從 NVIDIA GT200 到 Fermi 架構 標誌著 第三代 GPU 計算。雖然先前的架構是以圖形為主的單元「改造」用於數學運算，但 Fermi 是從零開始專為 GPGPU（通用型 GPU） 應用而設計。
1. 從圖形優先到計算優先
與只注重紋理單元和僵化資料平行化的 GT200 不同，Fermi 引入了統一的記憶體請求路徑。此轉變啟用了 計算思維，讓開發者得以超越簡單的二維網格映射，邁向複雜的 C++ 算法。
2. 記憶體階層的躍進
Fermi 引入了真正的 L1/L2 快取階層 以及符合 IEEE 754-2008 浮點運算標準。這代表研究人員不再需要為每個字節手動管理「暫存記憶體」（共享記憶體），從而能支援不規則的資料結構，並提供適合科學工程應用的雙精準度準確性。
main.py
TERMINALbash — 80x24
> Ready. Click "Run" to execute.
>
QUESTION 1
Which architecture is considered the true start of the 'Third Generation' of GPU computing?
GT200 (Tesla)
Fermi
G80
Fixed-function Pipeline
✅ Correct!
Fermi was the first architecture designed from the ground up for general-purpose compute rather than just adapting graphics units.❌ Incorrect
GT200 was still primarily a graphics-focused evolution; Fermi introduced the unified cache and compute-first SM design.QUESTION 2
What memory feature was introduced in Fermi to help handle irregular data patterns?
Manual Scratchpad only
Hardware-managed L1/L2 Cache Hierarchy
Write-only Texture Buffers
Disabling Global Memory
✅ Correct!
Fermi introduced a unified memory request path with dedicated L1 and L2 caches, drastically simplifying programming for non-grid data.❌ Incorrect
Previous generations relied almost entirely on manual Shared Memory management for performance.QUESTION 3
Fermi's compliance with IEEE 754-2008 was critical for which application type?
Simple 2D Sprite Rendering
High-precision Scientific Computing (FP64)
Text Scrolling
Basic Vertex Shading
✅ Correct!
Compliance with the 2008 standard ensured that GPUs could be used for rigorous engineering and physics simulations requiring high precision.❌ Incorrect
Graphics can often tolerate lower precision; scientific simulations cannot.QUESTION 4
What does 'Computational Thinking' refer to in the context of the Fermi shift?
Treating the GPU as a fixed-function signal processor.
Focusing on the physics of the problem rather than manual data movement.
Manually coding assembly for every pixel.
Using only 2D textures for storage.
✅ Correct!
It refers to the ability to use standard programming patterns (like C++ algorithms) because the hardware now handles the low-level data orchestration.❌ Incorrect
That was the limitation of earlier 'Graphics-First' generations.QUESTION 5
How did Fermi improve thread management?
It removed the concept of Warps.
It introduced sophisticated hardware thread scheduling.
It limited threads to only 32 per GPU.
It forced all threads to run the same instruction forever.
✅ Correct!
Fermi's redesign of the SM allowed for much faster context switching and better scheduling of massive thread counts.❌ Incorrect
Warps remained, but their management became significantly more efficient and flexible.Case Study: The Seismic Researcher's Dilemma
Architectural Transition Analysis
A researcher is porting a seismic imaging algorithm from a GT200-based cluster to a Fermi-based system. The algorithm uses irregular tree-based data structures that do not fit a 2D grid and requires high precision to avoid cumulative rounding errors.
Q
1. Why was the GT200 architecture difficult for this specific researcher's irregular tree data?
Solution:
GT200 lacked a hardware-managed cache. The researcher had to manually partition the tree into the 16KB Shared Memory (scratchpad) for every compute block, which is extremely difficult for irregular, non-coalesced access patterns found in trees.
GT200 lacked a hardware-managed cache. The researcher had to manually partition the tree into the 16KB Shared Memory (scratchpad) for every compute block, which is extremely difficult for irregular, non-coalesced access patterns found in trees.
Q
2. How does the 'Third Generation' (Fermi) architecture alleviate the manual memory burden?
Solution:
Fermi introduced a unified L1/L2 cache hierarchy. The hardware automatically caches frequently accessed nodes of the tree, allowing the researcher to use standard C++ pointers and logic without manually orchestrating every data move to Shared Memory.
Fermi introduced a unified L1/L2 cache hierarchy. The hardware automatically caches frequently accessed nodes of the tree, allowing the researcher to use standard C++ pointers and logic without manually orchestrating every data move to Shared Memory.
Q
3. Which hardware standard change ensures the researcher's results are scientifically valid?
Solution:
Fermi's adherence to the IEEE 754-2008 floating-point standard provided significantly faster and more accurate double-precision (FP64) performance compared to the GT200, which was primarily optimized for single-precision graphics.
Fermi's adherence to the IEEE 754-2008 floating-point standard provided significantly faster and more accurate double-precision (FP64) performance compared to the GT200, which was primarily optimized for single-precision graphics.