AMD has relied on chipmakers GlobalFoundries and TSMC for well over a decade now. Still, if a new report from J.P. Morgan turns out to be correct, the company might build entry-level 4nm processors for Chromebooks at Samsung Foundry. The firm also speculates that AMD could also use Samsung foundry for future GPUs.
"Our research indicates that AMD is likely to outsource a Chromebook CPU to Samsung on its 4nm (likely mass production in late 2022), and TSMC may have limited capacity to allocate for Chromebook projects, given the declining market demand," Gokul Hariharan, an analyst with J.P. Morgan, wrote in a note to clients shared by @MarcTheShark83.
AMD relied almost exclusively on GlobalFoundries at first, and then on TSMC, for two reasons. Firstly, AMD had a wafer supply agreement with GlobalFoundries that prevented it from using TSMC and other foundries extensively. Secondly, AMD controlled its development costs by re-using IP blocks tailored for a particular chip technology, greatly lowering the number of nodes it actually uses.
While AMD's profitability is now at historically high levels, the company continues to be very conservative in adopting new manufacturing processes. That's because using familiar IP not only reduces costs but also helps with yields.
If AMD proceeds with the plan, using Samsung Foundry's nodes even for some of its products will significantly change how AMD designs and builds its products.
Adopting Samsung Foundry's 4LPP (second-generation 4nm-class process) for a Chromebook-bound APU would require AMD to redesign its CPU and GPU IP for the new manufacturer, something that is rather costly at 4nm. Also, keeping in mind that the Chromebook market isn't growing, it's unclear how financially viable the project would be. The analysts believe that AMD could also use Samsung Foundry for some of its GPUs going forward, but didn't elaborate.
"In addition, we believe that AMD is likely to evaluate some projects (probably for GPU) at Samsung 3nm in 2023/2024, but vast majority of its core platforms (CPUs for server, mobile, and desktops) are likely to stay with TSMC N3." said the report.
Meanwhile, J.P. Morgan believes that AMD will continue to implement new manufacturing technologies rather conservatively and will not be among the first to start using leading-edge nodes. The analysts expect AMD to adopt TSMC's N3 technology (likely N3E, actually) only for products that will ramp in 2024.
"Note that AMD's N3 devices are likely to ramp up only in 2024, given that its N5 projects (Genoa and Zen 4 desktop variants) will only come to market in 2H 2022," the note reads.
IO dies are crap for freq and power usage on Ryzen
I think 5nm,3nm are already secured by everybody else but AMD. This is on AMD, they are really too slow for CPU releases.
Given that, I suspect that they are going to be releasing an ARM-based Chromebook SOC using some Samsung IP blocks along with the GPU IP blocks they are developing in partnership with Samsung. And when you think about it, that makes a lot of sense, since AMD has a lot of experience with doing laptop mobile GPUs. So it makes sense that they do a laptop SOC first, ie H1 2022 while they're on the road to developing a good phone GPU.
So yeah, I think the assumption that they're re-working Zen for Samsung's 4nm process is wildly incorrect.
Well, 2020 actually was an extremely good year for chromebooks.
i don't think GF even have 7nm process. let alone 5nm. so even if you want to put your low end device to be manufactured by GF it is something they cannot do.
AMD have some capacity at TSMC. but most likely not enough so they need to look for other foundries. and the competition to get TSMC capacity most likely high as well. want more capacity? then you probably need to bid the price against others as well.
Correct, IBM was suing GloFo in relatively recent news for being 2 years late to 14nm and abandoning 10nm when those nodes had been contracted as part of the deal of IBM selling their fabs to GloFo. GloFo in turn said that IBM had colluded against them by purchasing wafer starts elsewhere, which starved them of the resources to do the work, so they felt their non-performance of what had become a frustrated contract was perfectly legal. I don't remember how or whether this has been settled, but I know for sure GloFo's best node is 14nm.