Phanta Field unveiled Sophon PFG-1, a monolithic-3D AI accelerator that ditches High-Bandwidth Memory entirely. The chip integrates 330 GB of on-die DRAM using 2T0C 2D-TMD gain-cell technology, enabling both training and inference workloads on a single die. This represents a fundamental shift away from traditional HBM-based architectures that have dominated AI accelerators.

Built on a 750 mm² platform with 32-tier 2D-TMD Monolithic 3D stacking, Sophon delivers 2,100 TFLOPS BF16 and 4,200 TFLOPS FP8 performance. The pure digital Compute-In-Memory architecture pairs 256×256 DRAM subarray tiles with binary sense amplifiers and 8-level adder trees. Each tile operates at 500 MHz with bit-serial activation broadcast, achieving 0.620 pJ/MAC for BF16 forward passes.

Against 2026 competitors like NVIDIA Rubin and AMD Instinct MI455X, Sophon achieves 191-214× higher weight bandwidth by eliminating HBM bottlenecks. The chip processes 7,219 tokens/s for 80B models in BF16 decode mode, consuming 373 W. This translates to 38.7 tokens/s per watt—approximately 174× better efficiency than HBM4-bound alternatives.

Morgan Stanley estimates NVIDIA VR200 racks at $7.8M with $2.0M in HBM costs alone. Sophon's Bill of Materials drops to roughly $8,358, representing an 11.6× cost reduction versus competing solutions. The architecture fundamentally rethinks AI accelerator design by moving memory on-die rather than relying on expensive, bandwidth-constrained HBM packages.