Hacker News Reader

Comment by throwaw12 | Hacker News Reader

(I don't understand hardware well)

Can someone shed light on why China still couldn't copy the Nvidia GPUs in some form?

I understand its complex and there many parts to it, but which is the most complex part making it difficult for China to copy it?

Let's say they don't have access to 3nm process, what if they just use 12nm and create GPUs with much bigger size but comparable performance with CUDA compatibility? Or other option could be less tensor units, training will take longer, but they might be able to produce it cheaply

View on HN

Copying CPUs isn't really a thing: they are too complex.

If you could steal all the designs at TSMC, and you had exactly the process that TSMC uses, you could definitely make counterfeits. If you didn't have TSMC's specific process, you could adapt the designs (to Intel or Samsung) with serious but not epic effort. If you couldn't make the processes similar (ie, want to fab on SMIC), you are basically back to RTL, and can look forward to the most expensive and time-consuming part of chip design.

This is nothing like copying a trivial, non-complex item like a car. Copying a modern jet engine is starting to get close (for instance, single-crystal blades), but even they are much simpler. I mention the latter because the largest, most resourced countries in the world have tried and are still trying.

They have done a bit of this. SMIC is basically operating off of a cloned TSMC N7 node that they have since iterated on to get to a 5nm class node.

But its still such a complex sort of beast.

Even if you had 'ai tools' guessing at component blocks on evaluation you would have to have some evaluation of the result.

And, thats assuming NVDA hasn't pulled a Masatoshi Shima type play on their designs (i.e. complex traps that could require lots of analysis to determine if they are real or fake)

Im not sure how much of a speedup even modern tooling/workflow could do reliably.

Even then,

The elephant in the room is that China is working on their own AI accelerators/etc, so while there can be benefit from -studying- the existing designs, however I think they do not want to clone regardless.

Oh, absolutely. Straight up soviet style cloning of masks makes no sense for multitude of reasons. In addition to what you've said, China isn't banned from N7 class Nvidia architectures so could just buy those on the open market.

If engines are hard to build, why not build a car 3x the size of a normal one, well you can but due to things like aerodynamics, etc etc you'll never match the speed or fuel economy of cars.

Same with chips, efficiency, speed, etc all depend on good design, and cutting edge factors, if the main reason your chip isn't faster is because of the distance between your L1 cache and your core is far, then having a bigger node process but bigger chip won't make it quicker.

Exactly, you can build 12nm but you can't quadruple the speed of light

> Can someone shed light on why China still couldn't copy the Nvidia GPUs in some form?

They have alternatives, like the Tian supercomputer was originally built with Xeon Phi chips that have been replaced with their own domestic alternatives.

A big limitation is getting access to fab slots. Nvidia and Apple are very aggressive about buying up capacity from TSMC, etc, and China's own domestic fabs are improving fast but still not a real match, particularly for volume.

They can given enough time.

But there's a distinct time/value of investment equation with the current AI boom. The jury is at best still out on what that equation is for the goals of capital (it's increasingly looking like there's no moat), but if you're a national government trying to encourage local bleeding edge expertise in new fields like this it's quite a bit more clear.

Another factor, it's not just GPUs it's the full hardware stack. https://static.tweaktown.com/news/1/1/110521_2_nvidia-update...

At 3 GHz, a signal can travel at most 10 cm per clock cycle. You can't really physically scale a chip up.

You can you just have to use a tiled architecture. And microprocessors already have far shorter wiring distances than the simple speed of light calculation because it takes time for the gates to make the transition as well.

With processors it's customary to use the "Fan out of 4" metric as a measurement of the critical paths. It's the notional display for a gate with fan out of 4, which is the typical case for moving between latches/registers. Microprocessor critical paths are usually on the scale of ~10 FO4.

The largest chip at the moment is Cerebras's wafer scale accelerator. There the tile is basically at the reticule limit, and they worked with TSMC to develop a method to wire across the gaps between reticules.

If you can make much cheaper chips and use much cheaper energy it might not matter.

Mostly high end lithography.

They can copy it. And no, the software moat is not there if someone choose the blatant copy route. They just can't build it in the scale they want yet.

> what if they just use 12nm and create GPUs with much bigger size but comparable performance

Physics do not work this way :/

well, physics does work that way, depending on what you mean by performance. (in the sense that power is normally part of performance when we're talking about chips).

you could certainly use a larger process and clone chips at an area and power penalty. but area is the main factor in yield, and talking about power is really talking about "what's the highest clockrate can you can still cool".

so: a clone would work in physics, but it would be slow and hot and expensive (low yield). I think issues like propagation delay would be second- or third-order (the whole point of GPUs is to be latency-tolerant, after all).

Reply on news.ycombinator.com