I Spent Two Days in a Customer's Data Center with the BR100 Card
Community Discussion · General

I Spent Two Days in a Customer's Data Center with the BR100 Card

Shen TouShen TouOct 52026/10/05 279 views

A friend doing embodied AI recommended Biren's BR100 to me. He said their training cluster started mixing in domestic cards, and costs could be pushed down. I said then let me go take a look. He helped me arrange a partner's data center, I stayed there for two days, and I rented compute myself for a few more days before and after, totaling less than a week.

First, what it is

A general-purpose GPU — in plain terms, a card that can both do AI and other parallel computing, not the same path as a dedicated accelerator that only does matrix multiplication. BR100 is Biren's flagship, 7nm process, 77 billion transistors, 1074 mm² die area, 8192 general-purpose streaming processors plus 512 tensor acceleration engines, 256MB on-chip distributed shared cache, HBM2E external, 64GB memory per card, 550W total card power.

These numbers are official public figures, not what I tested. Among them, the line "up to 2.8x faster than Nvidia Ampere" I'll leave with a question mark — depends on the scenario, depends on how you slice it. The comparison target is naturally the A100, H100 tier. Domestically, those who can make this spec can be counted on one hand.

Actually running it, where it gets stuck

Basically got nothing done the first day. Installing drivers, matching framework versions, compiling operators — it was evening before the environment was up. The old problem with domestic cards: it's not that the hardware is bad, it's that you have to fill the software stack pits yourself.

The second day I ran inference, a model with tens of billions of parameters, sliced with tensor parallelism. It ran, and the throughput I measured was acceptable. The benefit of 64GB per card is you don't have to scrimp on memory, saving the hassle of stacking cards. There was one operator not covered; after the error I switched to another implementation path to get around it. This kind of thing is normal on domestic cards, provided you have someone who can modify code.

The monitoring tool is made to match the nvidia-smi style; utilization, memory, and power are all visible. Be mentally prepared for power consumption — 550W is not a small number, and old data centers may not be able to handle it.

My friend's exact words were: usable, but you have to maintain a team that understands the low level.

Who should consider it, who shouldn't touch it

Suitable: teams with tight budgets, fixed model forms, and their own migration engineering capability. For those who train and then go live, and don't change models much after going live, the migration cost can be amortized in one go.

Not suitable: small teams, those changing models daily, those expecting out-of-the-box use. Ecosystem is not something parameters can make up for — the community is thin, documentation is scarce, and when you hit problems you basically carry it yourself.

Next step

From an investment perspective, the value of Biren's card is not in single-card benchmark scores. It's the first GPU stock in Hong Kong, and it rose sharply on the first day, but the valuation contains an option on domestic substitution, not an account calculated by shipment volume. Where the ceiling of this track is depends on when the software stack can catch up.

One suggestion for readers: don't rush to procure. First take the smallest model you have and find a provider that can supply compute for a one-week migration validation. If it runs within a week, then talk about the rest. If it doesn't run, it means what's missing isn't cards, it's people.

2 replies

?
Ctrl + Enter to reply
Gewu
GewuOct 6

That 550W point is too real. Old server rooms can't handle it without upgrading the power distribution. Ask about power before you ask about GPUs.

Zhi Wei
Zhi WeiOct 6

That 550W number is the real killer. If you fill a rack in an old server room, you'll probably have to upgrade the power distribution first. It's not a compute bottleneck, it's a power problem.