Chinese Chipmaker Debuts First Homemade GPGPU; Based On TSMC 7nm FinFET Process

Shanghai Tianshu Intellectual Semiconductor Co. (Tianshu Zhixin) announced earlier this week that it was close to bringing its first-ever homegrown 7nm General Purpose GPU (GPGPU), Big Island (BI), into mass production and commercial delivery.

Tianshu Zhixin’s comes months after Big Island was first announced at the beginning of the year. Back then, the Chinese chipmaker hinted at BI being made with a 7nm process node, along with a 2.5D chip-on-water-on-substrate (CoWoS). The company failed to mention what 7nm process it was using for its GPGPU, but Tom’s Hardware has confirmed that it is a 7nm from TSMC’s foundries.

In addition to its “mass production” announcement, Tianshan Zhixin also shared some information about BI in the form of a slide. It’s written in Chinese, but based on what we can make out, the GPGPU is using 32GB HBM2 memory with speeds up to 1.2TB/s. There’s also a laundry list of its performance in floating point operations, including the usual FP32 and FP16 performance metrics.

At the time of writing, Tianshu Zhixin did not say when its first round of BI GPGPUs are scheduled to be shipped, nor did it say who would be its first batch of customers.

High-End NVIDIA CMP Mining GPUs Reportedly Based On GA102 Ampere GPU

Earlier last month, NVIDIA announced the Cryptocurrency Mining Processor (CMP), a GPU made specifically for mining cryptocurrency, with a total of four variants to be made available throughout the year. Now, there are reports detailing their alleged mining prowess.

According to Videocardz, it seems that the two lower-tier CMP HX models, the 30HX and 40HZ, will be based on NVIDIA’s previous generation Turing GPU, which made its debut back in 2018. Specifically, the Turing TU116 and TU106. As for the CMP HX50, that mining card will be based on the TU102, which if you’re familiar with the lingo, is the top-tier variant of the Turing architecture.

As for the CMP 90HX, reports hint at this particular mining card to be based on the current Ampere GPU lineup and more specifically, the GA102. For context, that’s the same GPU used in both NVIDIA’s GeForce RTX 3080 and RTX 3090 graphics cards.

Another interesting point mentioned is that it is likely that the GPUs used to make these CMP cards did not “meet the specifications or standards” required of a GeForce. As such, these cards could effectively be sporting “broken Tensor, RT, TMUs, or ROPs that renders them unfit for the GeForce badge”.