News The Rise of China GPU Makers: AI and Tech Sovereignty Drive New GPU Entrants

[Admin]

The number of GPU developers is increasing worldwide, but only a few can make a difference.

The Rise of China GPU Makers: AI and Tech Sovereignty Drive New GPU Entrants : Read more

 

[bit_user]

Thanks for the table, but it would've been interesting to see which GPU makers are using which IP. As mentioned in the article, at least a couple of the Chinese companies are using Imagination's. Another useful dimension that could've been added was to indicate the location of each. For instance Think Silicon is based in Greece. Think Silicon was bought by Synopsys? And Verisilicon owns the IP formerly known as Vivante? I was surprised to see how quickly Vivante came onto the scene, about a decade ago, and then it just seemed almost to go dormant. It seemed weird that more wasn't invested in continuing to build it up.

Regarding the section about applications, people should read "smart cities" mostly as "mass surveillance". The term means potentially more than that, but surveillance is a foundational element of smart cities, and AI-related technologies are key to that.

while there are a bunch of GPU developers, only two can actually build competitive discrete GPUs for PCs. That is perhaps because it is relatively easy to develop a GPU architecture, but it is truly hard to properly implement it and to design proper drivers.

I disagree. I think it's actually difficult to build competitive GPU hardware, because you need to do scaling really well, you need really good efficiency, and you need the right kinds and amounts of hardware-assist, to ensure the architecture isn't bottlenecked. Not to trivialize "drivers" (which encompasses a lot more than what people normally think of as a device driver), but the hardware aspect of GPUs isn't nearly as straight-forward as it used to be, a couple decades ago. The paragraphs following that statement seem to support me, on this point.

Anyway, thanks for the coverage. I look forward to seeing how this race shakes out. We should probably expect to see most of these new entrants specialize or die, with the field thinning out quickly, over the next few years.

P.S. I hope most of these new upstarts embrace OpenCL, for compute. That'd be a nice way to stick it to all the legacy players who've de-prioritized support for the standard.

 

[thisisaname]

P.S. I hope most of these new upstarts embrace OpenCL, for compute. That'd be a nice way to stick it to all the legacy players who've de-prioritized support for the standard.

I like standards but they do have to be kept up to date, else they become a burden rather than help.

 

[jp7189]

I remember when GPUs got stuck on 28nm for a few generations. They figured out how to do some great things at 28nm... not that those designs would be top of the heap today, but a 28nm wafer scale design (a la Cerebras) should be attainable inside China and quite potent. That of course assumes you have the budget of an entire country behind you.

 

[bit_user]

I remember when GPUs got stuck on 28nm for a few generations. They figured out how to do some great things at 28nm...

True, though mostly by making them huge.

Here are the 3 generations of AMD flagship GPUs that were made on 28 nm:

Launch (YYYY-MM)	Main Product	Codename	Architecture	Area (mm^2)	Shaders	Core Clockspeed (MHz)	fp32 GFLOPS	Power (W)
2012-01	HD 7970	Tahiti	GCN 1	352​	2048​	1050​	4301​	250​
2013-11	R9 290X	Hawaii	GCN 2	438​	2816​	1000​	5632​	320​
2015-06	R9 Fury X	Fiji	GCN 3	596​	4096​	1050​	8602​	275​

By using HBM, Fury was able to achieve some power savings. However, I also seem to recall the nominal 275 W was largely a fiction. I hadn't previously noticed how they stayed at ~1 GHz, but it makes sense.

Here are the 3 generations of Nvidia flagship GPUs that were made on 28 nm:

Launch (YYYY-MM)	Main Product	Codename	Architecture	Area (mm^2)	Shaders	Core Clockspeed (MHz)	fp32 GFLOPS	Power (W)
2012-03	GTX 680	GK104	Kepler	294​	1536​	1058​	3090​	195​
2013-02	GTX 780 Ti	GK110	Kepler	561​	2880​	980​	5121​	230​
2015-03	GTX 980 Ti	GM200	Maxwell	601​	3072​	1075​	6605​	250​

Note that Maxwell introduced tiled rendering. That was largely responsible for its roughly doubling of game performance over the GTX 780 Ti, even though the raw compute didn't increase much. It enabled the GTX 980 Ti to hold its own against Fury X, which was the much stronger GPU on paper.

a 28nm wafer scale design (a la Cerebras) should be attainable inside China and quite potent.

It would tie up quite a lot of production capacity.