AMD Has Plans To Create The ‘Starship’ Processor With 48 Cores / 96 Threads On The 7nm Node
Fudzilla reported on something pretty interesting a while back - and we never got around to covering it because it was about something that was pretty far out into the horizon. However, in light of recent announcements from AMD regarding the amended WSA agreement, I thought now would be a good time to finally go over it. I am talking, of course, about AMD's plans for a 48 core processor based on the 7nm process that has been code named Starship. Since all of this is basically just leaked paperwork at the moment, it would be prudent to take it with more than a pinch of salt.
AMD possibly planning to skip the 10nm process - 7nm Server processor codenamed 'Starship' planned for 2018
AMD is currently working on its 14nm based Zen architecture and the showcase product on the server side will be the 32 core Naples processor. However, they already have plans for a 48 core chip that will be based on the 7nm process - and since the company will continue to use an SMT (Simultaneous Multi Threading) based approach to its architecture - we are looking at a very impressive count of 96 threads. Considering the fact that Starship will be based on the 7nm process not only will it be more economically viable (in terms of cost effectiveness and $ per transistor) than the 14nm implementation of Zen but it will also be much more power efficient.
If AMD continues its value philosophy forward with Starship, then it can eventually go on to pose a significant challenge to Intel Xeon and Xeon Phi chips. The Xeon Phi co-processors will be the only offerings from Intel which will out strip the Starship in terms of the raw core count. As is usually the case with chips from the Blue giant however, they are very pricey and when it comes down to it, the use cases that these chips are targeting can be achieved just as easily by employing more than one processor as long as it is cost effective and power effective.
Now this leak as well as AMD's recent WSA announcement has one very interesting implication. It would appear that the company is going to be skipping the 10nm process. The amended agreement covers only 14nm and also 7nm chips and the fact that Starship is thought to be coming soon after the Naples platform both suggest the chip designer will be skipping the 10nm node. It also lends credence to this starship leak since AMD has already included 7nm processors in its amended WSA agreement. Meaning, it plans to roll them out sometime before the amendment expires - probably around the late 2018 to 2019 time frame.
It is worth adding at this point, that this could turn out to be completely inaccurate and simply a matter of omission due to 10nm node being not part of the contract - that said however, I must admit that the evidence currently points to AMD moving directly to 7nm. As is usually the case with such large processors, there will be smaller versions of the same (yield theory dictates that) and according to the source the TDP should vary between 35W and 180W for the entire range of servers. It is also not clear at this point whether AMD will be be using the Zen architecture for the Starship processor or whether it will be a new revision.
One thing is clear though, the 2018 time frame is probably a best case scenario. As the process node is shrinking in size - it is getting harder and harder to beat physics. If the 7nm FinFET process is anything like past nodes, it will be based on a 10nm backbone, and EUV lithography (which is technically good till 5nm) should be able to take us that far. Even then, late 2018 is the earliest we can expect any working x86 processor on the 7nm node to materialize otherwise we are looking at the same 2019 estimates for 7nm chips.