP32:    {float(x_fp32):.10f}")  # 0.3333333433
print(f"BF16:    {float(x_bf16):.10f}")  # 0.3320312500 (less precise, OK)
print(f"FP16:    {float(x_fp16):.10f}")  # 0.3334960938

# Range comparison (critical for training)
print(f"FP32 max:  {np.finfo(np.float32).max:.2e}")   # 3.40e+38
print(f"BF16 max:  {float(jnp.finfo(jnp.bfloat16).max):.2e}")  # 3.39e+38
print(f"FP16 max:  {np.finfo(np.float16).max:.2e}")   # 6.55e+04 << problem!

# In practice: training loss at step 10000
# FP32:  0.2341 (clean, stable)
# BF16:  0.2343 (virtually identical)
# FP16:  NaN    (gradient overflow in many models!)

TPU v4 Pod Topology:

Each TPU v4 chip connects to 6 neighbors: +X, -X, +Y, -Y, +Z, -Z
This forms a 3-dimensional torus network

Full TPU v4 Pod: 16 x 16 x 16 = 4,096 chips in 3D torus
Per-chip interconnect bandwidth: 600 GB/s

Why 3D torus?
- Max hops between any two nodes: O(N^(1/3))
- At 4096 nodes: max 24 hops (vs 2048 hops for a simple ring)
- Collective communication (AllReduce) is highly efficient
- Bisection bandwidth scales well with pod size

OCS (Optical Circuit Switch):
- Software-configurable optical switching fabric
- Dynamically reconfigures the torus topology per workload
- Eliminates electrical switch bottlenecks
- Enables full bandwidth between any chip pairs

# JAX code that compiles to XLA -> runs on TPU
import jax
import jax.numpy as jnp
from functools import partial

@jax.jit  # JIT compile via XLA
def transformer_forward(x, params):
    """Single transformer layer, JAX style"""
    w_q, w_k, w_v, w_o = params['attn']
    w_ff1, w_ff2 = params['ffn']

    # Layer norm
    x_norm = jax.nn.standardize(x, axis=-1)

    # Multi-head attention projections (GEMM -> Systolic Array)
    q = jnp.dot(x_norm, w_q)   # [batch, seq, d_model] x [d_model, d_head]
    k = jnp.dot(x_norm, w_k)
    v = jnp.dot(x_norm, w_v)

    # Attention (batched GEMM)
    d_head = q.shape[-1]
    scale = jnp.sqrt(float(d_head))
    scores = jnp.einsum('bqh,bkh->bqk', q, k) / scale
    attn_weights = jax.nn.softmax(scores, axis=-1)
    attended = jnp.einsum('bqk,bkh->bqh', attn_weights, v)

    # Output projection
    out = jnp.dot(attended, w_o) + x   # residual

    # FFN
    x2_norm = jax.nn.standardize(out, axis=-1)
    hidden = jax.nn.gelu(jnp.dot(x2_norm, w_ff1))
    ffn_out = jnp.dot(hidden, w_ff2) + out   # residual

    return ffn_out

# What XLA does to this code:
# 1. Operation fusion: LayerNorm (5 ops) -> single kernel
# 2. Layout optimization: pick memory layout for Systolic Array
# 3. Rematerialization: recompute activations vs store (memory/compute tradeoff)
# 4. Auto-sharding: split across TPU chips with minimal communication
# 5. Constant folding: pre-compute anything computable at compile time

Without fusion (naive):
  LayerNorm breakdown:
  Step 1: mean(x) -> write to HBM
  Step 2: x - mean -> write to HBM
  Step 3: variance(x - mean) -> write to HBM
  Step 4: normalize -> write to HBM
  Step 5: scale * x + bias -> write to HBM
  Total: 5 HBM round trips per layer

With XLA fusion:
  LayerNorm: single kernel
  - Read input once from HBM
  - Do all 5 computations in SRAM (on-chip)
  - Write output once to HBM
  Total: 1 HBM round trip per layer

Impact: 5x reduction in memory bandwidth usage
XLA applies hundreds of such fusion patterns automatically

import jax
import jax.numpy as jnp
from jax.experimental import mesh_utils
from jax.sharding import Mesh, PartitionSpec, NamedSharding
from functools import partial
import numpy as np

# Check available TPU devices
print(f"Available TPUs: {jax.devices()}")
# [TpuDevice(id=0, process_index=0, coords=(0,0,0), core_on_chip=0), ...]

# Create 8-TPU mesh for tensor parallelism
num_devices = 8  # one TPU v4 chip = 4 cores; 2 chips = 8 cores
device_mesh = mesh_utils.create_device_mesh((num_devices,))
mesh = Mesh(device_mesh, axis_names=('model',))

# Define sharding strategies
# Weight matrix W: shard along columns (model-parallel)
W_sharding = NamedSharding(mesh, PartitionSpec(None, 'model'))
# Bias: replicated on all devices
bias_sharding = NamedSharding(mesh, PartitionSpec(None))
# Activations: replicated (all devices see same input)
act_sharding = NamedSharding(mesh, PartitionSpec(None, None))

@partial(jax.jit,
         in_shardings=(act_sharding, W_sharding, bias_sharding),
         out_shardings=act_sharding)
def sharded_linear(x, W, b):
    """Column-parallel linear layer across 8 TPUs"""
    # Each TPU computes [batch, seq, d_model/8]
    local_out = jnp.dot(x, W)
    # AllReduce to sum partial results (automatic!)
    return local_out + b

# Real inference pipeline for a 7B model
def create_model_params(d_model=4096, d_ffn=16384, n_layers=32):
    """Create model params sharded across 8 TPUs"""
    params = {}
    for i in range(n_layers):
        # Each layer's FFN weights sharded across devices
        params[f'layer_{i}_ff1'] = jax.device_put(
            np.random.randn(d_model, d_ffn).astype(np.float16),
            W_sharding
        )
        params[f'layer_{i}_ff2'] = jax.device_put(
            np.random.randn(d_ffn, d_model).astype(np.float16),
            W_sharding
        )
    return params

# Inference: generate one token
def generate_next_token(params, input_ids, kv_cache):
    batch_size, seq_len = input_ids.shape

    # Run through all layers
    x = embedding_lookup(params['embed'], input_ids)  # [batch, seq, d_model]

    for layer_idx in range(32):
        # Attention with cached KV
        x = attention_with_cache(x, params, kv_cache, layer_idx)
        # FFN (sharded linear layers)
        with mesh:
            ffn_hidden = sharded_linear(
                x,
                params[f'layer_{layer_idx}_ff1'],
                params[f'bias_{layer_idx}_ff1']
            )
            x = x + sharded_linear(
                jax.nn.gelu(ffn_hidden),
                params[f'layer_{layer_idx}_ff2'],
                params[f'bias_{layer_idx}_ff2']
            )

    # Final logits
    logits = jnp.dot(x[:, -1, :], params['lm_head'])
    return logits.argmax(axis=-1)

PaLM 540B Training Configuration (Google, 2022):
- Hardware:      TPU v4 Pod x 6,144 chips
- Total compute: 6,144 x 275 TFLOPS = 1.69 ExaFLOPS
- Training data: 780 billion tokens
- Training time: ~57 days
- Batch size:    2,048 sequences x 2,048 tokens
- MFU:           46.2% (Model FLOPS Utilization)

Gemini Ultra Training (Google DeepMind, 2023):
- Hardware:      Multiple TPU v5p pods via OCS
- First model to outperform human experts on MMLU
- Context length training: up to 32,768 tokens

Why TPU Pods beat GPU clusters for this scale:
1. 3D torus: O(N^1/3) hop distance vs O(N) for ring-allreduce
2. OCS: full bandwidth between any chip pair (no switch bottleneck)
3. XLA: whole-model compilation, global optimization
4. bfloat16: stable training without loss scaling tricks

Llama 2 70B Inference Benchmark (batch_size=1, float16):

Hardware          | Mem Bandwidth | Throughput  | Latency
------------------|---------------|-------------|----------
NVIDIA H100 SXM   | 3.35 TB/s     | ~120 tok/s  | ~8.3 ms/tok
NVIDIA A100 80GB  | 2.0  TB/s     | ~72  tok/s  | ~13.9 ms/tok
TPU v5p (4-chip)  | 4x 1.2 TB/s  | ~160 tok/s  | ~6.3 ms/tok
AMD MI300X        | 5.3  TB/s     | ~190 tok/s  | ~5.3 ms/tok
Apple M3 Ultra    | 800  GB/s     | ~55  tok/s  | ~18.2 ms/tok

Key insight: throughput correlates almost perfectly with memory bandwidth!
This is the "memory-bound" nature of LLM inference — more in the NPU post.

TPU advantages at scale (beyond single-chip numbers):
- TPU Pod AllReduce: 600 GB/s per chip interconnect
- GPU NVLink: 900 GB/s (H100), but only within NVSwitch domain
- TPU OCS: reconfigurable topology for different workloads
- Cost: TPU v5p ~$2.20/hr vs H100 ~$3.50/hr on comparable clouds

# Profile your JAX/TPU code to find bottlenecks
import jax
import jax.numpy as jnp

# Method 1: JAX Profiler (produces Chrome Trace format)
with jax.profiler.trace("/tmp/jax_trace", create_perfetto_link=True):
    # Warmup (needed to exclude JIT compilation time)
    _ = my_model_fn(sample_input).block_until_ready()
    # Actual profiled run
    result = my_model_fn(real_input).block_until_ready()

# Open in Perfetto UI: https://ui.perfetto.dev
# Look for:
# - Long gaps between ops (memory bandwidth bound)
# - Unbalanced ops across devices (sharding inefficiency)
# - Frequent recompilation (dynamic shapes)

# Method 2: Measure Model FLOPS Utilization (MFU)
def compute_mfu(model_flops, elapsed_seconds, peak_tflops):
    """MFU = actual FLOPS / theoretical peak FLOPS"""
    actual_tflops = model_flops / elapsed_seconds / 1e12
    return actual_tflops / peak_tflops * 100

# Example: 7B model, 100 token/s throughput
flops_per_token = 2 * 7e9   # 2 * num_params (rough estimate)
elapsed = 1.0 / 100         # 1 sec / 100 tokens = 0.01 sec/token
tpu_v5p_tflops = 459

mfu = compute_mfu(flops_per_token, elapsed, tpu_v5p_tflops)
print(f"MFU: {mfu:.1f}%")
# Good MFU: >40% means you're using the hardware well
# Low MFU: memory bandwidth bound (typical for inference)

# Method 3: Identify recompilation events
# Add logging to catch shape changes that trigger recompiles
jax.config.update("jax_log_compiles", True)
# Check stderr for "Compiling..." messages during inference