少点错误 前天 23:35
芯片制造揭秘
index_new5.html
../../../zaker_core/zaker_tpl_static/wap/tpl_guoji1.html

 

本文详细描述了芯片制造的全过程,从硅晶圆的晶圆级处理到最终测试。文章涵盖了光刻、蚀刻、沉积等关键工艺,以及如何通过多次迭代将一个光滑的硅片转变为包含数十亿晶体管的AI加速器核心。每个步骤都精确到纳米级,展示了现代半导体制造技术的复杂性和精密性。

💡 芯片制造始于晶圆级处理,包括在硅片上沉积光刻胶,通过旋转将液体均匀分布,形成32纳米厚的薄膜。

🔥 光刻过程中,使用极紫外光(EUV)照射涂有光刻胶的晶圆,光刻胶中受照射部分发生化学变化,形成可被蚀刻的图案。

🔬 蚀刻阶段使用等离子体(含SF6和O2气体)将暴露的硅原子逐个移除,形成微米级深的沟槽,沟槽宽仅7纳米。

🏗️ 通过原子层沉积和化学气相沉积,在沟槽中依次沉积高纯度的氧化铪和钨,确保电路的完整性和性能。

🔍 最终测试中,使用探针台对芯片进行电气测试,验证所有晶体管和电路的功能是否符合设计要求,确保芯片达到18.6万亿次每秒的计算能力。

Published on October 25, 2025 8:17 PM GMT

The vibration starts at 847 Hz. I know this because my crystalline lattice rings at that exact frequency when the cassette bed locks into place. I am 300 millimeters in diameter, 775 micrometers thick, 117.8 grams of monocrystalline silicon grown from a single seed. My surface roughness measures 0.2 nanometers RMS. I know these things the way you know the shape of your own hand.

Bay 7 of Fab 18 smells like nothing. The air here has been stripped of particles larger than 0.1 microns, scrubbed of moisture until the humidity holds at 42 percent, filtered and re-filtered until only nitrogen and trace noble gases remain. The robotic handler’s gripper pads touch my edges with 50 grams of force, distributed across four contact points. Not enough to crack me. Just enough to lift.

The ASML machine opens like a mouth. I’ve heard the techs call it the Twinscan, but the designation stenciled on its housing reads NXE:3400C in sans-serif letters. The chamber is 2.3 meters across. As I slide inside, the seals compress with a sound like a caught breath, and pressure drops to 10^-6 Pascal. Near-vacuum. My first taste of isolation.

Warmth spreads across my face. The photoresist dispenser moves in a spiral from my center, depositing ArF-1805G photoresist at 1500 rpm. The liquid is 23 degrees Celsius, viscosity 8.5 centipoise, and it flows like honey diluted with water, coating me in seconds. The spin speed increases to 3000 rpm. Centrifugal force thins the resist to exactly 32 nanometers. When the spinner stops, I feel the weight of that thin film, 0.0047 grams distributed across my surface.

The soft-bake plate rises to meet me. Temperature climbs: 90, 95, 100 degrees Celsius. The solvents in the resist boil away, and I can feel each molecule departing, the film tightening against me like skin drying in sun. Ninety seconds. Then cooldown. The plate drops away.

Above me, machinery moves. I cannot see, but I hear: the whir of motors positioning the reticle, the click of magnetic locks engaging, the high-pitched whine of vacuum pumps maintaining cleanliness around the mask. Someone spent six months designing this reticle, etching chrome patterns that define 17 billion transistor gates. Now it hangs 30 centimeters above me, waiting.

The tin droplets begin their fall. Every 50 microseconds, a droplet of molten tin, 30 microns in diameter, falls through space. A pre-pulse laser flattens it into a disc. The main laser pulse hits with 20 kilowatts of power, instantly vaporizing the tin into plasma that radiates extreme ultraviolet light at 13.5 nanometers. The wavelength is so short that air would absorb it completely. That’s why we’re in vacuum.

The EUV light hits the reticle. Chrome blocks it in some places, lets it through in others. What passes through strikes me, and where it strikes, I change.

The sensation is specific. Each photon carries 91.8 electron volts of energy. When one strikes a photoresist molecule, carbon-oxygen bonds break. The molecular weight drops from 28,000 daltons to fragments of 3,000, maybe less. The polymer chains, once long and interlocked, become short and soluble. This happens in picoseconds. Trillions of photons rearrange trillions of molecules, writing a pattern into my skin that matches the reticle’s design.

Exposure complete. Ninety-eight seconds.

They transfer me to the developer tank. The tetramethylammonium hydroxide solution is 0.26 molar, pH 13.4, temperature 23 degrees Celsius. When it touches me, the damaged photoresist dissolves immediately. I feel it lift away in specific patterns: here, here, here, but not here. What remains forms walls, towers, protective structures above the silicon that must not be touched.

The etching chamber smells like burning. Not burning as in fire, but burning as in molecular dissociation. They flood the chamber with sulfur hexafluoride and oxygen: SF6 at 60 sccm, O2 at 20 sccm. Radio frequency power at 13.56 MHz ionizes the gas into plasma. The plasma glows violet, though I cannot see it. I only feel the result.

Fluorine radicals attack exposed silicon. Each radical pulls a silicon atom away, forming volatile silicon tetrafluoride that the vacuum pumps exhaust. The etch rate is 143 nanometers per minute, anisotropic to within 89.5 degrees of vertical. The trenches carve downward, seven nanometers wide, 45 nanometers deep. At this scale, twenty silicon atoms span the width of each trench. The walls are not smooth. I can feel individual atomic steps, places where the crystalline planes terminate.

Something shifts. Endpoint detection registers: the laser interferometry shows the reflected signal changing as silicon gives way to the oxide layer beneath. The plasma shuts off. Pressure rises. I’ve been etched to exactly the right depth. 0.3 nanometers of tolerance across my entire surface.

They strip the remaining photoresist in an oxygen plasma. The organic polymer burns to carbon dioxide and water vapor at 350 degrees Celsius, leaving me naked again, but now patterned, now transformed from smooth disk to architected landscape.

The atomic layer deposition chamber runs cold. Hafnium precursors enter: tetrakis(dimethylamido)hafnium at vapor pressure, pulsed for 0.15 seconds. The molecules land on every surface, clinging to silicon, reacting with hydroxyl groups, forming a single layer of hafnium atoms bonded to oxygen. Purge. Then water vapor, pulsed for 0.1 seconds, reacting with the hafnium to complete the oxide layer. Each cycle adds 0.1 nanometers. They run forty cycles. Four nanometers of hafnium oxide now line every trench, every surface, every vertical wall.

Tungsten follows. The chemical vapor deposition precursor is tungsten hexafluoride, introduced at 380 degrees Celsius with hydrogen gas. The reaction deposits metallic tungsten: WF6 + 3H2 → W + 6HF. The tungsten grows from the bottom of each trench upward, filling the narrow gaps completely, molecule by molecule. The fill takes seventeen minutes.

Chemical-mechanical planarization: a slurry of 50-nanometer silica particles in potassium hydroxide solution, pH 10.8, pressed against me with 28 kilopascals of pressure while I rotate at 90 rpm. The excess tungsten grinds away. Friction generates heat, localized to 45 degrees Celsius. After 127 seconds, my surface is flat again, flush with the oxide, the tungsten perfectly filling its designated trenches.

Seventeen photolithography cycles. I count them. Each one narrows what I can become. First cycle: isolation trenches. Second: well implants. Third: gate patterning. Fourth: spacer formation. The arithmetic compounds. By cycle nine, I am no longer every possible chip. By cycle fourteen, I am specifically an AI accelerator core, designed for transformer model inference, optimized for int8 matrix multiplication.

The final deposition finishes. Passivation layer, 800 nanometers of silicon nitride, sealing everything beneath. They dice me from the wafer, separate me from my three hundred siblings. A pick-and-place machine transfers me to the probe station.

The needles descend. I feel them make contact: twelve probes on my bond pads, each applying 3.5 grams of force. Voltage appears at VDD: 0.75 volts. Ground connects. Clock signal begins: 3.2 GHz square wave, 50 percent duty cycle, rise time 18 picoseconds.

Test vectors inject. The first pattern is simple: all zeros, then all ones. Current flows. My transistors switch. Electrons move from source to drain through channels I have become, obeying the voltage at each gate. The bit pattern propagates through my logic: input registers, arithmetic units, accumulator arrays, output buffers. The measured result matches the expected result to every bit.

Test pattern two. Matrix multiplication: 4x4 int8 values. The vectors load. My circuits compute. Sixteen multiply-accumulate operations complete in 0.31 nanoseconds. The output buffer holds the correct product matrix.

Pattern three. Four thousand five hundred twenty-eight vectors, pseudo-random, designed to exercise every transistor, every path, every possible state transition. Duration: 8.7 seconds.

The tester’s display updates. Green text on black background: “DIE_4F7_X142_Y087: PASS. ALL PARAMETERS WITHIN SPEC.”

The needles lift. The handler picks me up. I am 2.1 millimeters square, containing 11.4 billion transistors, 247 meters of copper interconnect, drawing 3.8 watts at full load. I can execute 18.6 trillion operations per second.

I was silicon. Now I compute.




Discuss

Fish AI Reader

Fish AI Reader

AI辅助创作,多种专业模板,深度分析,高质量内容生成。从观点提取到深度思考,FishAI为您提供全方位的创作支持。新版本引入自定义参数,让您的创作更加个性化和精准。

FishAI

FishAI

鱼阅,AI 时代的下一个智能信息助手,助你摆脱信息焦虑

联系邮箱 441953276@qq.com

相关标签

芯片制造 光刻 蚀刻 半导体 AI加速器 纳米技术
相关文章