Ultra-fast tokens · model-native silicon超高速 token · 模型原生芯片
A million tokens
a second.
For one user.
单用户,
每秒百万 token。
Little Gravity turns a fixed model’s heavy compute path into the chip itself. Weights become circuitry, history stays in a bounded on-chip state, and only live data moves. We build chips. What we deliver is ultra-fast tokens. 小有引力把确定模型的重计算路径直接变成芯片本体:权重成为电路,历史保存在有界的片上状态里,只有动态数据在流动。我们造芯片,交付的是超高速 token。
- Per-user speed单用户速度
- 1,000,000 tokens / s · design targettoken/s · 设计目标
- Weight traffic权重搬运
- 0 bytes at runtime · weights are circuitry运行时零字节 · 权重即电路
- Model state模型状态
- Bounded有界 no growing KV cache不再有增长型 KV 缓存
A model shapes the silicon. 图 01 — 质量弯曲时空,
模型塑造硅片。
01Thesis理念
Speed is a threshold.速度,是一道门槛。
Software is moving from fixed screens to agents generated on demand: a user states a goal, and the system assembles the tools, steps and interface in real time. The binding constraint is generation speed. 软件正在从预先写死的界面,走向按需生成的智能体:用户提出目标,系统实时组织工具、步骤与界面。这个未来最大的约束,是生成速度。
The user may read a few hundred words. Behind them, the agent proposes, checks, calls tools and revises. At a million tokens per second for one user, that inner loop runs continuously. The 8086 mattered not for its 5 MHz, but because it crossed the line where a personal computer became useful. 用户最终看到的也许只有几百字,背后的智能体却在不断提出候选、检查、调用工具并修改方案。达到单用户每秒百万 token,这个内部循环就能连续推进。8086 的意义不在于 5 MHz,而在于它跨过了让个人计算机真正可用的那道线。
Four design trades turn a model into a chip.四个设计取舍,把模型变成芯片。
-
01
Fixed weights become compute structure.固定权重成为计算结构。
W4 weights and fixed scales are compiled before tape-out into connections, selects, sign flips, shift-adds and local reductions. Inputs are still truly computed; what disappears is the repeated movement of constants.W4 权重与固定缩放在流片前被编译成连接、选择、取符号、移位加法与局部归约。变化的输入仍然经过真实计算,消失的只是常数的重复搬运。
-
02
Only dynamic data moves.只让动态数据流动。
Activations, local results and control relay through wires, pipeline registers, local FIFOs and accumulators. The main decode path never touches HBM or DDR.激活、局部结果与控制信号经由导线、流水寄存器、局部 FIFO 与累加器连续接力。主解码路径不依赖 HBM 或 DDR。
-
03
Bounded state replaces a growing KV cache.用有界状态取代增长型 KV 缓存。
History lives in a fixed-capacity recurrent state, not one more KV entry per token. Per-token state access stays constant, so timing and bandwidth can be budgeted exactly.历史保存在固定容量的循环状态中,而不是每个 token 追加一份 KV。每 token 的状态访问保持稳定,时序与带宽预算因此可以精确确定。
-
04
Generality moves up to the API.通用性上移到接口层。
Applications stay general through a standard LLM API. Underneath, the hardware serves one frozen model, and the area, bandwidth and scheduling budget of programmability goes to single-user speed.应用通过标准大模型 API 保持通用;底层硬件只服务一个冻结模型,把可编程性占用的面积、带宽与调度预算全部用于单用户速度。
02Technology技术
The model isn’t a workload.
It’s the chip.模型不是负载,
模型就是芯片。
Today’s AI chips treat the model as a runtime workload. While every token still re-reads weights, maintains a growing KV cache or passes through general scheduling, single-user decode cannot cross a million tokens per second. 今天的 AI 芯片把模型当作运行时负载。只要每个 token 仍要反复读取权重、维护持续增长的 KV 缓存,或经过通用调度,单用户解码就很难跨过每秒百万 token。
Four routes, and where each one stops.四条路线,各自停在哪里。
| Route路线 | Generality it keeps保留的通用性 | Main cost in single-user decode单用户解码的主要代价 |
|---|---|---|
| GPU / NPU | Many models and operators, mature software多模型、多算子、成熟软件生态 | HBM weight reads, KV cache, kernel and serving schedulingHBM 权重读取、KV 缓存、内核与服务调度 |
| Programmable dataflow可编程数据流 | Model graphs can be remapped模型图可重新映射 | Graph compilation, NoC routing, buffer exchange, sync图编译、片上网络路由、缓存交换与同步 |
| Fixed-weight Transformer固定权重 Transformer | Applications reuse one fixed model应用复用同一个固定模型 | Weight movement falls, but KV state still grows with context权重搬运下降,但 KV 状态仍随上下文增长 |
| Little Gravity小有引力 | General at the API: standard LLM interface在接口层保持通用:标准大模型 API | Only bounded dynamic data moves — never the model只移动有界的动态数据,从不移动模型 |
One token, end to end.一个 token,端到端。
- 1Input & scheduling.输入与调度。 A token-loop controller takes the new token and issues stage commands. The host only sends commands and receives results.token 循环控制器接收新 token,按层和阶段发令;主机只负责命令与结果。
- 228 layer clusters.28 个分层计算簇。 Fixed-coefficient issue, reduction, gating, normalization, requantization and state scan overlap inside a 30-cycle window per layer.固定系数发射、局部归约、门控/归一化、动态重量化与状态扫描在每层 30 周期窗口内重叠。
- 3Localized state.状态本地化。 Layered SRAM holds each layer’s cross-token state next to its cluster.分层 SRAM 紧邻各计算簇,保存各层跨 token 状态。
- 4No full logits.不搬完整 logits。 The output head is split into 116 shards, each producing a local top-8, merged by a 4-ary tree — about 64× less traffic than moving full logits.输出头拆为 116 个分片,各自产生本地 top-8,再经四叉树合并——候选流量比搬运完整 logits 降低约 64 倍。
- 5On-chip sampling.片上采样与反馈。 The next token is sampled on-chip and fed straight back with the new state version — a closed autoregressive loop.候选在片上完成采样,新 token 与状态版本立即反馈到下一轮,形成完整的自回归闭环。
The cycle budget.周期账。
At 1 GHz, a million tokens per second leaves 1,000 cycles per token. Our current architecture model fits 28 layers into 30-cycle fused windows and spends 44 cycles on the output head, candidate merge, sampling and feedback.在 1 GHz 下,每秒百万 token 意味着每个 token 只有 1,000 个周期。当前架构模型把 28 层压入每层 30 周期的融合窗口,再用 44 周期完成输出头、候选归并、采样与反馈。
The speed comes from freezing the model, unrolling it physically, fusing within layers and localizing state — together, not from accelerating any single operator.速度来自模型冻结、并行物理展开、层内融合与状态本地化的共同成立,而不是单一算子的加速。
- Fixed-coefficient ops / cycle每周期固定系数计算
- 8,388,608
- Margin vs 1.05M product point相对 105 万产品点余量
- ~68 cycles约 68 周期
03Products产品
Prove the chip.
Then optimize the model.先证明芯片,
再优化模型。
Two single-chip generations, not one bet. The first proves a strong open model can be cast whole into fixed-weight silicon. The second co-designs model and chip for the full million.两代单芯片产品,而不是一次豪赌。第一代证明优秀的开源模型可以完整芯片化;第二代让模型与芯片协同设计,冲击完整的百万 token。
-
Gen 1 · single chip第一代 · 单芯片LG1
A leading 27B open-weight model领先的 27B 开源权重模型
Unmodified. Proves a high-quality open model fits entirely in fixed-weight silicon, and builds a repeatable model-to-silicon delivery flow.不先改模型。证明高质量开源模型可以完整进入固定权重芯片,建立可复制的模型到硅交付能力。
10Ktokens/s per user · target单用户 token/s · 目标 -
Gen 2 · single chip第二代 · 单芯片LG2
Silicon-Optimized Model硅优化模型
Distillation, retraining and bounded-state design compress the KV path. Model and chip are designed together for extreme decode.通过蒸馏、重训练与有界状态设计压缩 KV 路径,让模型与芯片共同追求极致解码。
1Mtokens/s per user · target单用户 token/s · 目标 -
Then · multi-chip随后 · 多芯片LG×N
One larger model, sharded一个更大的模型,分片映射
One fixed model mapped across high-speed interconnected chips. A single chip delivers speed; many chips deliver scale — toward frontier-scale custom models.把同一个模型固定分片到多颗高速互连芯片上:单芯片负责速度,多芯片负责规模,逐步走向前沿规模的定制模型。
Scale规模frontier-class models前沿级模型
LG2 · product design pointLG2 · 产品设计点
- Per-user speed单用户速度
- 1.05M tokens/s design point设计点 105 万 token/s
- Process工艺
- 7 nm class, ~500 mm² core area7 nm 级,核心面积约 500 mm²
- Clock时钟
- 1 GHz
- Numerics数值
- W4 weights, A8 activations, local INT32 accumulationW4 权重、A8 激活、局部 INT32 累加
- Runtime SRAM运行时 SRAM
- 256 MiB: layered state + global staging, debug and replay256 MiB:分层状态 + 全局暂存、调试与回放
- External memory外部内存
- None on the main decode path主解码路径不使用
- Interface接口
- Standard LLM API for agent runtimes面向 Agent 运行时的标准大模型 API
Design targets from the current architecture model, not measured silicon.以上为当前架构模型的设计目标,非硅片实测。
04Evidence工程进展
One engineering chain,
model to metal.从模型到金属,
一条工程链路。
The moat is not one operator. It is the whole chain — model architecture, quantization and training, fixed-weight macros, token scheduling, state SRAM, physical implementation and post-silicon validation — owned by one team.真正的壁垒不是某一个算子,而是由同一支团队掌握的完整链路:模型架构、量化与训练、固定权重宏、token 调度、状态 SRAM、物理实现与硅后验证。
-
Complete已完成
Model & numerics模型与数值链路
- 2,069,873,088 parameters across 672 tensors inventoried, mapped to 1,827 fixed physical units完整盘点 672 个张量、2,069,873,088 个参数,映射为 1,827 个固定物理单元
- 452 / 452 known-answer tests pass on a 1.42B fixed-coefficient subset14.2 亿固定系数子集通过 452 / 452 项已知答案测试
-
Complete已完成
Architecture plan架构计划
- All 28 layers, state, output, sampling and feedback in one 884-cycle plan完整 28 层、状态、输出、采样与反馈纳入统一的 884 周期计划
- W4 / A8 datapath with local INT32 accumulationW4 / A8 数据通路,局部 INT32 累加
-
Complete已完成
Physical prototype物理原型
- Macro-level full-chip prototype on the ASAP7 predictive PDK, 28 × 28 mm top level基于 ASAP7 预测型 PDK 的宏级全芯片原型,28 × 28 mm 顶层
- 45 macro components, 687 / 687 nets globally routed, zero overflow45 个宏级组件,687 / 687 条网络全局布通,overflow 为零
-
Next下一步
Commercial silicon商业流片
- 7 nm-class implementation and sign-off7 nm 级工艺实现与签核
- Tape-out, packaging, bring-up against the same reference tests流片、封装,并用同一套参考测试完成硅后验证
05Applications应用
Frontier models raise the ceiling. Ultra-fast tokens make it real time.前沿模型抬高能力上限,
超高速 token 让能力实时发生。
-
On-demand agents按需生成的智能体
Software generated from a goal: the agent, its tools and its interface assembled while the user waits — without making them wait.从目标直接生成软件:智能体、工具与界面在用户提出需求的瞬间完成组织,而不让用户等待。
Loops, not waits循环而非等待 -
Physical AI物理智能
Robots and machines that generate actions continuously, on a fixed path with deterministic timing.机器人与各类机器持续生成动作,运行在时序确定的固定路径上。
Deterministic timing确定时序 -
Long-running sessions长时运行任务
A bounded state means memory does not grow with the conversation. An hours-long task runs on the same silicon as the first message.有界状态意味着内存不随对话增长。持续数小时的任务,与第一条消息使用同一块硅片资源。
Fixed memory内存恒定 -
Private deployment私有化部署
A model on a board instead of a GPU cluster. No weight files to ship or guard at runtime — the weights are the circuit.一块板卡替代一整套 GPU 集群。运行时没有需要分发和看护的权重文件——权重就是电路本身。
Single board单板交付
Fast decisions need firm boundaries.决策越快,边界越要刚性。
Software agents软件智能体
Tool allow-lists, permission isolation, action budgets, audit logs and human confirmation.工具白名单、权限隔离、动作预算、审计记录与人工确认。
Physical systems物理系统
Timing constraints, amplitude limits, interlocks, watchdogs and safe actuator states — enforced by runtime and hardware.时序约束、幅度限制、联锁、看门狗与执行器安全状态——由运行时与硬件持续执行。
06Company公司
Small team.
Strong pull.团队不大,
引力不小。
Little Gravity — 小有引力 — builds chips, but what we sell is ultra-fast tokens. That route demands one team fluent in both the model and the silicon.小有引力造芯片,但我们出售的是超高速 token 能力。这条路线要求同一支团队同时掌握模型与芯片。
If you have taped out silicon, trained models to fit a hardware budget, or written compilers that end in metal, we would like to hear from you.如果你有流片经验,做过面向硬件预算的模型训练,或者写过一路编译到金属层的编译器,欢迎来聊。
Open roles开放职位
Physical Design & Sign-off物理设计与签核 PnR · STA · DRC/LVS · IR/EM布局布线 · STA · DRC/LVS · IR/EM RTL & Fixed-Weight MacrosRTL 与固定权重宏 Microarchitecture · Timing微架构 · 时序 Model–Silicon Co-design模型与硅协同设计 Distillation · Bounded state · W4/A8蒸馏 · 有界状态 · W4/A8 Verification & DFT验证与可测性设计 KAT · Equivalence · BISTKAT · 等价性检查 · BIST Bring-up, Board & Runtime硅后调试、板卡与运行时 Silicon validation · LLM API硅片验证 · 大模型接口07Contact联系
Bring us
a workload.带上你的
任务来。
We work with agent builders, robotics teams, model developers and companies that want a custom model-native chip. Tell us what you run and where speed matters.我们与智能体开发者、机器人团队、模型开发者,以及希望定制模型原生芯片的企业合作。告诉我们你在跑什么任务,速度卡在了哪里。