Howdy, Stranger!
It looks like you're new here. If you want to get involved, click one of these buttons!
Quick Links
Intel's server roadmap is brutal.
The rumors of additional, severe delays had been out there for quite some time. Intel just made it official.
https://www.anandtech.com/show/13194/intel-shows-xeon-2018-2019-roadmap-cooper-lakesp-and-ice-lakesp-confirmed
First off is yet another generation of server parts on 14 nm, shipping in Q4 of this year. After that is still another generation of servers on 14 nm. The first server parts on a new process node won't come until 2020, and even that is assuming things go well for Intel. Considering how averse server buyers are to launching a new generation every several months, don't bet on early 2020 for Ice Lake, either.
Yes, I know Intel wants to say 14++ nm or whatever. And process nodes do improve some as they mature. But not nearly enough to be competitive with a real die shrink. That's why the state of the art isn't on a 90++^*&@ nm process node.
Intel is hoping to get the first real 10 nm Cannon Lake parts (as opposed to the press edition parts from earlier this year) out by the end of 2019. But a tiny dual core CPU together with a highly salvageable GPU is far, far away from the huge dies they use for servers. Even if Cannon Lake does launch as a nifty product that you can buy in time for Christmas of 2019, that would hardly point to wide availability for 28 core Ice Lake parts in 2020. (28 cores is the current top of the line for Sky Lake.)
Meanwhile, AMD claims that they're moving to 7 nm server parts in 2019. Apparently now they're going to use both TSMC and Global Foundries. In the halcyon days of the Opteron 64, AMD could sell their CPUs as fast as they could make them, but they just couldn't make enough to really take a ton of market share. Having both TSMC and Global Foundries available should mean that this time, fab production limits aren't such a concern.
It's possible that Intel smacked into a wall of physics and TSMC and Global Foundries are in the process of smacking into the same wall. There don't seem to be rumors to that effect, but you don't really know until there are parts that you can buy. But if AMD has their third generation of EPYC parts on 7 nm CPUs working as planned and available in 2019, then 2020 is going to be absolutely brutal for Intel's server division.
That wouldn't merely mean that AMD was competitive with Intel in servers. That would mean that Intel had nothing competitive to offer, and was relegated to being a distant also-ran. That has never happened in the history of servers, but it will next year if AMD delivers what they've promised.
There were claims a while back that AMD had to prioritize where to invest, and chose to starve their GPU division in order to do what they needed to do in CPUs. Since falling behind Nvidia with the launch of the GeForce GTX 980 nearly four years ago, AMD has done a die shrink, some minor tinkering around the edges, and that's about it--with nothing new looking imminent. I don't know how early they had some inkling that Intel was going to fall on their face. But as soon as they knew, that's pretty much why they had to make that decision.
Suppose that AMD had invested heavily in GPUs, for everything from graphics to compute. And suppose that those investments had gone really well. Maybe they could have swiped $1 billion/year from Nvidia in GPU revenue as compared to how things played out now--though you should keep in mind that the cryptocurrency craze meant AMD could easily sell everything they could make for the last year. Or maybe competitive high end compute GPUs would just force Nvidia to slash prices and there would be less revenue to go around in that market. And that's if everything went well for AMD.
Well, it's now looking very plausible that AMD will swipe several billion dollars per year worth of the server market from Intel. Depending on how long it takes for Intel to recover, it's possible that a few years from now, AMD will be the market leader in x86 CPUs for all of the major markets. Don't scoff at the idea that AMD could soon be the market leader in x86 CPUs. Just because Intel has been dominant for many years doesn't mean that it will always be that way. Remember when Blackberry dominated smartphones?
That said, just because Intel has stumbled doesn't automatically mean that AMD will reap the benefits. If ever there was an opportunity for servers to heavily move away from x86, this will be it. ARM has been trying to get into the server market for some years now, and the Cavium Thunder X2 recently brought a credible ARM chip to servers.
https://www.anandtech.com/show/13194/intel-shows-xeon-2018-2019-roadmap-cooper-lakesp-and-ice-lakesp-confirmed
First off is yet another generation of server parts on 14 nm, shipping in Q4 of this year. After that is still another generation of servers on 14 nm. The first server parts on a new process node won't come until 2020, and even that is assuming things go well for Intel. Considering how averse server buyers are to launching a new generation every several months, don't bet on early 2020 for Ice Lake, either.
Yes, I know Intel wants to say 14++ nm or whatever. And process nodes do improve some as they mature. But not nearly enough to be competitive with a real die shrink. That's why the state of the art isn't on a 90++^*&@ nm process node.
Intel is hoping to get the first real 10 nm Cannon Lake parts (as opposed to the press edition parts from earlier this year) out by the end of 2019. But a tiny dual core CPU together with a highly salvageable GPU is far, far away from the huge dies they use for servers. Even if Cannon Lake does launch as a nifty product that you can buy in time for Christmas of 2019, that would hardly point to wide availability for 28 core Ice Lake parts in 2020. (28 cores is the current top of the line for Sky Lake.)
Meanwhile, AMD claims that they're moving to 7 nm server parts in 2019. Apparently now they're going to use both TSMC and Global Foundries. In the halcyon days of the Opteron 64, AMD could sell their CPUs as fast as they could make them, but they just couldn't make enough to really take a ton of market share. Having both TSMC and Global Foundries available should mean that this time, fab production limits aren't such a concern.
It's possible that Intel smacked into a wall of physics and TSMC and Global Foundries are in the process of smacking into the same wall. There don't seem to be rumors to that effect, but you don't really know until there are parts that you can buy. But if AMD has their third generation of EPYC parts on 7 nm CPUs working as planned and available in 2019, then 2020 is going to be absolutely brutal for Intel's server division.
That wouldn't merely mean that AMD was competitive with Intel in servers. That would mean that Intel had nothing competitive to offer, and was relegated to being a distant also-ran. That has never happened in the history of servers, but it will next year if AMD delivers what they've promised.
There were claims a while back that AMD had to prioritize where to invest, and chose to starve their GPU division in order to do what they needed to do in CPUs. Since falling behind Nvidia with the launch of the GeForce GTX 980 nearly four years ago, AMD has done a die shrink, some minor tinkering around the edges, and that's about it--with nothing new looking imminent. I don't know how early they had some inkling that Intel was going to fall on their face. But as soon as they knew, that's pretty much why they had to make that decision.
Suppose that AMD had invested heavily in GPUs, for everything from graphics to compute. And suppose that those investments had gone really well. Maybe they could have swiped $1 billion/year from Nvidia in GPU revenue as compared to how things played out now--though you should keep in mind that the cryptocurrency craze meant AMD could easily sell everything they could make for the last year. Or maybe competitive high end compute GPUs would just force Nvidia to slash prices and there would be less revenue to go around in that market. And that's if everything went well for AMD.
Well, it's now looking very plausible that AMD will swipe several billion dollars per year worth of the server market from Intel. Depending on how long it takes for Intel to recover, it's possible that a few years from now, AMD will be the market leader in x86 CPUs for all of the major markets. Don't scoff at the idea that AMD could soon be the market leader in x86 CPUs. Just because Intel has been dominant for many years doesn't mean that it will always be that way. Remember when Blackberry dominated smartphones?
That said, just because Intel has stumbled doesn't automatically mean that AMD will reap the benefits. If ever there was an opportunity for servers to heavily move away from x86, this will be it. ARM has been trying to get into the server market for some years now, and the Cavium Thunder X2 recently brought a credible ARM chip to servers.
![[Deleted User]](http://wa.vanillicon.com/v2/a5231d9a9017cdf5f82125ccaf807f8f.svg)
Comments
Honestly, 2-3 years ago if you had told
me Crypto would have GPUs sold out across almost all product tiers, I would have laughed you out of the room.
Couple that with 32-core cores going main stream...
And the coup de grace would be Intel fab struggling to keep up.
No no way I would have bet on any one of those things occurring, let alone all three.
AMD played it all perfectly and if they play it well, could come out looking as good as nVidia does - just in a different way.