June 28, 2026
Chip or fever dream?
Sophon PFG-1: a monolithic-3D AI ASIC with 330 GB of on-die DRAM and no HBM
A giant AI chip claims to ditch pricey memory—and the internet is yelling “real or sci-fi?”
TLDR: Sophon PFG-1 claims it can put a huge amount of memory on one AI chip and avoid the expensive memory parts rivals rely on. Commenters are torn between calling it a bold breakthrough and side-eyeing it as a too-good-to-be-true paper fantasy.
A new AI chip called Sophon PFG-1 just dropped a truly chaotic promise: cram 330GB of memory directly onto the chip, skip the usual ultra-expensive memory stacks, and handle both AI training and answering prompts on the same piece of silicon. In plain English, the pitch is: one monster chip, less waiting around on memory, way cheaper than today’s premium AI hardware. And yes, the numbers are eye-popping enough to make even seasoned chip watchers do a double take.
That’s exactly what happened in the comments, where the vibe split into two camps: “this is genius” and “this is fan fiction with a BOM spreadsheet.” One user called the design “absolutely wild” and basically summed up the thread’s energy with, it probably won’t work, but wow, I respect the ambition. Another immediately asked the question lurking behind all futuristic hardware announcements: has this thing actually been tested, or is it still a beautiful paper dream?
Then came the comedy. Someone wondered if this is the start of a future where even your phone or laptop maker puts memory straight on the chip. Another went full scorched-earth with, “What is this? AI generated company?” Ouch. That line alone captures the suspicion: the claims sound so huge that people aren’t sure whether they’re witnessing a breakthrough, a moonshot, or the semiconductor version of a movie trailer with no release date. Either way, the community has spoken: the specs are impressive, but the real entertainment is watching everyone argue over whether Sophon is the future or just extremely expensive sci-fi cosplay.
Key Points
- •The article presents PFG-1 “Sophon” as a monolithic-3D AI ASIC with 330 GB of on-die 2T0C 2D-TMD DRAM that removes the need for external HBM.
- •Sophon is described as a unified chip for both training and inference, using pure digital compute-in-memory across 131,072 tiles to deliver 2,100 TFLOPS BF16 and 4,200 TFLOPS FP8.
- •The design is said to use a 28 nm silicon CMOS base tier, a 32-tier 2D-TMD CMOS MAC stack, and monolithic inter-tier vias with DRAM embedded in BEOL Metal-3 layers.
- •For an 80B-parameter model, the article reports 2,406 tokens/s training, 7,219 tokens/s BF16 decode, 14,438 tokens/s FP8 decode, and 72,188 effective tokens/s with INT4 plus speculative decoding.
- •The article compares Sophon with NVIDIA Rubin (R200) and AMD Instinct MI455X, claiming higher low-batch throughput, much greater effective weight bandwidth, and lower hardware BOM due to eliminating HBM.