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/SIMTNVIDIA Blackwell GB200TSMC 4N192GB HBM3e、8 TB/s、4500 TFLOPS FP16(s)、1.8 TB/s NVLink5、1000W TDP
GPU SIMD/SIMTAMD MI300XTSMC 5nm192GB HBM3、5.3 TB/s、256MB L3 Infinity Cache
SystolicGoogle TPU v75nm192GB HBM3、7.37 TB/s、~100B+ transistors
Many-core SRAM-centricGraphcore IPU-900MB on-chip SRAM、45 TB/s internal
Many-core SRAM-centricMeta MTIA v2-256MB shared SRAM (RISC-V + vector) + LPDDR5
Wafer-scaleCerebras WSE-35nm~900K core、44GB SRAM、SwarmX fabric
Deterministic pipelineGroq LPU v1-230MB on-chip SRAM80 TB/s internal、亚毫秒 latency

内存层次三分类

  1. HBM-focused: Blackwell / TPU v7 / MI300X — 192GB HBM
  2. On-chip memory focused: WSE-3 / IPU — 44GB / 900MB SRAM
  3. 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 parallelismbaseline1.0×
Pipeline parallelism1.5×1.3×
Expert parallelism (MoE/CoE)8.4×2.1×
Hybrid3-5×1.5×

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

Citations

[1] AI_Accelerators_LLM_Inference_2025.pdf — Sharma (2025)