China's first 7nm GPGPU is ready to go against AMD and Nvidia

nanoguy

Posts: 770   +12
Staff member
In brief: Tianshu Zhixin Semiconductor, a fabless joint venture between Via Technologies and the Shanghai municipal government, has revealed what appears to be China's first 7nm GPGPU, a gargantuan chip that's supposed to compete with the likes of Nvidia A100 and AMD MI100 in the data center.

China is on a quest for "technological self-sufficiency," and the trade war with the US has only accelerated that process. China is capable of covering over 20 percent of the chips needed in the local industry, but the government plans to increase that number to 70 percent by 2025. One example of that is using aging chips while its semiconductor industry catches up, turning to Japan for etching equipment that would otherwise be considered obsolete.

Last year, we saw the arrival of an impressive x86 CPU courtesy of government-backed Zhaoxin, which showed just how much progress has been made on that front. We've also heard about Huawei working hard to develop server GPUs, but things have been relatively quiet ever since the tech giant lost access to key suppliers such as TSMC.

Tianshu Zhixin was known to be making progress on the first general-purpose GPU designed in China, dubbed "Big Island." The ongoing chip shortage only allowed the company to do a paper launch in March, but what's interesting about these chips is that they're supposedly able to compete with Nvidia and AMD's GPU at the highest end.

Big Island GPUs aren't technically designed to go against RTX and Radeon gaming graphics cards, but instead they're geared towards machine learning, high performance computing, medical research, and security. This means they're meant to go against the likes of Nvidia's A100 and AMD's Instinct MI100, both of which are gargantuan chips designed for the data center and offer an order of magnitude the performance of previous generation architectures, while taking a lot less space and power to operate.

Big Island was developed between 2018 and 2020, built using TSMC's 7nm process node and 2.5D CoWoS packaging. They also utilize TSMC's 65 nm silicon interposer, and feature no less than 24 billion transistors. Just like AMD and Nvidia, Tianshu Zhixin has equipped its GPGPU with 32 GB of HBM2 memory (1.2 TB per second bandwidth) and made it compatible with the PCIe 4.0 standard.

The claimed performance of the Chinese GPGPU seems to be impressive. When it comes to FP16 (half precision math) performance, it is capable of 147 teraflops, which sits right between Nvidia A100's 78 teraflops and AMD MI100's 184.6 teraflops -- although it should be noted the A100's Tensor cores are capable of 312 FP16 teraflops.

When it comes to FP32 workloads, Big Island should be capable of up to 37 teraflops, which is higher than both the A100 and MI100, which offer 19.5 teraflops and 23.1 teraflops, respectively. Power consumption is rated at 300 watts, and Tianshu Zhixin says the new GPU will offer a better price-to-performance ratio than competing solutions.

Time will tell if this is indeed true and China has hit such an important milestone in its quest to reduce its reliance on foreign semiconductors.

Permalink to story.

 

umbala

Posts: 353   +460
Who care about IP? IP doesn't put food on my table or pay my bills. But cheap products help me.
You don't care about IP? Ok, what about the fact that every single item in every single store is made in China? There were certain things such as CPUs and GPUs China couldn't really make, but now they want to make those too. I have nothing against China, but this can't be good for any economy in the world. A lot of countries no longer have any manufacturing jobs because everything is manufactured in China at the lowest possible cost and lowest imaginable quality. But hey, as long as all your needs are met at the local dollar store, who cares about anything else, right?
 

neeyik

Posts: 1,877   +2,191
Staff member
I wonder how much IP they stole from the all the Chinese fabs making ARM/nVidia/AMD silicon?
AMD and Nvidia really only use manufacturers based in China for packaging, testing, and distribution. The GPUs themselves are fabricated at plants in Taiwan (TSMC) and South Korea (Samsung).

Given the resources involved in dissecting hardware and software for reverse engineering purposes available these days (example), it's not that hard to design and make a GPGPU. Making a good one, though, is an entirely different matter.

With regards to the Big Island chip, although it's clearly targeted at the likes of the A100 and Instinct MI100, it's claimed attributes and performance statistics suggest it's more akin to something like an RTX 3090:

chinas1stgpgpu.png

Tianshu Zhixin claim that their chip is 'half the size' of the competition and again, if one assumes they're referring to the likes of the A100, then their mockup picture isn't far off:

A100
F7oE-PUrSyJTa-LVxw50idY9qTDqcRx8V9uDTGyALeOc_pmozBBXYqi0xZGCTP0N4kZpaeSemafUmzSFLnevgh3VJi7FqQ4TwO7uqhrgrlsDt2DwZo5-MXidZiERVBnjqOY2rOqyiQ=w2400


Big Island
2021-04-07-image-29.jpg


BI is roughly 4 memory chips in area, whereas Nvidia's is about 9; the HBM2 chips in the two images don't exactly match in dimensions, but they're close enough. But can you have a GPGPU capable of the throughput figures being claimed in a chip that's somewhere between 350 and 400 mm2 in area?

Well, AMD's Vega 20 processor is 351 mm2 and the TSMC N7 version of it powers the Radeon Instinct MI60. But its performance figures are nowhere near what ZZ is claiming for their design (e.g. INT8 is only 59 TIOPS, vector FP32 is 14.7 TFLOPS), which is not surprising given that it's only a 13.23 billion transistor chip. AMD's Navi 21 has a similar transistor count to BI (at 26.8b) but again, it's throughput figures aren't anything like BI's and the chip is 520 mm2 in size.

So obviously something doesn't add up. Either the image is entirely hypothetical, the performance figures are, the physical dimensions are totally off, or even a combination of all of these. Sure, they could have shaved off a considerable portion of the transistor count by having very small caches and mediocre schedulers, but these would mean the claim peak throughputs aren't remotely unachievable.
 

orbital

Posts: 31   +30
You don't care about IP? Ok, what about the fact that every single item in every single store is made in China? There were certain things such as CPUs and GPUs China couldn't really make, but now they want to make those too. I have nothing against China, but this can't be good for any economy in the world. A lot of countries no longer have any manufacturing jobs because everything is manufactured in China at the lowest possible cost and lowest imaginable quality. But hey, as long as all your needs are met at the local dollar store, who cares about anything else, right?

This is not China's fault but one of the Big Tech (mostly US) manufacturers decided to outsource all their production (and in many cases certain R&D facilities) to Asia. China at this stage was eager to simply get any business on its territory lacking the industrial power, plants and absolutely basic know-how on how to do things, even the most simple ones. This was 30-40 years ago. Gaining more and more confidence and seeing its overall and IT infrastructure developing, also starting to get more and more engineers on its own able to independently think on their feet, it started first serving its basic economic and infrastructure needs which combined with its huge internal getting-richer market made them ask for more, with or without the blessing of their overseas investors.

What I see now is that every time I look for something online, be it an electric toothbrush, a clock with humidity and temperature indicators, or an electric kettle with display and temperature indicators, it's all China made and designed. Amazon is full of such things, so is every other retailer. EU do not design and sell such (or TBH any) products made on its territory, not to mention manufacture, so is USA. Who's fault is that I as an average consumer want to get the best for home and tech and no one apart from China itself is selling it to me? It would have been very funny if it wasn't pretty sad when a few months after China was devastated and gradually started recovering by the pandemic and Europe was hit even worse (Italy and Spain especially), the EU started importing en masse face masks, disinfectants, protective suits, and other PPE. Guess from where? Again, who's fault is this? EU stated they would arrange for their own PPE/Covid-19 related manufacturing facilities. Where are they 2 years after the very first lockdown? I bet in 10 years they will still import from China. 100%. For this or another virus...

We live in a dynamic world where empty spaces are getting filled in naturally. If profits and cost optimisation only rule the Big Tech, there is nothing good that comes in for Europe and the US. Actually I said it wrong, it's not cost optimisation - it's shareholders' dividends only and no care whatsoever to either paying customers or their own subcontractors of any type. Profits only and only for the few at the very top! Enough seeing what Intel did to us all as customers for the past 10 years! They would still be doing it to us if AMD did not intervene 4 years ago...

And if this doesn't change (I doubt it will), meanwhile I as an end user and consumer don't mind buying the best the tech and other worlds can offer at reasonable prices, no matter if it was manufactured and now even designed in China (more and more) or in the US/Europe (less and less). A simple rule of a dynamic economic, political and demand/supply matrix where actors and geographies change but interests, rarely.
 

Ludak021

Posts: 443   +303
What obsolete 7nm chips? No one is making 7nm chips for them so this can't be (entirely) true. There were "ASIC" ETH miners 2 something years ago. They consisted of nvidia mining cards and regular PC inside.
 

Mr Majestyk

Posts: 730   +618
Who care about IP? IP doesn't put food on my table or pay my bills. But cheap products help me.
Maybe you'll care when they put legit companies out of business and you can longer buy anything but a Chinese card where they can charge the west whatever they like.
 

mbk34

Posts: 179   +121
For some reason, the initial paragraph that explained what the hell a GPGPU is and what it's used for, was missing on my copy of this article.
 

pcmasterrrrrace9

Posts: 20   +5
In brief: Tianshu Zhixin Semiconductor, a fabless joint venture between Via Technologies and the Shanghai municipal government, has revealed what appears to be China's first 7nm GPGPU, a gargantuan chip that's supposed to compete with the likes of Nvidia A100 and AMD MI100 in the data center.

China is on a quest for "technological self-sufficiency," and the trade war with the US has only accelerated that process. China is capable of covering over 20 percent of the chips needed in the local industry, but the government plans to increase that number to 70 percent by 2025. One example of that is using aging chips while its semiconductor industry catches up, turning to Japan for etching equipment that would otherwise be considered obsolete.

Last year, we saw the arrival of an impressive x86 CPU courtesy of government-backed Zhaoxin, which showed just how much progress has been made on that front. We've also heard about Huawei working hard to develop server GPUs, but things have been relatively quiet ever since the tech giant lost access to key suppliers such as TSMC.

Tianshu Zhixin was known to be making progress on the first general-purpose GPU designed in China, dubbed "Big Island." The ongoing chip shortage only allowed the company to do a paper launch in March, but what's interesting about these chips is that they're supposedly able to compete with Nvidia and AMD's GPU at the highest end.

Big Island GPUs aren't technically designed to go against RTX and Radeon gaming graphics cards, but instead they're geared towards machine learning, high performance computing, medical research, and security. This means they're meant to go against the likes of Nvidia's A100 and AMD's Instinct MI100, both of which are gargantuan chips designed for the data center and offer an order of magnitude the performance of previous generation architectures, while taking a lot less space and power to operate.

Big Island was developed between 2018 and 2020, built using TSMC's 7nm process node and 2.5D CoWoS packaging. They also utilize TSMC's 65 nm silicon interposer, and feature no less than 24 billion transistors. Just like AMD and Nvidia, Tianshu Zhixin has equipped its GPGPU with 32 GB of HBM2 memory (1.2 TB per second bandwidth) and made it compatible with the PCIe 4.0 standard.

The claimed performance of the Chinese GPGPU seems to be impressive. When it comes to FP16 (half precision math) performance, it is capable of 147 teraflops, which sits right between Nvidia A100's 78 teraflops and AMD MI100's 184.6 teraflops -- although it should be noted the A100's Tensor cores are capable of 312 FP16 teraflops.

When it comes to FP32 workloads, Big Island should be capable of up to 37 teraflops, which is higher than both the A100 and MI100, which offer 19.5 teraflops and 23.1 teraflops, respectively. Power consumption is rated at 300 watts, and Tianshu Zhixin says the new GPU will offer a better price-to-performance ratio than competing solutions.

Time will tell if this is indeed true and China has hit such an important milestone in its quest to reduce its reliance on foreign semiconductors.

Permalink to story.

I honestly believe that China will inevitably become the biggest tech hub not only in Asia but also in the entire world.
 
But the big question is: how good is it for mining? :p
Its the best. Best mining GPU of all time. ALL the miners will want one and proudly say, "MY GPU was made in Cchhhyna!" Make mining great again...get it off gaming GPUs and get it on the Chyna GPU.
 

Markoni35

Posts: 1,075   +441
- And how much do you think IMB has stolen IP in 1960'es when they were the main producer of mainframe computers for the entire world? Whenever a disk would die, you'd send the old disk to IBM to get the new one. The old disk would then be read (even if it was erased) by the IBM experts to retrieve info from pharmaceutical corporations, software developers, military organizations and other users of their computers. Stealing all the data they could, and selling it to their partners (and US government).

- How much IP do you think Google and Apple steal now, when they can listen to almost everyone on the planet, and read their emails (with all the attachments) and know their schedules.

- Same goes for "secure" apps like Telegram. Maybe they are secure from my ISP, but how secure is my information from Telegram itself? If I'm sending passwords to my clients using a "secure" Telegram, Slack or Keybase, how secure is that data from them.

So, when it comes to spying and copying IP, everyone is stealing from everyone. It's just like in nature. Nothing is safe. Insects or reptilians may have invented wings, but that didn't prevent mammals like bats to develop their wings. And you can bet they didn't pay royalties to anyone.