All essays
RSS
FIELD NOTE / INFRASTRUCTURE · NO. 014 REV 2026‑07‑21
UNVERIFIED — GATES OPEN

Stationary Experts: rethinking trillion‑parameter inference on hardware you can actually afford

A look inside ExpertMesh, an open‑source project that trades a proprietary 64‑GPU scale‑up domain for refurbished AMD cards, a fixed home for every expert, and a data plane honest enough to say what it hasn't proven yet.

Repo: github.com/manishklach/expertmesh License: Apache‑2.0 Stack: Python, HIP, ROCm
The premise

A different question about where the weights should live

Serving a trillion‑parameter sparse mixture‑of‑experts model usually means one of two expensive answers: buy an enormous HBM supernode, or stream terabytes of weights across CXL fast enough to keep GPUs fed. ExpertMesh asks a third question — what if each quantized expert just stayed put on one inexpensive card, permanently, and the network only had to carry the small activations routed to it?

That single reframing is the whole architecture. Experts are stationary. The network moves tokens, not tensors. And the hardware underneath is deliberately unglamorous: refurbished AMD Instinct MI210 cards and used EPYC servers, not a vendor's flagship scale‑up rack.

How it moves

Attention, route, dispatch, reduce

Every decode step follows the same rail. A continuous‑batching scheduler feeds attention and a top‑k router, which hands routed activations to a dispatcher. That dispatcher checks who owns each expert: local traffic rides the in‑node Infinity Fabric hive, remote traffic goes out over RoCE. Either way it lands on a stationary, packed‑INT4 expert, and the results come back through a weighted top‑k reduction.

Request
Router (top‑k)
Dispatch by owner
Stationary INT4 expert
Weighted reduce

Local xGMI is preferred over the remote RoCE fabric wherever routing allows it — the scheduling layer is explicitly topology‑aware, and hot experts get replicated only after routing traces demonstrate real skew, not by default.

The reference cluster

What "low‑cost" means in practice

The Phase 1 reference system is a concrete, priced bill of materials, not a hand‑wave. Eight refurbished four‑GPU EPYC nodes, thirty‑two MI210 accelerators, dual‑port 100/200 GbE RoCE per node.

PHASE 1 REFERENCE — MI210 CLUSTER DATASHEET
Compute nodes8 × refurbished 4‑GPU EPYC systems
Accelerators32 × AMD Instinct MI210, 64 GB
Aggregate HBM2 TiB
Aggregate HBM bandwidth51.2 TB/s theoretical
Inter‑node fabricDual 100/200 GbE RoCE per node
GPU board power9.6 kW
Active‑weight planning range30–50 GB/token
Estimated acquisition cost$135k – $235k
Bandwidth roofline1,024 – 1,707 aggregate token/s
That roofline is an ideal bandwidth ceiling — it excludes communication overhead, routing imbalance, attention work, KV traffic, dequantization, kernel inefficiency, and synchronization. It is not a benchmark, and the repo says so in plain text.

The example sizing target is Kimi K3 — roughly 2.8T parameters, 896 experts, top‑16 routing, packed 4‑bit weights — used as a demanding stress case for the planner, not a claim of compatibility. That distinction matters more than it sounds: this codebase is built to be told which model it's sizing for, rather than assuming one.

Where it stands

What's actually proven vs. what's still a plan

The most unusual thing about ExpertMesh isn't the architecture — plenty of MoE serving projects propose expert placement schemes. It's that the project separates, in public, what's been verified from what's still an assumption, and refuses to let one borrow credibility from the other.

✓ Runs today, no GPU needed
  • 18 passing CPU unit/integration tests
  • Bit‑exact NumPy quantizer, dequantizer, routed expert, and MoE reference
  • Deterministic 896‑expert placement across 32 owners
  • expertmesh-plan and expertmesh-roofline CLIs
  • Full local coordinator/worker chat‑completions demo
△ Not yet verified
  • HIP kernel correctness on real MI210 silicon
  • Any multi‑node RoCE performance number
  • Compatibility with a real Kimi K2/K3 checkpoint
  • Whether the roofline survives real routing skew
  • Acquisition cost at today's used‑hardware prices
Straight from the validation table

Performance claims are listed as "intentionally withheld until measured." That's the line that separates this from most trillion‑parameter‑anything repos on GitHub.

The discipline

No model gets called "compatible" until it earns it

Rather than shipping a model integration and hoping it holds up, ExpertMesh defines a fixed sequence of release gates that any target checkpoint — Kimi K2, K3, or otherwise — has to clear in order, before the label "compatible" is allowed to appear anywhere.

  1. Provenance locked. Immutable checkpoint revision and license recorded.
  2. Shape discovery. Parameter count, layers, experts, top‑k, and shared‑expert behavior read from the artifacts themselves.
  3. Bit‑for‑bit format check. Weight packing, scale dtype/granularity, zero‑point semantics, activation dtype verified exactly.
  4. Semantic reproduction. Attention (MLA + RoPE for K2), INT4 packing, tokenizer, tool parser, residual behavior, and router outputs matched on small fixtures.
  5. Kernel parity. HIP expert output diffed against the official reference on adversarial values and odd dimensions.
  6. Transport at scale. Two‑node validation before anything touches all eight nodes.
  7. Instrumented measurement. TTFT, inter‑token latency, aggregate throughput, HBM and RoCE traffic, expert skew, errors, and wall power — all captured, not estimated.

Until every gate closes, the shipped config is labeled exactly for what it is: configs/phase1-mi210.json, an assumption profile.

Hands on

You can run the whole reference path without an AMD GPU

This is the part most "hardware‑aware" AI infrastructure repos don't offer: a Phase 1 you can actually exercise on a laptop, end to end, planner through a live chat‑completions endpoint.

# plan and check the roofline math
python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

pytest -q
expertmesh-plan configs/phase1-mi210.json
expertmesh-roofline configs/phase1-mi210.json

# spin up a tiny deterministic MoE end to end
expertmesh-checkpoint .demo --workers 2
expertmesh-serve .demo --workers 2
# → POST /v1/chat/completions on :8000

The generated model here is a plumbing fixture, not a useful language model — it exists to prove the coordinator/worker path, checkpoint sharding, and serving loop actually connect end to end.

The verdict

An honest early‑stage bet, not a finished product

ExpertMesh is early — one contributor, ten commits, a single star as of this writing — and its central bet, that commodity AMD hardware plus a stationary‑expert data plane can approach trillion‑parameter serving at a fraction of the usual cost, is still unproven on real silicon. The HIP kernel needs a ROCm runner it hasn't had yet. No checkpoint has cleared the release gates.

But the engineering posture is exactly right for a project at this stage: a real reference implementation you can run today, a cost model you can inspect line by line, and a validation table that says "not measured" instead of guessing. That's worth more, long‑term, than a flashy benchmark nobody can reproduce.

8/ 10 — strong scaffolding, hardware verification pending