Computer Science
Server Computing
How does modern server work from a pc user perspective?
This data is up to date on Jan 2026
Server vs PC
| Feature | Server Hardware | PC Hardware |
|---|---|---|
| Purpose | Resource and application management for multiple users/clients. | Personal productivity, gaming, and single-user tasks. |
| Processor (CPU) | Often multiple CPUs, optimized for parallel/multi-threaded workloads and sustained high-load performance. | Typically a single CPU, optimized for single-threaded speed and low latency in interactive tasks. |
| Memory (RAM) | Uses ECC RAM (Error-Correcting Code) to automatically detect and correct data errors, ensuring stability. | Uses standard RAM, which is faster but lacks error-correction capability. |
| Redundancy | High redundancy is standard: Redundant Array of Independent Disks (RAID) for storage, hot-swappable drives, and redundant power supplies to maximize uptime. | Minimal or no built-in redundancy; failure of a key component usually results in downtime. |
| Form Factor | Designed for mounting in equipment racks (e.g., 1U, 2U) to maximize density in data centers. | Designed for aesthetic and ergonomic use (tower, desktop case). |
| Networking | Fiber optic long range and high speed 100~400 Gb/s using QSFP-DD connector. Using core switches (Cisco Nexus or Arista 7000 series) | RJ45 ~1 Gb/s |
| OS Kernel | Optimized for throughput | Optimized for low latency and responsiveness |
- Computing on storage (CS): reduce the time spend on travelling
- HBM (High Bandwidth Memory): new type of RAM that is stacked vertically with 2+ TB/s data pipes compared to 1.5 TB/s of GDDR7.
CPU
| Tier | Category | Model (2026) | Cores / Threads | Best Use Case | Price (USD) |
|---|---|---|---|---|---|
| Highest (Server) | Extreme Data Center | AMD EPYC 9965 "Turin" | 192 | Cloud Providers, Massive AI | 9000 - 14000 |
| High (Server) | Enterprise Workhorse | Intel Xeon 6980P | 128 | Databases, Financial Modeling | 6,000 – 17,800 |
| High-End (PC) | Prosumer/Workstation | AMD Threadripper 9995WX | 96 | 8K Video Editing, 3D Rendering | 11,699 |
| High-End (PC) | Gaming/Productivity | Intel Core Ultra 9 285K | 24 | High-end Gaming, Streaming | 515 – 599 |
| Moderate (PC Desktop) | Mainstream | AMD Ryzen 7 9800X3D | 8 | Pure Gaming Performance | 469 – 479 |
| Moderate (PC Laptop) | Budget/Office | Intel Core Ultra 5 245K | 14 | General Office, Light Design | 299 – 315 |
GPU
| Year | GPU Model | Architecture | VRAM | CUDA | TFLOPS (FP32) | Price (MSRP) | Primary Use Case |
|---|---|---|---|---|---|---|---|
| 2016 | GTX 1080 | Pascal | 8 GB | 2,560 | 8.9 | $599 | Legacy Gaming |
| 2019 | RTX 2070 Super | Turing | 8 GB | 2,560 | 9.1 | $499 | Mid-range Gaming |
| 2025 | RTX 5090 | Blackwell | 32 GB | 21,760 | ~101 | $1,999 | Extreme Gaming / Local AI |
| 2017 | V100 | Volta | 32 GB | 5,120 | 14.1 | ~$10,000 | 2018 AI Standard |
| 2020 | A100 | Ampere | 80 GB | 6,912 | 19.5 | ~$15,000 | AI Training Workhorse |
| 2022 | L40 | Ada Lovelace | 48 GB | 18,176 | 90.5 | ~$8,000 | Pro Graphics / Rendering |
| 2023 | L40s | Ada Lovelace | 48 GB | 18,176 | 91.6 | ~$10,000 | Universal AI & Graphics |
| 2023 | H100 (SXM5) | Hopper | 80 GB | 16,896 | 67.0 | ~$30,000 | Modern LLM Training |
| 2024 | H200 | Hopper | 141 GB | 16,896 | 67.0 | ~$40,000 | Large-scale AI Inference |
In AI-specific tasks using FP8, an H100 is roughly 30x to 40x faster than a 1080.
Cloud Computing
Virtual Machine
Rent a virtual machine, storage, database, etc.
Serverless
Run code without managing servers. Paying only for the resources used. When an event happens, the cloud provider instantly spins up a tiny container to run the code and shutdown when done.