Modern CPU Architectures: From 2,300 Transistors to 65 Billion

Fifty Years in Fifty Seconds

In 1971, Intel released the 4004 microprocessor. It was the first commercially available CPU on a single chip — a remarkable achievement that fit the processor, memory controller, and I/O interface onto one piece of silicon the size of a thumbnail. It ran at 740 kHz, processed 4 bits at a time, and contained exactly 2,300 transistors.

In 2024, Apple's M4 chip contains 28 billion transistors. Intel's Core Ultra 9 285K contains over 23 billion. AMD's EPYC Genoa, designed for servers, contains 75 billion transistors across multiple chiplets.

That's a 10-million-fold increase in transistor count in 53 years — an engineering achievement with no parallel in human history. Understanding how we got here, and what the three dominant modern architectures — x86-64, ARM, and RISC-V — actually look like, is the final piece of the computer architecture puzzle.

---

The Process Node: Making Transistors Smaller

Every advance in CPU capability ultimately rests on the ability to etch smaller transistors onto silicon:

TSMC N3 (3nm, 2022): 292 million transistors per mm² — an almost incomprehensible density. The "nm" node numbers are now marketing labels rather than literal transistor gate lengths, but the density improvements they represent are real.

Why smaller is better:

• Faster switching (shorter electron travel distances)
• Lower power per transistor (smaller capacitance)
• More transistors per die (more features, more cores, bigger caches)

---

x86-64: The Architecture That Refused to Die

A Brief History

The x86 architecture began with the Intel 8086 in 1978 — a 16-bit processor designed for the emerging personal computer market. When IBM chose it for the IBM PC in 1981, x86's dominance was cemented by market forces that would prove nearly impossible to dislodge.

x86 accumulated 45 years of backward compatibility baggage: 16-bit real mode, 32-bit protected mode (IA-32, added with the 80386 in 1985), and finally 64-bit mode (AMD64/x86-64), introduced by AMD in 2003 with the Opteron server processor — later adopted by Intel as Intel 64.

The modern x86-64 ISA is famously complex: Variable-length instructions (1–15 bytes), hundreds of legacy instruction forms, CISC encoding that microcode-decodes into RISC-like micro-operations (µops) internally. This is why x86 CPUs contain a "front-end" that decodes complex instructions into simpler internal forms before the out-of-order execution engine sees them.

Key x86-64 Extensions

Intel's Hybrid Architecture (12th Gen+)

Intel's Alder Lake (2021) introduced a radical change: two types of cores on the same die.

• P-cores (Performance cores): Full Gracemont/Raptor Cove µarchitecture with hyperthreading, large caches, deep out-of-order execution. For foreground, latency-sensitive work.
• E-cores (Efficiency cores): Smaller, lower-power cores without hyperthreading. For background tasks, OS work, parallelizable workloads.

The Intel Thread Director hardware communicates with the OS scheduler, telling it which tasks are CPU-intensive vs. background, so the right tasks land on the right cores.

Intel Core i9-14900K (Raptor Lake Refresh, 2023):

• 8 P-cores (hyperthreaded → 16 logical) + 16 E-cores = 24 cores, 32 threads
• L1: 48KB I + 48KB D per P-core; L2: 2MB per P-core, 4MB per 4 E-cores; L3: 36MB shared
• Process: Intel 7 (~10nm equivalent)
• TDP: 125W base, 253W max turbo

---

ARM Architecture: The Power-Efficient Challenger

The Architecture License Model

ARM Holdings (now Arm Ltd, owned by SoftBank) designs CPU architectures but does not manufacture chips. It licenses its architecture to semiconductor companies who then design their own implementations. This model has made ARM the most widely deployed CPU architecture on Earth — billions of smartphones, tablets, embedded controllers, and increasingly, laptops and servers.

ARMv8-A / ARMv9-A: The Modern 64-bit ISA

ARMv8-A (introduced 2011) brought 64-bit computing to ARM with the AArch64 execution state:

• 31 general-purpose 64-bit registers (x0–x30) — compared to x86-64's 16
• More registers = fewer memory spills in compiled code = better performance
• NEON SIMD: 128-bit vector registers (v0–v31), supports 8/16/32/64-bit integer and 32/64-bit float operations
• Dedicated link register (x30): return address for function calls
• Fixed 32-bit instruction width (unlike x86's variable length) — simpler decode

ARMv9-A (2021) adds:

• SVE2 (Scalable Vector Extension 2): vectors from 128 to 2048 bits — the width is determined at runtime, making code automatically scale across different hardware implementations
• MTE (Memory Tagging Extension): hardware-enforced memory safety tags — helps catch buffer overflows and use-after-free bugs at hardware speed
• Realm Management Extension: hardware-enforced confidential computing

Major ARM Implementers

Apple Silicon: The Unified Memory Revolution

Apple's M-series chips (M1 through M4, 2020–2024) represent a fundamental architectural departure from traditional CPU+GPU systems:

• Unified Memory Architecture (UMA): CPU and GPU share the same physical memory pool — no PCIe DMA copies between CPU RAM and GPU VRAM
• Memory bandwidth: M4 Max — 546 GB/s (vs Intel Core i9-14900K's ~90 GB/s with DDR5)
• Process: TSMC 3nm (N3E) for M4
• Transistors: M4 = 28 billion; M4 Max = 92 billion
• Power efficiency: M4 Pro delivers Intel i9-level performance at ~30W vs 125W

---

RISC-V: The Open-Source ISA

RISC-V (pronounced "risk-five") is an open, free instruction set architecture developed at UC Berkeley starting in 2010. Unlike x86 (Intel/AMD proprietary) and ARM (licensed), RISC-V is free to implement — no licensing fees, no royalties, no restrictions.

Architecture Philosophy

RISC-V is a clean-slate RISC design incorporating 50 years of ISA lessons:

• Base integer ISA: RV32I (32-bit) or RV64I (64-bit) — just 47 instructions
• Modular extensions: M (multiply/divide), A (atomics), F (single-precision float), D (double), C (compressed 16-bit instructions), V (vector), H (hypervisor)
• A standard RISC-V chip for Linux: RV64GC (G = IMAFD, C = compressed)
• 32 general-purpose registers (x0 is hardwired to 0)

Who Uses RISC-V

RISC-V vs ARM vs x86-64

---

The Chiplet Revolution

Modern CPUs no longer require all transistors on a single monolithic die. Chiplet architectures connect multiple smaller dies via high-speed interconnects:

• AMD 3D V-Cache (2022): stacks a second L3 cache die on top of the compute die using TSMC's SoIC bonding. The Ryzen 7 5800X3D gained 15% gaming performance from 64MB of stacked cache with virtually no silicon area penalty on the main die.
• Intel Foveros (2019+): 3D stacking of tiles. Intel Core Ultra (Meteor Lake) uses separate compute, graphics, SoC, and I/O tiles connected via Foveros die-to-die links.
• AMD EPYC (Rome/Milan/Genoa): Server CPUs made from 8–12 compute chiplets + 1 I/O die. Allows mixing of TSMC (compute) and AMD's GF (I/O) process nodes.

Chiplets solve a fundamental yield problem: a 500mm² monolithic die has much lower yield than eight 60mm² dies — defects that would fail the whole large die now fail only one small chiplet.

---

Summary

Three architectures dominate modern computing: x86-64 (complex, backward-compatible, dominant in desktops and servers), ARM (lean, power-efficient, dominant in mobile and rapidly gaining in laptops and servers), and RISC-V (open, royalty-free, growing in embedded systems and academic research). All three trace their lineage to fundamental architectural choices made in the 1970s–1990s and have evolved through silicon process node improvements from 10 microns to 3 nanometers. The chiplet revolution is extending Moore's Law-era progress beyond what's achievable on monolithic dies, enabling CPUs with tens of billions of transistors that deliver unprecedented performance at competitive power levels.