Hulk
Diamond Member
- Oct 9, 1999
- 5,143
- 3,742
- 136
I doubt that Raptor Lake Refresh (if any) would be on Intel 4. If it were, be prepared for disappointment since there might be clockspeed regression.
I agree, no way RLR will be Intel 4. It's currently being "born" from the excellent yields they are getting on Intel 7 with perhaps some small structure changes.
Problem... or expected difficulty. Scaling down past where they currently are is getting massively more difficult. Witness other leading foundaries having their own problems with achieving better total processes and many others just throwing in the towel along the way. As it stands, all of the foundaries on the leading edge are essentially peddling half and quarter steps between node variants to eek out even the tiniest of improvements to fill the long gaps between full nodes. I don't see Intel as being in trouble as much as I see the crowd catching up to them and occassionally gaining a small lead. We may bag on Intel's 10nm process, but it was often within spitting distance or even better than competing nodes for absolute circuit density. When fed reasonable power, it returned reasonable performance that improved with each revision.
If you think Intel is doing so much worse, look at their competition. Samsung has struggled to stay within a full node of TSMC and Intel, underperforming all along the way for years. TSMC is experiencing delays and having to issue less aggressive subnodes on N3 (N3e being one) to keep customers happy. Who else is anywhere near these three?
Intel's schedule is, as usual, overly optimistic. They aren't falling behind on process tech though.
Good points and they lead me to the conclusion that as we approach the "zero limit" of node reduction we're not so much talking about numbers in terms of structure features but more how the process advances in terms of voltage reduction at given frequencies compared to previous nodes as well as increases in transistor density. As such Intel is probably on the right track by simply giving names to future nodes indicating that lower numbers are "better."
Let's look at 10nm with Intel. Cannon topped out at 3.2GHz and most likely had terrible yields. Raptor Lake, which is still technically 10nm is now hitting 6GHz. Same "node" almost twice the frequency. That's better than any node shrink has ever achieved.