Huawei's chip law looks less like Moore and more like marketing

Huawei is touting a semiconductor advance - including something it calls the "Tau Scaling Law" as a successor to Moore's Law - but at least one chip expert says the news is more branding than breakthrough. The Chinese tech titan unveiled its advances at the 2026 IEEE International Symposium on Circuits and Systems (ISCAS) in …

[Avon B7]

What's he smoking?

"That's density achieved though clever packaging, not transistor shrinking, and is not comparable to a real TSMC/Intel 1.4nm transistor,"

It was - explicitly - spelt out during the presentation that is wasn't transistor shrinking.

Did he even watch it? Did he read the accompanying paper?

The - whole point - was to get off the transistor shrinking train and onto time scaling. Something they claim everyone will be boarding at some point.

They revealed that hundreds of HiSilicon chips (across multiple fields) have already been produced over the last six years and that the Kirin 2026 will be the first to ship with LogicFolding. That chip is - already - in silicon but unreleased (next year's is currently in pre-silicon) so Huawei already has the production data with regards to metrics. He (I'm referring to her here, lol) also said the hybrid bonding was deliberately modest on the Kirin 2026 so there is known room for improvement there.

Cost was another major reason that transistor shrinking was said to be running out of steam.

Unified Bus was also part of the plan but that isn't new (it's already present on AI cluster systems).

All this without touching on the rumours of Huawei testing an EUV prototype since last year (something they obviously won't be talking about until it's ready - much like with this Tau Scaling Law).

He said that challenges remain but those can be resolved through partnering.

[Anonymous Coward]

Re: What's he smoking?

When defending your tribe you got to ignore facts and just glaze your side. It was really amusing seeing Jensen retort back to an interviewer about connectivity "Huawei is a networking company!!!"

[DS999]

Re: What's he smoking?

Yes it sounds to me like they are fabricating a second layer of transistors on the same die. That's always been "impossible" because the process steps used to lay down transistors are incompatible with the process steps used to lay down metal wires.

Presumably they've found a way around such issues - this is something others are working on as well. For example, Apple has a few recent patents discussing ways around this with ideas to add power related transistors or even SRAM within the metal stack, but not full on logic. But having it in a patent and having it in mass production are different things.

[DS999]

Re: What's he smoking?

OK so it turns out no it isn't a second layer of transistors. It is die stacking like what TSMC does for AMD's X3D CPUs that have stacked cache, but with a much finer pitch and Huawei is stacking logic dies which is a lot more complex if the logic really is treated as if it has two levels (i.e. not just stacking additional cores on top but having everything truly split between the dies)

That's a complicated design scenario but they say they've written software to handle that side of it. It is also a problem for heat but so far they've only done it in two layers and for a mobile SoC. Whether they can expand that to more layers and do it at the power levels of AI chips is another matter. I guess we'll see.

[mevets]

Good grief....

It wasn't enough that the East trumped us in manufacturing, design, research, science, technology and clean power, but now they are trying to steal the marketing bullshit crown?

What will we be left with? A gay hockey soap opera and low grade pornography.

The West really is over.

[Anonymous Coward]

Huawei -- It's Not The Tool -- It's The Job The Tool Is Used To Do!

See title.

[Anonymous Coward]

Intel's 14A isn't really 1.4nm either

It's a bit odd for an industry expert to claim that the new process is "not comparable to a real TSMC/Intel 1.4nm transistor." Intel's "14A" and "18A" processes are marketing names - they are claimed to be "equivalent to" 1.8nm and 1.4nm, but the gate lengths are longer. The site

https://semiwiki.com/wikis/industry-wikis/intel-18a-process-technology-wiki/

claims that the Intel processes are equivalent to the ones TSMC and Samsung call "2nm" (which aren't 2nm feature-size either). My recollection is that the process-node names stopped referring to actual sizes some time ago (14nm?), but I can't dig up a good reference. Anyone?

[Sil]

Haven't TSMC and Intel already transformed nodes from an engineering concept to a marketing one a few years ago ?

[Henry Wertz 1]

If it's good enough for Intel...

If it's good enough for Intel... I'll just comment Intel referrs to their 'enhanced superfin' 10nm process as 'Intel 7' (implying it's equivalent to a 7nm process) and renamed their 7nm process 'Intel 4.'.. With the aame reasoning (we can do all this adnvanced stuff to get more speed and desnity out of the process). Of course that Intel 4 being comparable to 4nm is assuming those actually producing 4nm chips wouldn't just use similar methods themselves.

So although I thinkl it's silly for Huawei to do this, Intel pulls similar shenanigans with marketing their foundry tech.