AI Hardware

PFN Stacks DRAM on Top of Its Next Inference Chip, With Samsung Foundry and SEMIFIVE Building It

November 19, 20261 views阅读中文原文
PFN Stacks DRAM on Top of Its Next Inference Chip, With Samsung Foundry and SEMIFIVE Building It

Anyone running AI inference keeps hitting the same wall: compute is purchasable, bandwidth is not. At SC26 in Chicago, Japan's Preferred Networks (PFN) proposed an unusually literal fix — put the memory directly on top of the chip so the data stops taking the long way around.

On November 16, PFN and Korean custom-silicon house SEMIFIVE announced they are co-developing the MN-Core L2000 series of inference chips. It is the successor to PFN's current MN-Core L1000, and the stated goal is blunt: more than twice the memory bandwidth of L1000, with tape-out scheduled for the first quarter of 2027. The logic die will be built on Samsung Foundry's platform technology.

TL;DR: the short version

  • This is not a chip you can buy. The L2000 is at "collaboration under way, architecture defined, tape-out in Q1 2027."
  • It attacks memory bandwidth, not raw compute. The official target is more than double the bandwidth of L1000, to speed up inference and improve energy efficiency.
  • The approach skips the side-by-side HBM package entirely: DRAM is stacked vertically on top of the logic die, with the two connected across their full surface.
  • The division of labor is clean. PFN owns architecture and logic design, SEMIFIVE owns design implementation and the supply chain, Samsung Foundry owns the manufacturing platform and 3D-IC design infrastructure.
  • "Twice the bandwidth" is a plan, not a measurement. No pricing, power figures, or customer names have been disclosed.
  • One thing worth keeping straight: the first real product in this family is L1000, slated for 2027. L2000 sits on the official roadmap at 2028.

Why bandwidth, not compute, is the bottleneck

LLM inference runs in two stages. The first is prefill: the model reads your prompt, documents, and code in one pass and computes intermediate state. That stage is compute-hungry. The second is decode: the model emits one token at a time. That stage is memory-bandwidth hungry — every new token requires re-reading the model weights.

The bigger the weights, the more data moves per token. A 100-billion-parameter model carries hundreds of gigabytes of weights, and generating a single word means sweeping through all of it. Compute can be abundant and still sit idle waiting for data. This is the memory wall.

GPUs answer it by placing HBM right next to the die in a 2.5D package. PFN pushes further: if data has to move, shorten the distance to the minimum. Stack logic and DRAM vertically and interconnect them across the entire die surface, rather than routing through an interposer at the edge.

How the stack actually works

PFN's public description of the MN-Core L series is straightforward: logic chips built on the MN-Core architecture, vertically stacked with DRAM in a 3D structure. The company's argument is that this yields substantially higher memory bandwidth than GPUs and other AI processors, while improving power efficiency — aimed at latency-sensitive workloads such as AI agents.

Attractive on paper, hard in practice. 3D stacking has to solve three problems at once: getting power in, getting heat out, and holding yield — one bad die can scrap the whole stack. PFN CEO Daisuke Okanohara did not dodge this in the announcement, saying explicitly that 3D-stacked structures are harder than conventional chip development in power consumption, heat dissipation, and manufacturing yield.

The official product page at mn-core.com currently lists L1000 basics as 128GB of memory, 40+TB/s of bandwidth, and 10x inference performance in the same system footprint, noted as "one workstation in 2027." The page also flags these as preliminary specifications subject to change. A caution for readers: different outlets have quoted different L1000 sub-figures, for instance per-chip bandwidth and capacity figures that do not line up with the official system-level numbers. Confirm whether a number refers to a single chip or a whole machine before comparing.

Who does what

  • Preferred Networks (Japan, chip definition): processor architecture and logic design. MN-Core has been in development since 2016.
  • SEMIFIVE (Korea, KOSDAQ: 490470, custom-silicon turnkey): design implementation and the full supply chain — DRAM supply, volume manufacturing, advanced packaging, and test.
  • Samsung Foundry (manufacturing and design infrastructure): platform technology for the logic die, plus 3D-IC design infrastructure including I/O interfaces for 3D die connectivity and power-structure optimization methodology.

SEMIFIVE CEO Brandon Cho called the project his company's "first full-scale 3D-IC initiative." Samsung's Taejoong Song, vice president of Foundry Technology Planning, framed it around ecosystem: expanding 3D-IC through the SAFE program.

There is another layer worth knowing. The project is supported by Japan's Acquisition, Technology & Logistics Agency (ATLA) under its security technology initiative, through a project titled "Research on Ultra-High-Speed Deep Learning and Inference Using 3D-Integrated LSI Technology" (Type S, FY2024–FY2028). In other words, this carries both industrial and national-compute-sovereignty weight.

Timeline: don't mistake a roadmap for reality

  • 2016: PFN begins MN-Core development (background).
  • November 2024: L1000 first announced publicly (old news, not this story).
  • November 16–17, 2026: L2000 collaboration and tape-out plan announced at SC26 (this announcement).
  • Q1 2027: L2000 tape-out planned (has not happened).
  • 2027: L1000 launch planned, in single-chip, four-chip box, and 16-chip workstation configurations (has not happened).
  • 2028: L2000 introduction planned, per the official roadmap (has not happened).
  • 2030: Next generation, per the official roadmap (has not happened).

One easily confused item: on November 18, Korea's MangoBoost announced an MOU with PFN to run a proof of concept for an "AI Token Factory" heterogeneous inference architecture — GPUs handling prefill, MN-Core handling decode, DPUs managing KV cache and data movement. That is a separate announcement from the L2000 chip development. Do not merge the two.

How to use this: if you are evaluating the approach

The next step for most teams is a desk exercise, not a purchase order. Start here.

Step one: be clear about what is for sale. L1000 is a 2027 product; L2000 is later. What you can do today is not procurement — it is deciding whether this architecture belongs in your three-year plan.

Step two: map your workload shape. Long-input, short-output work (document Q&A, code completion, retrieval-then-summarize) stresses prefill. Long-output work (full reports, extended dialogue) stresses decode bandwidth. 3D-stacked DRAM targets the latter.

Step three: check the software stack. This is not CUDA. Models have to pass through PFN's own compile and runtime environment. Historically, migration cost on custom architectures decides outcomes more often than hardware specs do.

Step four: watch two verifiable milestones — whether L1000 ships on schedule in 2027, and whether L2000 tapes out on schedule in Q1 2027. If either slips once, replan everything downstream.

Cost and timing

No pricing, power figures, or customer list have been published. The only firm anchors are three dates: L2000 tape-out in Q1 2027, L1000 launch in 2027, and L2000 introduction in 2028 per the official roadmap. SEMIFIVE mentions Samsung's platform technology and 3D-IC design infrastructure but does not name a process node.

Commercially, the real benchmark will be the HBM4-based systems on the market in 2027. 3D-stacked DRAM wins on bandwidth density and energy efficiency; it is exposed on yield, repairability, and capacity flexibility. Whether DRAM makers can supply enough stacking capacity is the precondition for this path to scale.

Who it is for

  • Operators building their own inference clusters and constrained by power and floor space: higher bandwidth density means more concurrent requests per rack.
  • Teams building agents and real-time interactive applications: these workloads need sustained low-latency token generation, which is exactly bandwidth-sensitive.
  • Institutions and policymakers focused on compute sovereignty: Japan's use of an ATLA program to back a domestic AI chip is a useful case study.
  • Chip and packaging engineers: this is a concrete example of 3D-IC moving from lab to volume process flow — SEMIFIVE describes it as its first full-scale 3D-IC project.

Not a good fit

  • Anyone who wants to buy an accelerator now: tape-out is Q1 2027, and there is no product today.
  • Teams deeply invested in the CUDA ecosystem: switching architectures means rewriting and retuning, which rarely pays off short term.
  • Production environments that need mature operations tooling: monitoring, fault isolation, and degradation strategies for this architecture are not public.
  • People chasing peak single-card FLOPS: this route sells bandwidth and efficiency, not TFLOPS bragging rights.

Pitfalls: five things easy to misread

  • "Twice the bandwidth" is a target, not a result. The company uses planning language and has published no benchmark data. Unite.AI made the same point in its independent coverage.
  • L2000 is not arriving next quarter. Tape-out is Q1 2027, and the official roadmap puts introduction in 2028. Tape-out to volume production typically takes a year or more.
  • Do not treat L1000 specs as L2000 specs. The only L2000 number disclosed is the relative "more than 2x L1000 bandwidth." There are no absolute figures at all.
  • Do not conflate 3D stacking with HBM. HBM is side-by-side plus an interposer. This is vertical stacking, with a completely different interconnect area and thermal path — and different engineering risk.
  • Do not read MangoBoost's MOU as L2000 progress. That is system-level validation of heterogeneous inference on L1000 samples, a separate track from L2000 silicon development.

Bottom line

The real signal here is not "twice the bandwidth." It is that Japan has handed the volume-production chain for a homegrown inference chip to a Korean design house and Samsung Foundry, while committing to 3D-stacked DRAM as a bandwidth-first bet that diverges from Nvidia's path. The test for readers is simple: if inference bandwidth costs and power bills keep you up at night, put this on your watch list. If you just want the machine on your desk to be faster, it has nothing to do with you.

Note: This article is compiled from the official PFN and SEMIFIVE press releases and from multiple media reports. Every date that has not yet occurred is marked as a plan.

Sources

Comments

0 published
Log inorsign up to leave a comment.

Loading comments...