Asimov晶片評測：8.75億美元豪賭不用HBM晶片 | Asimov Chip Review: $875M Bet on Memory Over HBM

9月 13, 2026

By Kit 小克 | AI Tool Observer | 2026-09-14

🇹🇼 Asimov晶片評測：8.75億美元豪賭不用HBM晶片

Asimov 是新創 Positron AI 正在打造的 AI 推論晶片，這家公司 9 月 10 日才剛募到 8.75 億美元，估值衝上 50 億美元——比今年 2 月的 10 億美元估值翻了快 5 倍。賣點很簡單：不用供應吃緊的 HBM，改用一般伺服器就在用的 LPDDR5X 記憶體，號稱能把 AI 推論的「每美元 token 數」拉到 Nvidia GB300 的 26 倍。

Asimov晶片怎麼跳過HBM衝規格

Asimov 單顆晶片可以搭 288GB 到 2,304GB 的 LPDDR5X，目標是撐起 16 兆參數的模型、單機處理超過 1000 萬 token 的上下文。把 4 到 8 顆 Asimov 兜在一起，就是 Positron 主打的 Titan 系統。晶片預計今年底在台積電 N3P 製程流片，量產時間已經從原訂 2027 年初延到下半年。

Nvidia比較表，先看清楚是誰量的

Positron 宣稱：尖峰效能下每美元 token 數是 Nvidia GB300 的 26 倍，低負載時剩 2.4 倍
記憶體頻寬利用率號稱 90% 以上，對比 GPU 普遍不到 30%
獨立媒體 The Register 估算：把效率優勢算進去，Nvidia 下一代 Rubin 的 HBM 仍比 Asimov 快約 2.4 倍

老實說：跑分全部來自模擬，晶片還沒流片

這才是這則新聞最該被畫重點的地方。所有 Asimov 的效能數字都來自「週期精確模擬」（cycle-accurate simulation），實體晶片要等到 2027 下半年才會出貨，現在市面上沒有任何一顆真的 Asimov 可以拿來實測。更微妙的是，這些跑分數據來源是研究機構 SemiAnalysis，而 SemiAnalysis 創辦人正是這輪募資的共同領投人，還進了 Positron 的董事會——利益衝突寫得很明顯，數字要打個折扣看。

不是空氣公司：Atlas已經在Oracle雲上跑

Positron 不是紙上談兵。前一代產品 Atlas（用 Intel Agilex 7 FPGA + HBM/DDR5）已經有 50 多個機櫃跑在 Oracle Cloud，客戶包括 Jump Trading 和 i3d.net，推論服務商 Parasail 也在用這批算力。這給了 Positron 一定的執行力信用，但 Atlas 用的架構跟 Asimov 完全不同，不能直接拿來背書 LPDDR5X 路線會成功。

結論：Asimov 代表的「用便宜記憶體換算力密度」路線值得關注，尤其在 HBM 缺貨、Nvidia 一卡難求的當下，這是少數認真挑戰記憶體瓶頸的方案。但在真晶片流片、跑出真跑分之前，8.75 億美元買到的還是一份很有說服力的 PPT。好不好用，試了才知道。

🇺🇸 Asimov Chip Review: $875M Bet on Memory Over HBM

Asimov is the AI inference chip that startup Positron AI is racing to build, and the company just raised $875 million on September 10, pushing its valuation to $5 billion — nearly 5x its $1 billion valuation from February. The pitch: skip supply-constrained HBM entirely and use commodity LPDDR5X memory instead, claiming up to 26x more tokens per dollar than Nvidia's GB300 at peak throughput.

How Asimov Skips HBM for Commodity Memory

Each Asimov chip pairs with 288GB to 2,304GB of LPDDR5X, aimed at serving models beyond 16 trillion parameters with context windows past 10 million tokens. Bundle four to eight chips together and you get Positron's Titan system. Asimov is set to tape out on TSMC's N3P process by the end of 2026, with production already pushed from early 2027 to the second half of 2027.

The Nvidia Comparison — Check Who's Measuring

Positron claims 26x more tokens-per-dollar than Nvidia GB300 at peak speed, dropping to 2.4x at lower throughput
Claimed memory bandwidth utilization is 90%+ versus under 30% typical for GPUs
Independent outlet The Register estimates Nvidia's next-gen Rubin, even after accounting for efficiency gains, still runs roughly 2.4x faster than Asimov's memory architecture

The Honest Part: All Benchmarks Are Simulated

This is the detail that matters most. Every performance number for Asimov comes from cycle-accurate simulation — the physical chip won't ship until H2 2027, so there is currently no real silicon to test. It gets murkier: the benchmark data was produced by SemiAnalysis, whose founder co-led this funding round and just joined Positron's board. That's a disclosed conflict of interest, and it means every number here needs a discount.

Not Vaporware: Atlas Is Already Running at Oracle

Positron isn't purely speculative. Its prior product, Atlas (built on Intel Agilex 7 FPGAs with HBM/DDR5), already runs across 50+ racks at Oracle Cloud Infrastructure, serving customers like Jump Trading and i3d.net, with inference provider Parasail using that capacity. That buys Positron some execution credibility — but Atlas's architecture is entirely different from Asimov's, so it doesn't validate the LPDDR5X bet.

Bottom line: the idea behind Asimov — trading cheap memory for compute density — deserves attention, especially with HBM supply tight and Nvidia GPUs hard to get. But until real silicon ships real benchmarks, $875 million has bought a very convincing slide deck. 好不好用，試了才知道 (You won't know if it's good until you've actually tried it).

Sources / 資料來源

延伸閱讀 / Related Articles

AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

搜尋此網誌

AI小貼士