内容导航
1. Graphics Card (GPU)
1.1 What You Need to Know
The GPU is the most critical component for AI workloads. Prioritize VRAM specifications in this order:
- VRAM Capacity (Determines maximum model/data size)
- Memory Bus Width (Affects data transfer efficiency)
Analogy
- VRAM is like a water reservoir – insufficient capacity leads to overflow (VRAM exhaustion). Many AI applications enforce minimum VRAM requirements, similar to system RAM. Insufficient VRAM may prevent software from launching or cause severe lag.
- Bus width is like a pipeline – a narrower pipe slows data transfer, reducing efficiency.
1.2 Recommendations
- Minimum new GPU: RTX 4060 Ti 16GB.
- Lower-tier GPUs (e.g., 8GB models from RTX 3000/4000 series) are usable but compromise performance. “Adequate” and “optimal” are distinct tiers.
- Ideal choice: RTX 4090 or future RTX 5090 (highest-end models).
- Brand preference: NVIDIA. Most AI tools are optimized for NVIDIA, with better documentation and community support. AMD GPUs are theoretically compatible but often suffer from lower efficiency and troubleshooting hurdles (not recommended for non-experts).
1.3 Series-Specific Notes
- RTX 30-series: Performance is acceptable, but mining-used (“used in cryptocurrency mining”) cards pose significant risks. Models like the 3060 Ti 12GB may be worthwhile if the seller is trustworthy and pricing is reasonable.
- RTX 40-series: Marginal performance gains over the 30-series (except the 4090).
- RTX 50-series: Expected in 2025, likely to be the new flagship.
1.4 GPU Tier Breakdown by Brand
1.4.1 ASUS
- Budget: DUAL
- Entry-level: TUF Gaming (formerly ATS)
- Mid-range: ProArt (designer-focused)
- High-end/Sub-flagship: TUF Gaming (gamer-focused)
- Flagship: ROG Strix
- Ultra-flagship: ROG Matrix
1.4.2 MSI
- Budget: Ventus
- Mid-range: Gaming (discontinued)
- High-end/Sub-flagship: Gaming X Trio
- Flagship: Suprim X
- Ultra-flagship: Lightning (discontinued)
1.4.3 GIGABYTE
(Refer to ASUS/MSI tiers for analogous product lines.)
1.4.4 Colorful
Colorful is a reputable second-tier brand with reliable quality and after-sales support. Its iGame series is the flagship lineup:
- Budget: Netral/Lei Feng
- Entry-level: BattleAx
- Mid-range: Advanced (triple-fan) / Ultra (dual-fan)
- Flagship: iGame Vulcan/Neptune
- Ultra-flagship: Kudan
1.4.5 ZOTAC
A Hong Kong-based manufacturer with strong R&D capabilities (Sapphire, its subsidiary, dominates AMD’s GPU market).
- Budget: Twin Edge
- Mid-range: X-Gaming
- High-end: Trinity OC
- Flagship: AMP Extreme/PGF
1.4.6 GALAX
- Budget: Black/White General
- Mid-range: Metal Master
- Flagship: HOF (Hall of Fame)
1.5 Our Commentary
- Component Breakdown
A graphics card comprises:- GPU Die: Processing core (NVIDIA/AMD)
- VRAM: High-speed memory (GDDR6X/GDDR7)
- PCB Backplate: Circuitry and power delivery
- Cooling System: Heatsink/fans or liquid cooling
- Structural Parts: Casing, RGB lighting
- Core Architecture
The GPU die (designed by NVIDIA/AMD) is the performance-defining component. AIBs (Add-in Board partners) like ASUS/MSI license these dies for retail production. - Manufacturing Process
Brands integrate proprietary:- Thermal solutions (e.g., ASUS Axial-tech fans)
- PCB layouts (e.g., MSI TRI FROZR 3.0)
- Aesthetic designs (ROG RGB vs ZOTAC minimalist)
- Performance Variance
Identical GPU dies ≠ identical performance. Factors:- Power Limits: 450W (flagship) vs 350W (budget)
- Cooling: 3-8°C under load
- VRM Quality: 20+ phase vs 14-phase designs
- Price-Performance Example
Metric ASUS ROG Strix 4090 ZOTAC AMP Extreme 4090 Boost Clock 2640 MHz 2580 MHz VRM Phases 24+4 20+3 Price $1,999 $1,599 *Performance gap: <5% Price gap: 25%* - Value Proposition
Second-tier brands (Colorful, PNY) often deliver 95% of flagship performance at 70-90% cost.
2. GPU Benchmark Hierarchy
3. Specifications
3.1 Compare 50 Series Specs
3.2 Compare 40 Series Specs
| GPU Engine Specs: | |||||||||
| NVIDIA CUDA® Cores | 16384 | 10240 | 9728 | 8448 | 7680 | 7168 | 5888 | 4352 | 3072 |
| Shader Cores | Ada Lovelace 83 TFLOPS | Ada Lovelace 52 TFLOPS | Ada Lovelace 49 TFLOPS | Ada Lovelace 44 TFLOPS | Ada Lovelace 40 TFLOPS | Ada Lovelace 36 TFLOPS | Ada Lovelace 29 TFLOPS | Ada Lovelace 22 TFLOPS | Ada Lovelace 15 TFLOPS |
| Ray Tracing Cores | 3rd Generation 191 TFLOPS | 3rd Generation 121 TFLOPS | 3rd Generation 113 TFLOPS | 3rd Generation 102 TFLOPS | 3rd Generation 93 TFLOPS | 3rd Generation 82 TFLOPS | 3rd Generation 67 TFLOPS | 3rd Generation 51 TFLOPS | 3rd Generation 35 TFLOPS |
| Tensor Cores (AI) | 4th Generation 1321 AI TOPS | 4th Generation 836 AI TOPS | 4th Generation 780 AI TOPS | 4th Generation 706 AI TOPS | 4th Generation 641 AI TOPS | 4th Generation 568 AI TOPS | 4th Generation 466 AI TOPS | 4th Generation 353 AI TOPS | 4th Generation 242 AI TOPS |
| Boost Clock (GHz) | 2.52 | 2.55 | 2.51 | 2.61 | 2.61 | 2.48 | 2.48 | 2.54 | 2.46 |
| Base Clock (GHz) | 2.23 | 2.29 | 2.21 | 2.34 | 2.31 | 1.98 | 1.92 | 2.31 | 1.83 |
| Memory Specs: | |||||||||
| Standard Memory Config | 24 GB GDDR6X | 16 GB GDDR6X | 16 GB GDDR6X | 16 GB GDDR6X | 12 GB GDDR6X | 12 GB GDDR6X | 12 GB GDDR6 / 12 GB GDDR6X | 16 GB GDDR6 or 8 GB GDDR6 | 8 GB GDDR6 |
| Memory Interface Width | 384-bit | 256-bit | 256-bit | 256-bit | 192-bit | 192-bit | 192-bit | 128-bit | 128-bit |
| Display Support: | |||||||||
| Maximum Digital Resolution (1) | 4K at 240Hz or 8K at 60Hz with DSC, HDR | 4K at 240Hz or 8K at 60Hz with DSC, HDR | 4K at 240Hz or 8K at 60Hz with DSC, HDR | 4K at 240Hz or 8K at 60Hz with DSC, HDR | 4K at 240Hz or 8K at 60Hz with DSC, HDR | 4K at 240Hz or 8K at 60Hz with DSC, HDR | 4K at 240Hz or 8K at 60Hz with DSC, HDR | 4K at 240Hz or 8K at 60Hz with DSC, HDR | 4K at 240Hz or 8K at 60Hz with DSC, HDR |
| Standard Display Connectors | HDMI(2), 3x DisplayPort(3) | HDMI(2), 3x DisplayPort(3) | HDMI(2), 3x DisplayPort(3) | HDMI(2), 3x DisplayPort(3) | HDMI(2), 3x DisplayPort(3) | HDMI(2), 3x DisplayPort(3) | HDMI(2), 3x DisplayPort(3) | HDMI(2), 3x DisplayPort(3) | HDMI(2), 3x DisplayPort(3) |
| Multi Monitor | up to 4(4) | up to 4(4) | up to 4(4) | up to 4(4) | up to 4(4) | up to 4(4) | up to 4(4) | up to 4(4) | up to 4(4) |
| HDCP | 2.3 | 2.3 | 2.3 | 2.3 | 2.3 | 2.3 | 2.3 | 2.3 | 2.3 |
| Technology Support: | |||||||||
| NVIDIA Architecture | Ada Lovelace | Ada Lovelace | Ada Lovelace | Ada Lovelace | Ada Lovelace | Ada Lovelace | Ada Lovelace | Ada Lovelace | Ada Lovelace |
| Ray Tracing | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes |
| DLSS 3 | DLSS 3 | DLSS 3 | DLSS 3 | DLSS 3 | DLSS 3 | DLSS 3 | DLSS 3 | DLSS 3 | |
| Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | |
| Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | |
| PCI Express Gen 4 | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes |
| Resizable BAR | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes |
| Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | |
| Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | |
| NVIDIA FreeStyle | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes |
| Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | |
| NVIDIA Highlights | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes |
| Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | |
| Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | |
| Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | |
| Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | |
| Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | |
| Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | |
| NVIDIA GPU Boost™ | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes |
| NVIDIA NVLink™ (SLI-Ready) | No | No | No | No | No | No | No | No | No |
| Vulkan RT API, OpenGL 4.6 | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes |
| 2x 8th Generation | 2x 8th Generation | 2x 8th Generation | 2x 8th Generation | 2x 8th Generation | 1x 8th Generation | 1x 8th Generation | 1x 8th Generation | 1x 8th Generation | |
| 5th Generation | 5th Generation | 5th Generation | 5th Generation | 5th Generation | 5th Generation | 5th Generation | 5th Generation | 5th Generation | |
| Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | |
| Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | |
| 8.9 | 8.9 | 8.9 | 8.9 | 8.9 | 8.9 | 8.9 | 8.9 | 8.9 | |
| Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | |
| Card Dimensions: | |||||||||
| Length | 304 mm | 304 mm | 304 mm | Varies by manufacturer | Varies by manufacturer | 244 mm | 244 mm | 244 mm | Varies by manufacturer |
| Width | 137 mm | 137 mm | 137 mm | Varies by manufacturer | Varies by manufacturer | 112 mm | 112 mm | 112 mm | Varies by manufacturer |
| Slot | 3-Slot | 3-Slot | 3-Slot | Varies by manufacturer | Varies by manufacturer | 2-Slot | 2-Slot | 2-Slot | 2-Slot |
| – | Varies by manufacturer | Varies by manufacturer | Varies by manufacturer | Varies by manufacturer | Founders Edition Yes Varies by manufacturer | Founders Edition Yes Varies by manufacturer | – | – | |
| Thermal and Power Specs: | |||||||||
| Maximum GPU Temperature (in C) | 90 | 90 | 90 | 90 | 90 | 90 | 90 | 90 | 90 |
| Total Graphics Power (W) | 450 | 320 | 320 | 285 | 285 | 220 | 200 | 165 or 160 | 115 |
| Required System Power (W) (5) | 850 | 750 | 750 | 700 | 700 | 650 | 650 | 550 | 550 |
| Required Power Connectors | 3x PCIe 8-pin cables (adapter in box) OR 1x 450 W or greater PCIe Gen 5 cable | 3x PCIe 8-pin cables (adapter in box) OR 1x 450 W or greater PCIe Gen 5 cable | 3x PCIe 8-pin cables (adapter in box) OR 1x 450 W or greater PCIe Gen 5 cable | 2x PCIe 8-pin cables (adapter in box) OR 300 W or greater PCIe Gen 5 cable | 2x PCIe 8-pin cables (adapter in box) OR 300 W or greater PCIe Gen 5 cable | 2x PCIe 8-pin cables (adapter in box) OR 300 W or greater PCIe Gen 5 cable. Certain manufacturer models may use 1x PCIe 8-pin cable | 2x PCIe 8-pin cables (adapter in box) OR 300 W or greater PCIe Gen 5 cable Certain manufacturer models may use 1x PCIe 8-pin cable | 1x PCIe 8-pin cables (adapter in box) OR 300 W or greater PCIe Gen 5 cable. Certain manufacturer models may use 1x PCIe 8-pin cable |
3.3 Compare 30 Series Specs
3.4 Compare 20 Series Specs
3.5 Technologies Compares Between Different GPU Series
Generally speaking, the newest is always the best.
| RTX 50 Series | RTX 40 Series | RTX 30 Series | RTX 20 Series | GTX 16 Series | GTX 10 Series | ||
|---|---|---|---|---|---|---|---|
| NVIDIA Architecture | Architecture Name | Blackwell | Ada Lovelace | Ampere | Turing | Turing | Pascal |
| Streaming Multiprocessors | 2x FP32 | 2x FP32 | 2x FP32 | 1x FP32 | 1x FP32 | 1x FP32 | |
| Ray Tracing Cores | Gen 4 | Gen 3 | Gen 2 | Gen 1 | – | – | |
| Tensor Cores (AI) | Gen 5 | Gen 4 | Gen 3 | Gen 2 | – | – | |
| Platform | NVIDIA DLSS | DLSS 4 Super Resolution DLAA Ray Reconstruction Frame Generation Multi Frame Generation |
DLSS 3.5 Super Resolution DLAA Ray Reconstruction Frame Generation |
DLSS 2 Super Resolution DLAA Ray Reconstruction |
DLSS 2 Super Resolution DLAA Ray Reconstruction |
– | – |
| NVIDIA Reflex | Reflex 2 Low Latency Mode Frame Warp (Coming Soon) |
Reflex 2 Low Latency Mode Frame Warp (Coming Soon) |
Reflex 2 Low Latency Mode Frame Warp (Coming Soon) |
Reflex 2 Low Latency Mode Frame Warp (Coming Soon) |
Reflex Low Latency Mode |
Reflex Low Latency Mode |
|
| NVIDIA Broadcast | Yes | Yes | Yes | Yes | – | – | |
| NVIDIA App | Yes | Yes | Yes | Yes | Yes | Yes | |
| Game Ready Drivers | Yes | Yes | Yes | Yes | Yes | Yes | |
| NVIDIA Studio Drivers | Yes | Yes | Yes | Yes | Yes | Yes | |
| NVIDIA ShadowPlay | Yes | Yes | Yes | Yes | Yes | Yes | |
| NVIDIA Highlights | Yes | Yes | Yes | Yes | Yes | Yes | |
| NVIDIA Ansel | Yes | Yes | Yes | Yes | Yes | Yes | |
| NVIDIA Freestyle | Yes | Yes | Yes | Yes | Yes | Yes | |
| VR Ready | Yes | Yes | Yes | Yes | GTX 1650 Super or higher | GTX 1060 or higher | |
| NVIDIA Omniverse | Yes | Yes | Yes | Yes | – | – | |
| RTX Remix | Yes | Yes | RTX 3060 Ti or greater | – | – | – | |
| Additional Features | PCIe | Gen 5 | Gen 4 | Gen 4 | Gen 3 | Gen 3 | Gen 3 |
| NVIDIA Encoder (NVENC) | Gen 9 | Gen 8 | Gen 7 | Gen 7 | Gen 6 | Gen 6 | |
| NVIDIA Decoder (NVDEC) | Gen 6 | Gen 5 | Gen 5 | Gen 4 | Gen 4 | Gen 3 | |
| AV1 Encode | Yes | Yes | – | – | – | – | |
| AV1 Decode | Yes | Yes | Yes | – | – | – | |
| CUDA Capability | 12.0 | 8.9 | 8.6 | 7.5 | 7.5 | 6.1 | |
| DX12 Ultimate | Yes | Yes | Yes | Yes | – | – |
3.6 显卡购买参考
-
- 华硕（ASUS）ATS GeForce RTX 4060 O8G
- 华硕（ASUS）TUF GeForce RTX 4070 SUPER O12G
- 技嘉魔鹰 GeForce RTX 4060 Ti GAMING OC 16G
- 华硕（ASUS）TUF GeForce RTX 4070 Ti SUPER O16G
- 华硕（ASUS）ROG STRIX GeForce RTX 4090 D 24G
- 4090D是为了规避贸易战限制在4090基础上小幅度削弱的版本，专门针对中国市场，如果可以尽量买不带D的版本，两者价格几乎没有区别.
- 华硕ROG STRIX GeForce RTX 4090 O24G GAMING
4. 处理器
-
人工智能和CPU关系不大, i3也能勉强用但不推荐, 建议i5及以上, i7，i9更好。
-
CPU不太关键, 但是部分AI相关应用可以使用CPU的计算能力, 另外更强的CPU肯定体验更好。
4.1 因特尔CPU问题介绍
- 英特尔13代和14代处理器存在性能衰退现象，主要表现为处理器在使用一段时间后性能显著下降，甚至导致系统不稳定、蓝屏等问题。
- 性能最多可能损失达30%，且无法恢复。
- 问题不是一定发生，长时间高负载更容易出问题。日常办公这种低负载情况出问题的概率会降低很多。
- 一般来讲，要么选择英特尔12代CPU，要么选择AMD。
4.2 AMD芯片能否用于人工智能？
- 英特尔由于之前的强势，在个人消费和服务器领域占有率较高，部分软件或者框架都英特尔CPU有些专门的优化，因此可能存在一定优势。
- 但是英特尔13代和14代存在的性能衰退现象，实在是过于抽象。
- 所以在2025年，AMD处理器已经是个很好的选择了。
5. 内存
-
16G起步, 建议双通道32G(16+16), 64G(32+32)更好, 多多益善.
-
内存条是非常容易购买的产品, 和CPU一样在主板内被保护得非常好, 几乎没有损坏的可能性. 购买时注意下区分下ddr4或ddr5, 最大容量等即可.
- 没必要买所谓极客级别的产品, 也没必要为了特挑颗粒高出的一点时序而支付额外成本, 64G的普通内存条远比32GB的顶级内存更加实用.
- 参考
6. 固态硬盘
-
容量越大, 速度越快, 性能就越强.
-
很多应用和模型动辄五六G, 建议2T起步, 最好4T.
- 参考
7. 散热系统
随着显卡，处理器，PCIe 5.0固态硬盘等硬件的性能越来越高，发热量也随之水涨船高。所以散热系统的重要性，也愈发重要。轻则影响硬件性能释放，严重的话可能会损害硬件。
7.1 CPU散热
- 水冷散热器
- 优点
- 理论上水冷散热系统更好，散热效率更好。
- 更节省空间，有些ITX只能使用水冷
- 缺点
- 使用寿命存在不确定性，产品品质、电脑平时的负载情况，以及搬家托运这种情况发生的震动，都可能影响可用年限。两三年，三五年，七八年都有可能。
- 某些特殊情况下，可能会产生漏液的风险，继而可能进一步导致其他硬件受损。
- 优点
- 风冷散热器
- 优点
- 皮实耐用。
- 缺点
- 高端风冷散热器，为了提高性能，提及较大，有些机箱可能无法安装。
- 散热效率不如水冷。
- 优点
我的选择
- 个人偏好风冷散热，买一个顶级性能风冷散热器，可以基本上不会坏，可以使用很多年。
7.2 固态硬盘散热
现在有些高性能固态硬盘，新能很强，但发热非常夸张。现在有专门给固态硬盘散热的产品，如果机箱和主板空间支持安装，建议安装。
7.3 显卡散热
显卡的散热系统由厂家设计和决定，市面上没有为显卡单独额外安装的散热器。所以关于显卡散热，一是了解显卡的散热设计，二是注意机箱的风道等散热设计。
8. 操作系统
-
普通用户尽量windows 10 & windows 11.
-
操作系统不像显卡那样是一个生死问题, 大部分应用也支持Linux, MacOS, 或者使用Docker也行, 但对于普通用户较为麻烦.
9. 其他
9.1 购买整机还是装机
- 尽量自己购买各种配件装机, 使用体验和硬件升级等各方面都比整机好太多。
- 整机厂商作为一个商业公司，商业利润总是第一考虑要素。部分厂家为了榨取利润，可能会牺牲品质。
10. 硬件优先级
显卡 > 内存 >= 硬盘 > CPU > 主板 >= 电源
