HBM

AI Memory Stack Economics

Stacked DRAM vs Commodity NAND Offload Simulator
Throughput / GPU 191.4 tok/s Bandwidth Limited
HBM Utilization 37.9% Capacity: 640 GB Total
NAND Stall Penalty 0.0 ms No Storage Spill
Total System Memory BOM $15,200 HBM $12.8k | NAND $2.4k
2D Silicon Interposer & Multi-Tier Memory Topology
Stacked HBM3e (DRAM) PCIe NVMe (NAND Flash)
HBM Bandwidth-Bound System: Stacked DRAM interposer bandwidth dominates token speed; NAND bulk storage offload introduces 100x IO stall penalty if weights spill over HBM boundary.
All weights and KV cache fit cleanly within primary HBM3e stacks. Interposer TSVs operate at optimal bus throughput.
Economic Breakdown: Non-Commodity Stacked DRAM vs Commodity Bulk NAND
"Storage is Commodity, HBM is Not. Be careful choosing between NAND capacity vs DRAM... particularly when stacked as HBM." — Jim Liu (@jiahanjimliu)
Structural Economic Dynamics:
HBM (Stacked DRAM): Requires advanced 2.5D/3D silicon interposers & Through-Silicon Vias (TSVs). Yield bottlenecks limit supply, maintaining high gross margins (~65%).
NAND Flash: Standardized multi-layer planar/3D die stacking without complex silicon interposers. Highly elastic supply leads to rapid commodity pricing cycles.
Workload & Model Specs
Model Parameters 70 B
Quantization 16-bit (FP16)
Context Window Length 8,192 tok
Batch Size 32
Hardware Stack Configuration
GPU Count 8 GPUs
HBM Capacity / GPU 80 GB
HBM Bandwidth / GPU 3.35 TB/s
NAND Flash Read Bandwidth 14.5 GB/s
Actions & Export
true 37.88 191.40 12.80 140.00 242.40 12800 2400 HBM Bandwidth-Bound System