AMD is the first to launch the world's first 7nm GPU, but can it be shipped on time?

As an annual ICT event, chip giants will bring attractive new products to the Taipei Computer Show. Following Intel, Nvidia, and Qualcomm, AMD also launched a number of new products on the 6th. AMD said it expects to release the AMD Radeon Vega GPU using the 7nm process and the 32-core second-generation Ryzen Threadripper processor in 2018. Interestingly, Intel just said the day before AMD that it will launch a 28-core processor by the end of this year, and the 7nm Vega is likely to become the world's first 7nm GPU. However, can AMD, which likes to be one step ahead in numbers, successfully launch the 7nm GPU?

At Computex 2018, AMD unveiled its first 7nm Vega with 32GB HBM2 video memory, and demonstrated photo rendering in Cinema4D at the event, supporting real-time ray tracing. At the same time, AMD announced that Radeon Instinct, based on the 7nm Radeon Vega architecture, has begun to provide samples to initial customers, and will begin shipping in the second half of 2018 in server and workstation specifications for critical computing application cases.

Dr. Lisa Su, President and CEO of AMD

“We demonstrated the industry’s strongest CPU and GPU portfolio, with upcoming 7nm and 12nm products building on and continuing the momentum of our Ryzen, Radeon and EPYC products, positioning AMD for the next generation of high-performance computing, from premium devices and gaming to machine learning and data centers,” said Dr. Lisa Su, AMD president and CEO.

According to Leifeng.com (official account: Leifeng.com), Vega is designed for servers and workstations, will add new optimizations for processing deep learning and artificial intelligence tasks, and will also become part of AMD's AMD Instinct series, a product line dedicated to accelerating machine learning. Dr. Su Zifeng also assured gamers that this technology will also be included in Radeon gaming graphics cards in the future.

From the products AMD has revealed this time, whether it is the 32-core Threadripper processor and the number of cores is more than Intel's 28 cores, or the world's first 7nm process GPU, it seems that AMD's love for numbers has continued. However, compared with the stacking of processor cores, to take the lead in process manufacturing, it not only requires AMD's advanced design capabilities, but also cooperation with wafer foundries is very important, especially in the context of the increasing difficulty of semiconductor process improvement.

AMD will choose TSMC even if it pays GF fine

In fact, AMD's 7nm GPU progress is not the first time to be made public. The most recent time was in a conference call after the release of AMD's first quarter 2018 financial report. Dr. Su Zifeng said that the production of 7nm GPU is going smoothly. She also said that the production of the chip is not produced by GlobalFoundries (GF) but in cooperation with TSMC.

While Dr. Lisa Su is not worried about the progress of AMD's 7nm GPU, AMD's choice of TSMC does not seem to be a good sign. Since AMD first spun off its foundry and created GlobalFoundries, it signed an agreement that it would move all 28nm CPU and GPU products to the new foundry. But what happened next was a disaster for both companies. GlobalFoundries hoped to gain revenue from AMD's first Fusion APU designed for notebooks (codenamed Llano), but AMD said that GlobalFoundries could not produce qualified products within the required time and had to hand over the Krishna and Wichita APU chips to TSMC for foundry. In addition, AMD's 28nm GPUs have never been manufactured at GlobalFoundries, and AMD has paid a lot of fees to GlobalFoundries for failing to meet its original agreement. At the 14nm stage, GlobalFoundries became Samsung's second foundry.

Later, AMD's Ryzen7 2700X proved that GlobalFoundries' 12nm process was up and running smoothly, but there have been questions about whether the 7nm process technology can be delivered on time. Now, AMD has chosen TSMC for its 7nm GPU. According to the terms of the original agreement between AMD and GlobalFoundries, the agreement is still valid, so AMD will have to pay a penalty for choosing TSMC. This alone is enough to raise questions about why AMD should pay a penalty to build Vega with other foundry partners. Another disturbing fact is that GlobalFoundries announced earlier this year that its CEO would be replaced by Sanjay Jha by Thomas Caulfield, and this is also the fourth CEO of GlobalFoundries in ten years.

The outside world has been worried about the transition from 14/12nm to 7nm since the launch of Ryzen. The installation of GlobalFoundries EUV equipment and the smooth production of 12nm and 14nm factories have alleviated these concerns, but the decision to replace Jha with Caulfield has once again increased the outside world's concerns about GlobalFoundries' progress. AMD's final decision to choose TSMC for some products is also a pessimism about GlobalFoundries' 7nm process.

Unlike AMD's spin-off of foundries, AMD's rival Intel's chips have always been manufactured by its own foundries and have advantages in advanced processes. However, Intel has been stuck at the 14nm process since 2014, which means that Intel's same process has gone through the fourth (or fifth) iteration, hindering its new architecture from being brought to market.

As for the reasons for the repeated delays, CEO Brian Krzanich explained that Intel's goal of increasing transistor density by 2.7 times on the 14nm process for the 10nm process is "a little bit aggressive." In contrast, Intel only increased density by 2.4 times when it upgraded to 14nm. Although the numerical difference seems small, the industry average is only 1.5-2 times. Due to the difficulties of 10nm manufacturing, Intel has revised its density target to 2.4 times in order to transition to the 7nm node, and will also rely more on its EMIB technology.

Leifeng.com also learned that the naming of semiconductor process nodes is different from the previous naming method based on scientific measurement. Recently, the naming of semiconductor processes is more inclined to marketing, so TSMC's 7nm and Intel's 10nm process are not exactly the same. However, TSMC is starting to mass produce 7nm products, which shows that foundries represented by TSMC are successfully solving production challenges.

As the world's largest wafer foundry, regardless of whether the naming of advanced processes is biased towards marketing, TSMC's 7nm process will be mass-produced this year, and more than 50 chips will be taped out one after another, involving industries such as cryptocurrency chips, networks, games, 5G, and autonomous driving chips.

The industry generally believes that Apple's A12 processor will be the first mass-produced 7nm processor and will occupy a large part of TSMC's production capacity. In addition, Xilinx's FPGA chips, Qualcomm's Snapdragon 855 mobile phone chips, and HiSilicon's Kirin 980 will also use TSMC's 7nm process.

TSMC's advanced process layout

According to Leifeng.com, Taiwanese media recently reported that the co-founder and CEO of the world's largest mining machine manufacturer Bitmain, Zhan Ketuan, came to Taiwan last week and visited TSMC to seek cooperation for a new generation of high-performance mining chips. At the same time, the supply chain reported that another mining machine manufacturer Canaan Creative's new generation of ASICs was also produced at TSMC, also using 7nm, and planned to go into mass production in July, which is faster than its competitor Bitmain.

Then, with the release of the new generation of iPhones in the second half of the year and the demand for 7nm from new products from Nvidia, Xilinx, Qualcomm, HiSilicon, Bitmain, Canaan Creative, etc., TSMC's production capacity will be tight, and the yield problem of the 7nm process in the early stage will also affect TSMC's production capacity. At the beginning of the year, it was reported that major chip manufacturers were competing for TSMC's production capacity. In 2018, when the 7nm advanced process has just started mass production, can AMD's 7nm GPU be mass-produced as scheduled? Can this help AMD seize the market opportunity?

Article reference: extremetech, tomshardware

• It is said that the next generation iPhone processor has been put into production, using TSMC's new 7nm process
  

• Mining machine giants compete for TSMC's production capacity, Canaan's 7nm ASIC chip mass production may be earlier than Bitmain
  

• Intel's 10nm process mass production pain, the goal is somewhat "radical"