Search AI Chips News Today

Manufacturing

US-China Export Controls on AI Chips: BIS TPP Thresholds, HBM Rules & Foundry Compliance

How US export controls on China AI chips enforce Total Processing Performance (TPP), performance density, HBM memory stack rules and foundry KYC verification.

AMIS. Photo: Tvien (CC BY-SA 3.0, Wikimedia Commons).

Tracking US export controls on AI chips news today requires looking past diplomatic headlines to the exact silicon engineering metrics enforced by the U.S. Department of Commerce’s Bureau of Industry and Security (BIS). Under Export Control Classification Numbers (ECCN) 3A090 and 4A090, regulatory compliance for data centre accelerators shipped to China and other Country Group D:1/D:4/D:5 destinations is determined not by marketing model names, but by two physical hardware formulas: Total Processing Performance (TPP) and performance density, alongside direct export restrictions on high-bandwidth memory (HBM) stacks and advanced 300mm foundry tape-outs.

How Total Processing Performance (TPP) and performance density gate AI silicon

When initial export controls focused primarily on raw compute combined with bidirectional interconnect bandwidth, chip designers were able to preserve compute FLOPs while throttling NVLink or SerDes I/O lanes to create compliant variants (such as the A800 and H800). Subsequent BIS revisions replaced interconnect ceilings with a two-axis boundary based on Total Processing Performance (TPP) and performance density:

  • Total Processing Performance (TPP) definition: Calculated as 2 × MacTOPS × bit length aggregated across all processing units on the package, using the highest figure achieved across FP16, BF16, FP8, INT8 or INT4 numerical formats. Under ECCN 3A090.a, non-primitive integrated circuits achieving a TPP of 4,800 or more (or 1,600 or more with a performance density above threshold) are controlled.
  • Performance density ceiling: Calculated by dividing TPP by the applicable die area measured in square millimetres (TPP / mm²). This rule specifically blocks chiplet workarounds where a designer might split a monolithic GPU into smaller dies or stitch multiple sub-threshold chiplets onto a single CoWoS interposer to evade the per-chip compute cap.
Regulatory Parameter Engineering Metric Architectural Impact on AI Chips
ECCN 3A090 Total Processing Performance (TPP) 2 × MacTOPS × bit width (e.g., 4,800 TPP threshold) Caps total dense/sparse tensor-core throughput per packaged accelerator shipped without a license.
Performance Density TPP ÷ die area (mm²) Prevents splitting high-FLOP compute engines across multi-die chiplets or dense 3nm/5nm logic dies.
HBM Memory Bandwidth Density Memory bandwidth density (GB/s per mm²) on stacked DRAM Restricts standalone exports of advanced HBM3, HBM3E and HBM4 stacks used in custom AI ASICs.
Foundry Logic & Packaging KYC FinFET/GAA nodes (≤16/14nm) + 2.5D CoWoS/interposer transistor counts Requires foundries to verify transistor counts (≥30 billion) and HBM pad layouts before mask tape-out.

HBM memory restrictions and advanced packaging verification

For readers following export controls on AI chips and China news today, memory supply has become just as decisive as logic lithography. Modern large language model (LLM) inference is bottlenecked by memory bandwidth during autoregressive token decoding. Even if a domestic China AI chip design uses multiple lower-density compute dies, training or serving frontier models requires High Bandwidth Memory (HBM).

Recent BIS foreign direct product rules extend export licensing requirements to standalone HBM stacks exceeding specified memory bandwidth density thresholds. At the same time, leading-edge contract foundries — led by TSMC, Samsung Foundry and Intel Foundry — operate mandatory Know-Your-Customer (KYC) and physical layout verification checks on sub-7nm FinFET and Gate-All-Around (GAA) tape-outs. Any customer design incorporating CoWoS-class 2.5D interposer bond pads, HBM PHY interfaces, or package transistor counts above 30 billion triggers automated compliance screening before photomask generation.

How China AI chip architectures and cluster operators are adapting

Within China’s domestic semiconductor and cloud ecosystem, engineering teams have responded across three hardware layers:

  1. Scale-out optical networking over larger node counts: Because individual compliant accelerators carry lower per-package FP8/BF16 compute, data centre operators deploy wider scale-out clusters using 400G and 800G optical transceivers, trading higher cluster power consumption for aggregate throughput.
  2. Aggressive quantization and MoE sparsity: Running Mixture-of-Experts (MoE) architectures and 4-bit/8-bit weight quantization reduces active parameter memory traffic per token, extracting higher utilization from memory-constrained accelerators.
  3. Domestic DUV multi-patterning and packaging investment: Local fabs rely on immersion DUV (193i) self-aligned quadruple patterning (SAQP) and domestic 2.5D silicon interposer assembly while domestic HBM2E/HBM3 memory lines work to improve stack yields and thermal reliability.

Huawei AI chips and NVIDIA China-compliant accelerators under BIS rules

In Huawei AI chips news today and NVIDIA China AI chips news today, the export-control boundary has split the China data centre market into two competing tracks: compliant merchant accelerators (such as NVIDIA’s H20 and Blackwell-derived export-compliant configurations) that remain strictly below BIS Total Processing Performance (TPP) thresholds while preserving high memory bandwidth for LLM decoding, versus domestic accelerators led by Huawei’s Ascend 910B and dual-die Ascend 910C packages connected via high-radix CloudMatrix optical scale-out racks.

Practical takeaways for global hardware buyers and supply-chain planners

  • Audit full-system BOM compliance: Multinational enterprises deploying AI inference nodes across global regions must verify both the accelerator ECCN classification and the server’s aggregate NAC/ACA notification status before cross-border shipment.
  • Expect bifurcated SKU roadmaps to persist: Global hyperscalers standardize on unconstrained Blackwell, Vera Rubin, Instinct MI355X and custom HBM3E/HBM4 ASICs, while export-compliant SKUs remain bound by strict TPP and die-density ceilings.

For related foundry and packaging analysis, explore our Manufacturing coverage hub, read our breakdown of TSMC CoWoS capacity and HBM stack roadmaps, or consult the AI Chip Buyer’s Guide.