Intel Abandoning 10nm After DG1, Planning To Use TSMC’s 6nm And 3nm For Next-Generation Xe GPUs
A relatively obscure Taiwanese publication is exclusively reporting that Intel is planning to ditch its own 10nm process immediately after the DG1 GPU and shift over to TSMC's process for the fabrication of its next-generation parts. The jump will happen by 2021 and utilize TSMC's 6nm process at the start. The transition to TSMC for GPU production is something that has been rumored before and just like before, we do not have independent validation of this information - so this post has been marked rumor and to be taken with a grain of salt.
Report: Intel tapping TSMC's sub 7nm processes for the production of high-end Xe GPU flagships
Intel will be holding its first GPU trials with the DG1 this year and this is a graphics card that is based on the 10nm process. Consuming roughly 25W of power and performance levels higher than an NVIDIA MX 250, the DG1 is essentially the TGL iGPU in a discrete form factor. Featuring 96 EUs it is clear that Intel just wants to get its feet wet and learn the ropes before it moves on to more ambitious variants like Xe HP. Interestingly, however, Intel's CFO has previously admitted that 10nm yields aren't great and will actually be lower in profitability than their older 22nm process so I don't see Intel tapping the node for high volume production of a mainstream GPU.
Before we go any further, here is a translated version of the relevant portion of the article:
Intel expects to hand over its self-developed independent display GPU to TSMC's 6-nanometer manufacturing process, and it will also use TSMC's 3-nanometer manufacturing process by 2022. Intel's Chief Financial Officer George Davis previously stated publicly that in response to the insufficient capacity of the 14nm process, Intel will increase more capacity to fill the gap in 2020. As for the part of the 10nm process, it is not expected that it will be produced on a large scale as the 22nm or 14nm process. In the face of more new products coming out in the future, the problem of advanced processes must be solved.
According to the report, relevant sources point out that Intel will begin to produce its own research GPUs and chipsets using TSMC's 6nm EUV process in 2021. The reason why TSMC is used to produce GPUs is mainly because the production of GPUs is relatively simpler than the CPU, and TSMC is also very experienced in GPU production. According to official Intel data, Intel ’s self-developed Xe architecture DG1 independent display GPU uses its own 10nm process technology and will be available at the end of 2020. It has 96 sets of execution units with a total of 768 cores, a base frequency of 1GHz, an acceleration frequency of 1.5GHz, and 1MB. Secondary buffer memory and 3GB display memory, maximum power consumption is 25W. In terms of external performance, the Xe-based DG1 alone will be between Nvidia GeForce GTX 1050 and GeForce GTX 1650.
Although DG1 positioning is not high-end, there will be new generations of DG2 independent display GPU products in the future. It is estimated that DG2 will be a high-performance GPU. According to previous market news, Intel DG2 will use TSMC's 7nm process technology, but according to market news, it may eventually adopt TSMC's 6nm EUV process. However, Intel expects to mass-produce its 7nm process technology in 2021, and officials have already announced that the Ponte Vecchio acceleration card for data centers will use its own 7nm process. At that time, whether Intel and TSMC can really join forces to record a new page for Intel's independent display GPU production, which needs to be further developed. Source: Tech News Taiwan
Intel's 10nm problems places the GPU architects in a very unique conundrum. While 7nm is likely going to be a breeze since the company is planning to shift to EUV (and reset the difficulty curve), 10nm is going to be very hard to squeeze in terms of high volume production. Considering TSMC has a past fabricating GPUs and they are (relatively) easier to fab then CPUs, it seems like the company might actually have a perfectly rational reason to ditch Intel TMG for TSMC.
Also, keep in mind that demand for CPUs is through the roof, and with manufacturing running at 100% and still being nowhere close to exceeding demand, it wouldn't make sense to put the strain of mainstream GPUs produced in high volume when you can offload the same to TSMC. That said, the company has confirmed that Ponte Vecchio will be manufactured on the 7nm process, so either Intel is going to come up with a very flexible arrangement tapping both TMG and TSMC or this rumor is all fluff.