AI Accelerators for LLM Inference: Architecture Analysis and Scaling Strategies
Author: Amit Sharma (IEEE Member) | arXiv: 2506.00008 (2025) | PDF: AI_Accelerators_LLM_Inference_2025.pdf
一句话总结
第一篇真正定量横评 GPU / TPU / LPU / WSE / RDU 跑 LLM inference 的论文。结论:无单一架构统治所有 workload,batch × seq 维度上架构间最大 3.7× 性能差;软件>硬件(同硬件不同栈达 40% 性能差)。
五类架构 + 关键数字
| 架构 | 平台 | 工艺 | 关键数字 |
|---|---|---|---|
| GPU SIMD/SIMT | NVIDIA Blackwell GB200 | TSMC 4N | 192GB HBM3e、8 TB/s、4500 TFLOPS FP16(s)、1.8 TB/s NVLink5、1000W TDP |
| GPU SIMD/SIMT | AMD MI300X | TSMC 5nm | 192GB HBM3、5.3 TB/s、256MB L3 Infinity Cache |
| Systolic | Google TPU v7 | 5nm | 192GB HBM3、7.37 TB/s、~100B+ transistors |
| Many-core SRAM-centric | Graphcore IPU | - | 900MB on-chip SRAM、45 TB/s internal |
| Many-core SRAM-centric | Meta MTIA v2 | - | 256MB shared SRAM (RISC-V + vector) + LPDDR5 |
| Wafer-scale | Cerebras WSE-3 | 5nm | ~900K core、44GB SRAM、SwarmX fabric |
| Deterministic pipeline | Groq LPU v1 | - | 230MB on-chip SRAM、80 TB/s internal、亚毫秒 latency |
内存层次三分类
- HBM-focused: Blackwell / TPU v7 / MI300X — 192GB HBM
- On-chip memory focused: WSE-3 / IPU — 44GB / 900MB SRAM
- Hybrid: MTIA v2 / Inferentia2 — 256MB SRAM + LPDDR5/HBM2e
四个 Scaling 策略(trillion-param 模型)
| 策略 | 参数-计算比 | 延迟方差 |
|---|---|---|
| Tensor parallelism | baseline | 1.0× |
| Pipeline parallelism | 1.5× | 1.3× |
| Expert parallelism (MoE/CoE) | 8.4× | 2.1× |
| Hybrid | 3-5× | 1.5× |
关键洞察:Expert parallelism 参数容量 8.4× 是其他策略 2-5×,但延迟方差 2.1× —— 对 agentic/interactive LLM 场景(需要稳定 per-token latency)是隐患。
Architecture-workload 匹配
- 长 context / 单请求低延迟 → Groq / WSE(deterministic + on-chip SRAM)
- 大 batch / 高吞吐 → Blackwell / TPU(HBM 大容量 + 高 inter-GPU 带宽)
- 大模型 + 中 batch → WSE(weight streaming) + Hybrid
- 稀疏激活(MoE/CoE) → SambaNova RDU / Groq(数据流 + 多档内存)
关键结论
“Performance variations of up to 40% were observed for the same hardware with different software stacks.”
→ 编译器 / runtime 优化空间 ≥ 硬件设计空间。这是 Direction 2(compiler-aware decode)选择的最强论据。
“no single architecture dominates across all workload categories, with performance variations of up to 3.7× between architectures depending on batch size and sequence length.”
→ 没有”最优架构”,只有”匹配 workload 的架构”。这给”通用 LLM 推理编译器”留空间。
与 wiki 已有内容关联
- Cerebras WSE — WSE-3 数据
- Nvidia Groq 3 Lpx — LPU 数据(v1 vs Groq 3 Lpx)
- WaferLLM System — 系统级 evaluation
- Prefill-Decode Resource Divergence — 阶段差异
- Deterministic Execution — Groq/WSE 共同范式
- Cerebras WSE vs Groq Network Comparison — 双架构横评
Citations
[1] AI_Accelerators_LLM_Inference_2025.pdf — Sharma (2025)