Update 9/21/21 5:30am PT: Intel reached out to clarify Raja's response, which was misquoted in the referenced article. We have amended the text accordingly:
Raja Koduri, Intel's Senior Vice President and General Manager of the Accelerated Computing Systems and Graphics (AXG) Group, recently shed some light on its decision to contract TSMC to manufacture its ARC GPUs on its 6nm process.
"It is necessary to first determine the process that can be assumed at the start of design," said Raja. "Other features, such as how much operating frequency can be used are also important factors. Cost is also an issue. These three, that is, the cost-performance-capacity, is taken into consideration when deciding which process to use. "
Intel concluded that the first product, Alchemist, happened to find the best balance on TSMC's N6 node. And of course, it can't be understated that due to Intel's series of process delays, competing with AMD or NVIDIA with a new product on a not-leading-edge node could certainly bite into the performance potential for ARC.
The semiconductor manufacturing market has been hit repeatedly over the last 18 months with supply issues, stemming mostly from the complexity of its long supply chain. However, Intel seems to have escaped relatively unscathed, mostly owing to its vertical strategy. Intel designs and manufactures its CPUs fully in-house, giving it more flexibility than so-called "fabless" manufacturers (such as AMD) to prioritize its manufacturing lines and adapt to shifting market conditions.
The complex equation to choose a certain node also encompasses other aspects. For example, the CPU cores for AMD's Ryzen 7 5800X CPU (8-cores, 16-threads) feature a roughly 80.7 mm² surface area on TSMC's 7 nm process for a total of 4.15 billion transistors. AMD's top GPU offering, the RX 6900XT, manufactured on the same process, absolutely dwarfs that at more than six times the die size (520 mm²) and 26.8 billion transistors. In other words, GPUs are in another league, in general, when it comes to silicon manufacturing simply because they are larger, resulting in fewer chips per wafer.
Of course, opting for TSMC also means there's less capacity for its rivals to produce GPUs and CPUs on the best manufacturing nodes available, as well. So moving GPU production to TSMC means Intel will be able to produce more devices in a silicon-hungry market while simultaneously assuring its competition can't make as many chips as they could sell. That's one way to beat the competition.
Of course. The example was given because A) Intel's GPUs will be manufactured by TSMC; B) I could show the difference between a pure CPU on TSMC's 7 nm and a GPU on the same manufacturing node. I think this is the best way to showcase the pure complexity difference between both. Don't you agree?
And I agree, the potential for a complete market failure is not unknown to Intel. Having this built on someone else's fab means they lose a lot less if it completely fails to gain any market traction. Which this first round may have a lot of trouble with.
Unless they get the drivers right this could all be a waste of time.
Though like what AMD seems to be planning, slapping one of those dies into a true Intel APU would not be beyond them (heck they might have that in mind, Intel has used Vega once and it was pretty decent) Intel 8+4 + 128 EU, that would be something else.
Intel has huge amounts of clout which makes it very difficult for TSMC or anybody else to simply tell them no. In the short term, sending Arc over to TSMC is like an 8x win-win-win for Intel. Even if Arc is an under-performer in the short term this will end up constraining GPU availability even further ensuring that every single Arc sells out. The one that's available is the one you'll buy - somebody will.
In the long run once Intel gets their fabs up to date and good enough to build GPUs this will be a good thing and free up massive amounts of GPU manufacturing capacity.
In the mean time, its a good thing that Nvidia (and seemingly AMD is following now) is using failed GPUs re-spun as miner cards instead of simply throwing them in the trash. Hopefully Intel eventually does the same and we get three brands of miner cards out of it all instead of just CMP HX.
The decision to build Xe-HPG at TSM was probably made three years ago, before Intel had their 10nm yield issues solved.
I say "solved" because Intel stated recently that they are now building more 10nm chips than 14nm. Also, the 10nm chips they are building include server chips with relatively high core counts.
Intel builds other chips at TSM. They have made statements about 20% of their volume previously. We know about the eyeq5, habana gaudi and perhaps the easic chips. Now that they are more open about the GPUs being built there, maybe there will be more visibility about the other products.