Mirsee · Silicon Research · July 2026

Weights, etched.
Fixed-weight LLM silicon, 1–10B.

Can a custom ASIC physically embed the weights of one trained language model — fixed at manufacture in ROM or compute-in-memory — and serve inference of only that model? Yes: the digital mask-ROM + on-chip-SRAM route is credible today, and one company has already fabricated it. This is the full verified evidence base.

107research agents
25sources fetched
124claims extracted
22verified 3-vote
3refuted
Verdict Synthesis

Feasible — and already done once.

The viable route is a digital hybrid: model weights baked into mask ROM, paired with on-chip SRAM for the dynamic state (KV cache) and for small updatable adapter weights. It works only if the model is aggressively quantized — ternary to 4-bit — via quantization-aware training, which currently holds accuracy at the 1–4B scale.

For a 1–3B ternary model, fixed-weight silicon looks practical today: a ~57 mm² die at ~5 W with enormous batch-1 speedups. An 8B model is at the edge — it needs 2-bit weights (accuracy unproven at that scale) or a near-reticle-limit die. The analog compute-in-memory alternative is roughly three orders of magnitude short of the needed capacity.

Method

How to read this page

A deep-research workflow decomposed the question into five search angles, fetched 25 sources, extracted 124 falsifiable claims, and adversarially verified the top 25 with three independent votes each. Every tile carries its evidence tier:

  • ✓ Verified 3-vote  survived adversarial refutation
  • Fabricated silicon  real chip, vendor-claimed figures
  • △ Extracted, unverified  sourced, not fact-checked
  • ✗ Refuted  killed by ≥2 of 3 verifiers
Landscape · Existence proof Fabricated silicon△ Vendor-claimed figures

Taalas HC1 — Llama 3.1 8B, hardwired in mask ROM

The single most important finding is a real, fabricated, commercially demoed chip (EE Times, Feb 2026): Taalas' HC1 hardwires the entire Llama 3.1 8B model, weights included, into mask ROM on TSMC N6, storing a 4-bit parameter and performing its multiplication on a single transistor. No HBM, no external DRAM. A public demo runs at chatjimmy.ai.

TSMC N6process node
815 mm²die area
~53 Btransistors
~16–17 ktokens/s per user
~250 Wpower draw
2 layersmasks changed per model
~2 monthsnew-model turnaround
0.75 ¢per M tokens, claimed TCO

Two details answer the hardest design questions. KV cache: the ROM holds only fixed weights; a small on-chip programmable SRAM holds the KV cache and optional LoRA fine-tune weights. Obsolescence: a structured-ASIC-like approach means only two mask layers change to customize the chip for a new model — model-to-RTL in about a week, new silicon in about two months. The company has raised over $200M with ~25 employees and claims ~30 chips could serve DeepSeek R1 671B at ~12,000 tok/s per user.

All performance figures are Taalas' own claims; no independent benchmarks exist yet.

Feasibility · Motivation ✓ Verified 3-0

Why baking in weights wins at batch 1

Autoregressive decode at low batch is memory-bandwidth-bound: fetching weights dominates cost, leaving GPU tensor cores around 28% utilized in decode versus ~92% in prefill. A contrast design (VitaLLM, TSMC 16nm, 0.223 mm²) that streams a 3B ternary model from external DRAM caps at 70.7 tok/s purely on LPDDR bandwidth. On-die weights remove that wall entirely — TOM's aggregate on-chip bandwidth is 200 TB/s, about 41× an H100's HBM3e.

Qualification: the advantage is for batch-1 / low-batch serving. At batch ≥ 32 decode becomes compute-bound; at very long contexts KV traffic dominates instead of weights.

Feasibility · Die-area math ✓ Verified

The density math works at 7nm

⅓×ROM bit-cell area vs 6T SRAM (TSMC 7nm)
15.0MB/mm² — ternary weights as sparsity-aware logic (~70% zeros)
25.3MB/mm² at 95% sparsity
8.4MB/mm² — 3D-stacked DRAM, for comparison

Compiler-generated ROM reaches ~57.8 Mbit/mm² versus ~25.7 for SRAM. Better still, synthesizing ternary weights directly as standard-cell logic — exploiting BitNet's natural zero-density — beats even 3D-stacked DRAM by ~75%. This is the arithmetic that makes billions of on-die parameters plausible.

Synthesis and place-and-route estimates from preprints, not measured silicon.

Feasibility · Design studies ✓ Verified△ Simulation only

Two complete 7nm design studies

Both are synthesis + place-and-route studies at TSMC 7nm — no fabricated chip exists for either — but together they bracket the design space.

DesignTOM (Microsoft Research, 2026)ROMA (SJTU/PKU/MSRA, 2025)
ModelBitNet-2B, ternary4-bit 3B or 2-bit 8B LLaMA
Storage498.5 MB ROM + 37.5 MB SRAM1.86 GB ROM + 304 MB SRAM
Die area56.9 mm²503.7 mm²
Power5.33 W power-gated (25.8 W without)33.1 W
Throughput3,306 tok/s peak; ~64× A100 end-to-end, batch 1>20,000 tok/s short-context; ~70× RTX 4090 avg
Context on-chip1,024 tokens (FP8 KV)up to 4K tokens

Verifier flags: TOM's headline "465×" is a peak corner figure (typical decode gap vs published bitnet.cpp A100 numbers ≈ 13×); ROMA's 20K tok/s falls to ~10K at 4K context; TOM's density multipliers appear transposed in the paper's prose.

Feasibility · Comparison △ Mixed provenance

Batch-1 tokens/s, single user

~2–8B-class models; figures are vendor-claimed or simulated, not independently benchmarked, and model sizes differ — read as order-of-magnitude.

Feasibility · Architecture ✓ Verified 3-0

The two hard problems have answers

Dynamic KV cache. Both design studies hold it in distributed on-chip SRAM, tiled across the context dimension with a flash-decoding-style dataflow. Demonstrated on-chip context is 1–4K tokens — the likely product-defining constraint when users expect 32K–128K windows.

Model obsolescence. ROM base weights are immutable, but ternary QLoRA adapters living in SRAM (reusing the KV-cache arrays) let the chip specialize after manufacture. This fixes task drift, not full base-model obsolescence — Taalas' cheap two-mask-layer respin is the commercial answer to the rest.

Feasibility · Quantization ✓ Verified 3-0

Quantization is the enabler — QAT, not PTQ

BitNet-style ternary (1.58-bit) quantization-aware training maintains competitive accuracy, with the gap versus full precision shrinking as models grow — strongest evidence at 2–4B parameters. LUT-based ternary matmul cores are 2.2× smaller than dequantize-to-FP16 designs. Fabricated analog silicon (NeuRRAM) has validated 4-bit weights-in-memory inference, though only at CNN/LSTM scale.

The refuted-claims tile matters here: generic 2–4-bit post-training quantization being lossless was killed in verification. The feasibility case rests on QAT. Evidence for ternary parity at 7B+ with trillion-token training remains thin.

Landscape · The other route ✓ Verified 3-0

Analog compute-in-memory: real silicon, wrong scale

Analog CIM has actually shipped: IBM's Hermes (14nm + phase-change memory, 17M parameters), NeuRRAM (3M RRAM devices), Mythic's M1076 (weights as analog levels on flash, 40nm). But capacities top out at millions of parameters — about three orders of magnitude short of 1B+.

And NVM can't practically serve attention: dynamic operands must be written into the crossbar, but ReRAM endurance is ~10⁵ cycles versus SRAM's >10¹⁶. Every credible NVM-CIM transformer design routes attention into digital CMOS anyway. Instructively, d-Matrix started analog and pivoted to digital in-memory compute; Mythic ran out of money in 2022 ($140–170M raised) before reaching revenue — reportedly a commercial failure rather than a technical one.

Landscape · Specialization spectrum △ Extracted, unverified

Where fixed weights sit on the spectrum

ApproachWhat's fixedWeights live in
Nvidia GPUNothingHBM (off-die)
Groq / CerebrasDataflow styleOn-chip SRAM, reprogrammable
Etched SohuTransformer architecture144 GB HBM3E (off-die)
d-Matrix CorsairDigital IMC fabric2 GB on-die SRAM + DDR
Taalas HC1Architecture + weightsMask ROM, on-die

Even Etched — the most aggressive transformer ASIC, ~$800M raised — keeps weights in off-die HBM. Architecture-fixed silicon already carries obsolescence risk: Sohu cannot run MoE routing, Mamba/SSM, or vision models. Taalas is the only player past the weights-in-silicon line.

Economics · NRE △ Extracted, unverified

What a tape-out costs

Node classMask setTotal NRE
3–5nm leading edge$10–40M$30–100M+
7–28nm mid-range$1–10M+ (7nm >$10M)$5–30M
28nm mature~$1.5Mdominated by engineering,
IP and verification,
not masks
55/65nm~$0.5M
180nm<$100K

Sources repeatedly stress that verification, not masks, is the most underestimated NRE component — mid-range totals assume 20–50 engineers for 18–24 months. The break-even heuristic from the ASIC-Clouds literature (the "two-for-two rule"): if annual TCO on existing hardware exceeds the ASIC NRE by 2× and the ASIC delivers ≥ 2× TCO-per-op improvement, the ASIC pays. Weight density pushes this design toward advanced nodes — Taalas' 815 mm² N6 die reflects that tension — but their two-mask-layer respin converts "new model = full tape-out" into a cheap increment.

Prototype · Small-team path △ Extracted, unverified

Path to first silicon

  1. FPGA emulation of the ROM + SRAM dataflow — prove the architecture and the ternary compute pipeline before spending on masks.
  2. MPW shuttle test chip proving ROM weight-fabric density and power. EUROPRACTICE 2026: GF 22nm FDSOI ≈ €17.8K/mm², 4 mm² minimum (~€65–71K); GF 12LP+ ~€100–110K; TSMC 7nm MPW runs $100–500K. Academic discounts don't apply to commercial teams.
  3. Production tape-out only with volume commitments — MPW parts prove concepts but are explicitly not a production path.

Warning: the cheapest route died. Efabless shut down in March 2025, taking chipIgnite and stranding Tiny Tapeout shuttles mid-fab (~$100/design is gone). Tiny Tapeout is migrating to IHP 130nm. The ultra-low-cost open-silicon path is fragile and grant-dependent.

Verification · Killed claims ✗ Refuted

Three claims that didn't survive

"~35% of transformer ops can't be weight-stationary at seq-512." Overstated — the dynamic-attention share does not undermine the fixed-weight case the way the source implied. 1–2
"2–4-bit quantization is generically lossless." False for post-training quantization. The feasibility case rests specifically on quantization-aware training (BitNet-style). 1–2
"In-flash computing already runs real-time 7B inference (36.3 tok/s LLaMA2-7B)." Those were simulations on hypothetical compute-enabled NAND, not hardware. 0–3
Open questions

What the evidence couldn't settle

Real unit economics. Actual NRE and minimum viable volume for a ROM-heavy die versus amortized GPU serving — no verified cost data survived.
A second existence proof. Does anyone besides Taalas get true weights-in-ROM silicon to market? TOM/ROMA-class designs could tape out within a year.
Ternary at 7–10B. Can ternary/2-bit QAT reach parity at 7B+ with trillion-token training budgets? ROMA's 2-bit-8B configuration depends on it.
Context scaling. On-chip SRAM KV caches top out at 1–4K tokens today; users expect 32K–128K. This may be the real product-defining constraint — flash-resident KV (KVNAND-style) is the speculative escape hatch.
Sources · Key primary references

Read the primary sources