Howdy, Stranger!
It looks like you're new here. If you want to get involved, click one of these buttons!
Quick Links
Intel Rocket Lake is a backport of 10 nm hardware to 14 nm
https://www.anandtech.com/show/16205/intels-11th-gen-core-rocket-lake-detailed-ice-lake-core-with-xe-graphics
In a move that may be unprecedented in the whole history of computing, Intel is backporting Sandy Cove (Ice Lake) CPU cores and their Xe GPU architecture used in Tiger Lake from 10 nm to 14 nm. They'll launch it as Rocket Lake on their 14 nm process node in Q1 2021.
So why would Intel do this? In going from Sky Lake to Kaby Lake (Sky Lake Refresh) to Coffee Lake (Sky Lake Refresh Refresh) to Coffee Lake Refresh (Sky Lake Refresh Refresh Refresh) to Comet Lake (Sky Lake Refresh Refresh Refresh Refresh), Intel has squeezed pretty much everything that they could have out of their 14 nm process node. They've gone with a larger die for more CPU cores, and improved per-core performance by increasing clock speeds as the expense of blowing out the power budget.
That has allowed Intel to steadily improve performance even without the benefit of a new process node or a new architecture, which are the two ways that you usually improve performance. It has also allowed Intel to hang onto the best single-threaded CPU performance even as AMD has passed them by in just about any other metric of CPU goodness. But with that single-threaded advantage due to turn into a deficit next week with the launch of AMD's Zen 3 CPUs, Intel had to do something or else accept being relegated to an inferior, budget option.
Intel does have a newer process node, as well as newer architectures. Ice Lake and Tiger Lake use very similar CPU cores, and Xe is their new GPU. They're all designed for Intel's 10 nm process node. Ice Lake launched about a year ago, and Cannon Lake "launched" more than a year before that. Normally, that would mean that the 10 nm process node is very mature by now, all products that need a leading edge process node would already be on 10 nm, and Intel would be preparing to move to 7 nm.
But it's not just that Intel isn't imminently going to move to 7 nm. Their 10 nm node is still dreadful, so much so that they don't want to actually use it for CPUs if they don't absolutely have to. So they're taking their 10 nm parts and backporting them to 14 nm, which will use a lot more power and die space for the same design. Building Rocket Lake on 14 nm will surely cost a lot more to build each die than keeping everything on 10 nm.
So why would Intel do that? The most plausible answer is that yields on 10 nm are still awful. If it costs 50% more to build each die even to make an inferior product, but you have to throw most of the 10 nm dies in the garbage, then it can be cheaper per good die on 14 nm, where you can get good yields. If 10 nm is not in a usable state more than two years after its nominal launch, it's fair to wonder whether it will ever be a worthwhile process node. Intel's 7 nm node can't come soon enough, but apparently that's delayed, too.
The reason backporting launched products that need the best process node they can get to a much older process node is likely unprecedented is that Intel's travails on their 10 nm node are comparably unprecedented. I'd say that this is how Moore's Law dies except that TSMC and Samsung don't seem to be having quite the same problems, which is why Intel has been passed by AMD.
We don't know how high Rocket Lake will be able to clock, but it might be able to wrest the single-threaded CPU performance crown back from AMD's Zen 3 CPUs. Or it might not. Intel is claiming "Double Digit IPC Instructions Per Clock (IPC) improvement", whatever it means when you untangle that mess of typos. So single-threaded performance could go up even if clock speeds go down. But they're claiming 8 cores rather than 10, while still using the same power budget as before. On net, Rocket Lake will probably be slower than Comet Lake in well-threaded code--while both are already far behind AMD's Zen 2 there, and will be further behind Zen 3.
Intel is also porting the Xe GPU architecture back to 14 nm, though that doesn't matter much, unless you actually care about Intel GPUs. Most people who care about a good GPU buy a discrete card, and most people who specifically want a good integrated GPU buy AMD.
The other notable improvement is that the Rocket Lake platform will have 20 PCI Express 4.0 lanes. So in 2021, Intel will finally offer PCI Express 4.0, after AMD got there in mid-2019. And Intel will still offer fewer lanes of it than AMD does on mainstream consumer parts, so they'll still be trailing behind AMD, but just not nearly as far behind. But that's still progress, right?
At this point, one has to wonder how much longer Intel will remain in the foundry business. Unless they can deliver a competitive, new process node, they'd be better off just getting out (or perhaps rather, continuing to use 14 nm for products such as chipsets that don't need a cutting edge process node) and building their CPUs at TSMC or Samsung. That would be quite the earthquake in the industry, but if they can't deliver a competitive process node and don't shift production to a foundry that can, they're headed for bankruptcy. They've already announced that they're moving at least one discrete GPU to be produced at TSMC. We'll see if that's just a temporary stopgap or the first of many.
In a move that may be unprecedented in the whole history of computing, Intel is backporting Sandy Cove (Ice Lake) CPU cores and their Xe GPU architecture used in Tiger Lake from 10 nm to 14 nm. They'll launch it as Rocket Lake on their 14 nm process node in Q1 2021.
So why would Intel do this? In going from Sky Lake to Kaby Lake (Sky Lake Refresh) to Coffee Lake (Sky Lake Refresh Refresh) to Coffee Lake Refresh (Sky Lake Refresh Refresh Refresh) to Comet Lake (Sky Lake Refresh Refresh Refresh Refresh), Intel has squeezed pretty much everything that they could have out of their 14 nm process node. They've gone with a larger die for more CPU cores, and improved per-core performance by increasing clock speeds as the expense of blowing out the power budget.
That has allowed Intel to steadily improve performance even without the benefit of a new process node or a new architecture, which are the two ways that you usually improve performance. It has also allowed Intel to hang onto the best single-threaded CPU performance even as AMD has passed them by in just about any other metric of CPU goodness. But with that single-threaded advantage due to turn into a deficit next week with the launch of AMD's Zen 3 CPUs, Intel had to do something or else accept being relegated to an inferior, budget option.
Intel does have a newer process node, as well as newer architectures. Ice Lake and Tiger Lake use very similar CPU cores, and Xe is their new GPU. They're all designed for Intel's 10 nm process node. Ice Lake launched about a year ago, and Cannon Lake "launched" more than a year before that. Normally, that would mean that the 10 nm process node is very mature by now, all products that need a leading edge process node would already be on 10 nm, and Intel would be preparing to move to 7 nm.
But it's not just that Intel isn't imminently going to move to 7 nm. Their 10 nm node is still dreadful, so much so that they don't want to actually use it for CPUs if they don't absolutely have to. So they're taking their 10 nm parts and backporting them to 14 nm, which will use a lot more power and die space for the same design. Building Rocket Lake on 14 nm will surely cost a lot more to build each die than keeping everything on 10 nm.
So why would Intel do that? The most plausible answer is that yields on 10 nm are still awful. If it costs 50% more to build each die even to make an inferior product, but you have to throw most of the 10 nm dies in the garbage, then it can be cheaper per good die on 14 nm, where you can get good yields. If 10 nm is not in a usable state more than two years after its nominal launch, it's fair to wonder whether it will ever be a worthwhile process node. Intel's 7 nm node can't come soon enough, but apparently that's delayed, too.
The reason backporting launched products that need the best process node they can get to a much older process node is likely unprecedented is that Intel's travails on their 10 nm node are comparably unprecedented. I'd say that this is how Moore's Law dies except that TSMC and Samsung don't seem to be having quite the same problems, which is why Intel has been passed by AMD.
We don't know how high Rocket Lake will be able to clock, but it might be able to wrest the single-threaded CPU performance crown back from AMD's Zen 3 CPUs. Or it might not. Intel is claiming "Double Digit IPC Instructions Per Clock (IPC) improvement", whatever it means when you untangle that mess of typos. So single-threaded performance could go up even if clock speeds go down. But they're claiming 8 cores rather than 10, while still using the same power budget as before. On net, Rocket Lake will probably be slower than Comet Lake in well-threaded code--while both are already far behind AMD's Zen 2 there, and will be further behind Zen 3.
Intel is also porting the Xe GPU architecture back to 14 nm, though that doesn't matter much, unless you actually care about Intel GPUs. Most people who care about a good GPU buy a discrete card, and most people who specifically want a good integrated GPU buy AMD.
The other notable improvement is that the Rocket Lake platform will have 20 PCI Express 4.0 lanes. So in 2021, Intel will finally offer PCI Express 4.0, after AMD got there in mid-2019. And Intel will still offer fewer lanes of it than AMD does on mainstream consumer parts, so they'll still be trailing behind AMD, but just not nearly as far behind. But that's still progress, right?
At this point, one has to wonder how much longer Intel will remain in the foundry business. Unless they can deliver a competitive, new process node, they'd be better off just getting out (or perhaps rather, continuing to use 14 nm for products such as chipsets that don't need a cutting edge process node) and building their CPUs at TSMC or Samsung. That would be quite the earthquake in the industry, but if they can't deliver a competitive process node and don't shift production to a foundry that can, they're headed for bankruptcy. They've already announced that they're moving at least one discrete GPU to be produced at TSMC. We'll see if that's just a temporary stopgap or the first of many.
![[Deleted User]](http://wa.vanillicon.com/v2/a5231d9a9017cdf5f82125ccaf807f8f.svg)
