Your browser is not supported. Please update it.

16 July 2020

Zhaoxin preps homebrew GPU, as China cold war looms

What seems like an inevitable technological cold war between the West and China means that the arrival of Zhaoxin onto the global stage of GPU production is an important milestone. The Chinese vendor is preparing to offer both rivals to Intel’s x86 CPUs and now Nvidia’s GPUs, and will likely find itself supplying a huge domestic market, if the US begins restricting its firms from selling into China.

The news comes as another milestone has come to pass, where Nvidia passed Intel’s market cap for the first time. As of the time of writing, this reversal has been rectified, with Nvidia at $225 billion and Intel just under $250 billion, but Nvidia’s future looks much rosier than Intel – selling its GPUs into new AI and video applications, housed in cloud data centers and new edge locations too, while Intel faces an onslaught of ARM-based CPUs in both the data center and personal computing markets.

Meanwhile, Zhaoxin is preparing to launch its first standalone GPU late this year, or early next. Early indications suggest it is not going to be as powerful or efficient as Nvidia’s offerings. It has a 70 watt Thermal Design Power (TDP), which is significantly lower than the 250 watt TDP used in high-end desktop GPUs and server GPUs.

This suggests that Zhaoxin is aiming for the low-end of desktop performance, as 70 watts is still far too high for laptops, and would melt holes inside smartphones and tablets. But performance isn’t so much of a concern if you are the only supplier in town, and if Chinese firms find themselves cut off from Nvidia and AMD GPUs, they will have to look elsewhere.

To this end, Zhaoxin is only going to be using the 28nm TSMC planar fabrication process, for the GPU. For comparison, both Nvidia and AMD use 7nm FinFET for their latest designs. With the lower TDP and the larger manufacturing process, this new GPU would get thrashed in head-to-head benchmarks.

Zhaoxin’s CPUs aren’t much to write home about either – generally thought to be around two to three generations behind what Intel currently provides. It’s worth noting, however, that the vast majority of personal computing customers are not buying the bleeding edge of performance, and so again, that’s not something Zhaoxin has to be overly concerned about, if Chinese consumers can no longer purchase Intel and AMD parts.

Zhaoxin’s KaiXian KX-6000 is the top-end CPU, and has eight processing cores, with support for DDR4 memory, made using TSMC’s 16nm FinFET process, and is thought to be comparable to Intel’s seventh-generation i5 line – which has only has four cores and does not support hyperthreading (virtualized cores).

This brings us to an interesting wrinkle in the discussion. Much of the rhetoric surrounding US trade restrictions on China has focused on the theft of intellectual property and patent infringement. If this is the case, why are the likes of Zhaoxin so far behind the curve, in both CPU and GPU development? If China is already stealing designs, shouldn’t its domestic firms be doing a much better job?

If it is not already stealing designs, then similarly, if the cold war breaks out in earnest, then this should mean that it can simply spin up production facilities to churn out ‘stolen’ copies of the Intel, Nvidia, and AMD designs – with absolutely no fear of patent cases or embargos being used in retaliation.

This makes the justifications somewhat messy. Trying to protect US businesses by cutting them off from one of their largest markets isn’t a great idea, nor is provoking the country that houses the majority of their production infrastructure – as retaliation from China, such as blocking or seizing factories, would then damage these US firms in the rest of their markets.

Such a cold war would be damaging to all involved, and would upset Nvidia shareholders. This past week, the stock price reached an all-time high of $415, and the consensus among Wall Street types is that there is still room to grow. Data center and video gaming are the trends driving this surge, and Nvidia seems to have recovered from the fallout owing to the bitcoin crash.

Cloud gaming, where video games are rendered remotely and then streamed as video to the players, have threatened to take off for a few years now. However, the latest battle between Xbox and PlayStation is about to commence, and the death of consoles seems to be a prerequisite for the arrival of cloud-gaming dominance. Nvidia’s arch-rival AMD has wrapped up both console GPU deals, which means that you can broadly look at this battle as one between GPU providers – one in the cloud, the other in the living room.

Elsewhere, Nvidia beat out Intel to acquire Mellanox, a specialist vendor in the data center interconnect business, but paid some $6.9 billion to do so, back in March 2019. That deal turned the screw on Intel, and dealt another blow to a company that has had very little take up of its AI-focused products. For most in the AI game, it’s easier to run workloads on general purpose GPUs, and eventually move to dedicated silicon designs where needed.

FPGAs provide a sort-of middle ground, and while Intel paid $16.7 billion for Altera’s FPGA expertise back in 2015, it was not enough to secure the nascent AI market – letting Nvidia sweep in on a GPU wave. There is a glut of specialist providers too, targeting that exact market, and Graphcore just announced it had raised $150 million in funding, to reach a valuation of $2 billion. Do not be surprised if Graphcore and its IPUs also end up in Intel’s Programable technology group, eventually.