Intel already offering reference design to board partners, custom designs planned
Intel’s Senior VP and GM of Graphics Group, Raja Koduri, has been interviewed by a Japanese site ASCII.
Intel chose TSMC N6 due to the insufficient capacity of its own nodes
Raja answered a number of questions in regard to the upcoming Xe-HPG architecture which is now set to debut in the first quarter of 2022 under the Arc Alchemist series. Raja was asked why has Intel decided to use TSMC N6 process technology instead of Intel’s own ‘7’ node. According to his response, Intel’s manufacturing capacity simply wasn’t ready to mass-produce discrete GPUs. Intel clearly does not want its discrete GPUs to compete with CPU series, such as Alder Lake, for capacity.
It was necessary to first determine the manufacturing capacity of the process that can be assumed at the start of design, and (Intel’s) advanced process did not have sufficient capacity yet.
— Raja Koduri
Intel has not yet determined whether Alchemist’s successor will rely on Intel own’s or external foundry. However, the company might want to switch to more advanced nodes once NVIDIA Lovelace and AMD RDNA3 series launch using TSMC N5 process.
Intel’s supersampling technology (XeSS) will be compatible with Xe-LP architecture
Without much of a surprise, Raja reaffirmed what was already announced by Intel. Existing Xe-based architectures such as Xe-LP will support XeSS technology, Intel’s own version of super-resolution based on artificial intelligence algorithms. The Xe-LP version is likely to use DP4a instruction, considered backward compatible with all new GPUs. The DP4a version of XeSS is not as efficient as XMX based, but it should be an interesting addition to these low-power graphics nonetheless.
By the way, regarding the super sampling technique XeSS, it was said that it was “backward compatible”. In other words, the Iris Xe DG1 that has already been released can also use the XeSS. It was also said that it would work on 11th generation based GPUs including Tiger Lake.
— ASCII quoting Raja Koduri
Ponte Vecchio’s Xe-Link will not be used by the HPG series
Another interesting tidbit is that Raja did not explicitly confirm whether the Xe-HPG series will support multi-GPU, such as NVIDIA SLI and AMD Crossfire. Both technologies are pretty much dead at this point, not only because of subpar performance but also due to the fact that multiple GPUs from the same vendor are expensive and hard to obtain. Intel clearly has no intention to compete in the declining multi-GPU market.
Custom cards and workstation Arc series planned?
By far the most interesting part of the interview is the question about custom card launch based on Arc Alchemist GPUs. As revealed earlier through leaks, Intel is already offering its reference design to board partners. Intel had actually confirmed this leak during its drone show. Exactly the same design, which was first shown by Moore’s Law is Dead, is now expected to be offered to board partners who may or may not use it for their cards, alternatively, they might want to adopt it for their own semi-custom versions.
Partners and I think there will be a differentiation of ODM, and that will lead to the ultimate customer interest
— Raja Koduri
When asked about the possibility of Intel launching Intel Arc workstation cards, an equivalent of NVIDIA (Quadro) and AMD Radeon Pro, Koduri only mentioned that Alchemist can support the workstation market and high-end applications such as 3DX Max. While this does not ultimately confirm Intel will launch its own “Arc Pro” series, there is really no reason not to launch any, unless Intel is devoting its whole TSMC capacity for gamers.
Intel DG2-512EU engineering sample, Source: Moore’s Law is Dead