$AMD Helios networking explained:
Scale-up: inside the rack
- 72 MI455X GPUs
- 12 Broadcom Tomahawk 6 ASICs
- 36 parallel network planes
- 3.6 TB/s bidirectional per GPU
- 260 TB/s aggregate
The network is based on two standards:
> UALoE defines how GPUs communicate, including memory reads and writes, transaction ordering, and how traffic is distributed across the 36 network planes
> ESUN provides the Ethernet fabric underneath, handling forwarding, flow control, congestion management, link-level retries, and lower-overhead transmission for small messages
The path is:
MI455X → UALoE → ESUN Ethernet → Tomahawk 6 → MI455X
The 31 TB of HBM4 stays physically distributed across 72 GPUs. UALoE lets any GPU read or write any other GPU's memory directly, making the whole rack act as a single computer
Scale-out: between racks
The racks are connected with Ethernet scale-out through 800G Pensando Vulcano AI NICs into a UEC-ready fabric
The path is:
Helios rack → Vulcano NIC → UEC Ethernet → another Helios rack
The price paid for laser reliability in external lasers for CPO is dire.
See the optical link budget below.
The laser puts out 24.5 dBm, which is equivalent to 280 mW.
The coupling and connector losses, fiber losses, 4-way split, Sipho modulator losses, and multiplexer losses drop about 21.6 dB before the signal reaches TP2.
In terms of power, only 2mW of the generated 280mW is actually available at the optical engine output.
99.3% of the generated power is simply lost.
All because you can't integrate the laser, and have to keep it outside so it stays alive.
Similar to the panic over DeepSeek R1, some uneducated people think Kimi K3’s use of linear attention (KDA) is bad for NVIDIA, HBM, DRAM, and networking because it has relatively lower KV-cache requirements. The opposite is true, and we explain why below. 👇️ 1/8🧵
PicoJool's Al Yuen: The Case for GaAs VCSELs in Scale-Up Interconnects
Austin talks with PicoJool CEO Al Yuen about 200G/lane GaAs VCSELs for scale-up AI interconnects.
- GaAs unconstrained, InP sold out: 8-10 weeks vs 8-18 months
- Supply: InP single-mode built for 100K links, not millions/month. GaAs VCSEL supply chain can handle it.
- Three ways to 1.6T: 8×200, 16×100 LPO, 32×50 NRZ
- 12.8T needs no new physics, just more channels and bi-di
- Capacity isn't the wait, qual is: WIN runs 1,000 wafers/week. GaAs wafer has 240K VCSELs
Chapters:
0:00 Meet Al Yuen and PicoJool
2:29 Inventing the active optical cable
5:03 Engineering mindset, copper limits
8:43 Why VCSELs
13:45 Scale-up and bit error rate
20:09 Unconstrained vs constrained supply
21:53 Indium phosphide bottleneck
25:49 VCSEL design and foundry handoff
31:59 Product road map, 200G launch
34:16 Path to 3.2T and 12.8T
40:10 Ordering a million VCSELs
45:10 Ramp timing and training new engineers
Follow Chipstrat:
Newsletter: https://t.co/bTnS0luOzO
X: https://t.co/2lQ9JpAQUU
Follow Vik:
Newsletter: https://t.co/YRSASRTEfB
X: https://t.co/hcA6nqVq8q
Follow Semi Doped:
Get more of Austin and Vik daily, free!
Sign up: https://t.co/7trwA2esP9
关于美光超高毛利是否合理,云厂商这样抢存储是不是不理智,我还是想拿上次类比的中国房地产周期来类比。
大家应该都有这个记忆,就是过去中国20年的地产周期期间,市场多少次也在质疑:地价都超过楼面价了,这生意怎么做?这个现在存储看似不合理的价格何其相似,但后来真正发生的是,土地变贵以后,房价、融资、周转、城市预期一起被重新定价。不是每个开发商都赢,但能拿到核心地块的人,反而掌握了下一轮扩张的门票。
AI 现在也类似。
如果只看今天的 HBM、DRAM、NAND 价格,确实很离谱。客户像是在高位抢货。但云厂商买的不是一颗颗存储芯片,而是未来 AI 服务的“楼面”。
GPU 没有足够 HBM,算力跑不满;模型没有足够 DRAM/SSD,长上下文、缓存、RAG、推理吞吐都上不去。对云厂商来说,存储不是成本表里一个普通零件,而是决定未来 AI 产能、客户锁定和收入天花板的地基。
所以他们愿意高价锁货,本质上不是因为今天便宜,而是怕明天没货。
如果未来 AI token、agent、企业订阅、广告、云服务价格继续水涨船高,今天看起来很贵的存储,可能就像当年的核心土地:当时看是高价,事后看是稀缺资产。
这也是为什么我觉得这轮存储周期不能只按传统库存周期看。真正的问题不是“价格贵不贵”,而是“下游有没有能力把更贵的存储变成更高价值的 AI 服务”。
如果答案是有,那高价抢存储就不是疯狂,而是抢下一轮 AI 产能的土地。
所以现在重点盯的就是Openai和Anthropic两家的算力紧张程度,模型除了写代码以外能否扩展到其他行业和领域。