从 NVIDIA GT200 到 Fermi架构 标志着 第三代GPU计算。此前的架构以图形处理为核心，经过“改造”用于数学计算；而Fermi则是从零开始为 通用GPU（GPGPU） 应用而设计的。
1. 从图形优先到计算优先
与专注于纹理单元和严格数据并行性的GT200不同，Fermi引入了统一的内存请求路径。这一转变开启了 计算思维，使开发者能够突破简单的二维网格映射，转向复杂的C++算法开发。
2. 内存层次结构的飞跃
Fermi引入了真正的 L1/L2缓存层级结构 并符合 IEEE 754-2008 浮点标准。这意味着研究人员不再需要为每个字节手动管理“临时存储”内存（共享内存），从而能够支持不规则的数据结构，并实现适合科学工程领域的双精度计算精度。
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.