It’s no longer about doing better chips, but connecting half a million

The new battle of the AI ​​is no longer earned only with integrated circuits, but orchestrating complete systems that connect half a million chips working at the same time.

Why is it important. For decades, the technology industry has been governed by the Moore law: The number of transistors in a chip doubled every two years.

Today, calculation power advances come from the ability to connect endless amounts of chips between them, an effort that does not follow any predictable “law”.

The turn. AMD has presented Your mi400 chip Not as an isolated piece, but as a complete system called helios Rack of the pre -installed data center. It is the implicit recognition that the era of individual chips is over.

The context. Nvidia dominates the AI ​​because Jensen Huang had the vision of using GPUS for these calculations, which gave him the advantage of those who hit first.

But AMD is shortearing fast distances (Bet on smooth his usually goes wrong), and the inherent change of the industry has also changed competition.

Between the lines. Companies no longer hire chips engineers, but also more concrete profiles:

• Specialists in Design of racks.
• Systems architecture specialists.
• Specialists in data centers optimization.

AMD is now using More than 1,000 specialized engineers in design of racks. Architecture chiefs who previously focused on concrete domains should now be somewhat more general to create more integrated systems.

The threat. The biggest competitive pit (‘Moat‘) Nvidia is no longer hardware, but software. His suite CUDA It is the standard in the AI ​​industry to train models.

AMD competes with ROCM. And although changing system can be too expensive, especially for small companies or independent developers, large companies can justify spending if the greatest efficiency gives them the right return.

Yes, but. Companies are packing so much hardware in the racks that High soils and electrical networks They begin to be too tension. The question here is when they can sustain this race without finding physical limits.

Deepen. The CEO of COPEAidan Gomez, said In the event that was “very easy” to migrate one of your company’s models to the AMD platform.

Openai, Oracle and other great technological ones are looking for alternatives to Nvidia since he boom of chatgpt began to make your chips scarce. Supplier diversification is also strategic for projects like Stargatethe initiative of half a billion Openai dollars next to Oracle.

Outstanding image | İsmail Enes Ayhan in Unspash

In Xataka | Taiwan has decided to prevent by all means that TSMC technology arrives in China. Your future is at stake