Howdy, Stranger!
It looks like you're new here. If you want to get involved, click one of these buttons!
Quick Links
Intel launches Ice Lake SP, gets crushed by AMD's EPYC Milan
For as much as Intel and AMD like to sell desktop CPUs for $300 each, they really like to sell server CPUs for $4000 each. With Comet Lake and then Rocket Lake, Intel has managed to be kind of competitive in desktop performance at the expense of blowing out the power budget. But you can't do that in servers, as if AMD is willing to burn 280 W, Intel can't just burn 500 W per socket and call it a day. That's too impractical to cool, especially when you want racks full of servers.
Even so, Ice Lake SP has one big advantage that Rocket Lake doesn't: it's on a newer 10 nm process node, rather than Intel's aging 14++*^#% or whatever they're calling it these days. That offers some major efficiency advantages both in performance per watt and performance per mm^2, if you can get reasonable yields. It's probably still behind TSMC 7 nm, but it's not nearly as far behind as Intel's 14 nm node is.
As such, Intel was able to make some major improvements over their previous generation Cooper Lake... err wait, Cooper Lake was mostly canceled. ... as compared to their previous generation Cascade Lake... wait, Cascade Lake was basically just a refresh of Sky Lake. Well anyway, Intel was able to offer some major improvements over their previous generation, whatever it was.
For example, AMD had 64 x86 CPU cores per socket, and Intel only had 28. Now Intel has 40. Which is still less than 64, but at least a two socket Intel server can finally beat a one socket AMD server.
Meanwhile, AMD had 128 PCI Express 4.0 lanes, while Intel had 48 PCI Express 3.0 lanes. Now Intel offers 64 PCI Express 4.0 lanes. Which is still fewer than 128, but hey, progress. Though to be fair, most servers don't need a ton of PCI Express connectivity.
Intel actually did catch up in some ways. They offered up to 1.5 TB of memory per socket, while AMD offered 4 TB. Now Intel also offers support for 4 TB of memory per socket. AMD had 8 DDR4 memory channels, while Intel had six. Now Intel also has 8 DDR4 memory channels.
The upshot is that Ice Lake SP is competitive with EPYC Milan in single-threaded CPU performance, for those people who buy 40 core CPUs because they care primarily about single-threaded performance. It's really not competitive in highly threaded CPU performance, because 40 cores is still a lot fewer than 64. Oh, and it's also more expensive than EPYC Milan.
Intel still offers huge, monolithic dies, which makes it a lot more efficient to transfer massive amounts of data between random CPU cores than AMD's chiplet approach. So there's still that, even if Milan has greatly reduced the inter-die communication problems that EPYC Naples had. This may be a contributing factor to Ice Lake SP having much lower idle power consumption than EPYC Milan. Intel is going chiplets (well, Intel calls them "tiles"), too, but hasn't done so yet.
Even so, Intel still has most of the x86 server market, partially because of inertia. Additionally, AMD has focused on the higher end of the server market, and doesn't really have a competitive product for low end servers. You can get a lower end Intel server with 8 or 10 cores, suitably low price and power consumption, and common server features like ECC memory. AMD's approach with the huge I/O die just doesn't lend itself to that very well. But AMD is gaining on Intel, and the longer that AMD has clearly superior products, the more market share they're likely to claim.
Even so, Ice Lake SP has one big advantage that Rocket Lake doesn't: it's on a newer 10 nm process node, rather than Intel's aging 14++*^#% or whatever they're calling it these days. That offers some major efficiency advantages both in performance per watt and performance per mm^2, if you can get reasonable yields. It's probably still behind TSMC 7 nm, but it's not nearly as far behind as Intel's 14 nm node is.
As such, Intel was able to make some major improvements over their previous generation Cooper Lake... err wait, Cooper Lake was mostly canceled. ... as compared to their previous generation Cascade Lake... wait, Cascade Lake was basically just a refresh of Sky Lake. Well anyway, Intel was able to offer some major improvements over their previous generation, whatever it was.
For example, AMD had 64 x86 CPU cores per socket, and Intel only had 28. Now Intel has 40. Which is still less than 64, but at least a two socket Intel server can finally beat a one socket AMD server.
Meanwhile, AMD had 128 PCI Express 4.0 lanes, while Intel had 48 PCI Express 3.0 lanes. Now Intel offers 64 PCI Express 4.0 lanes. Which is still fewer than 128, but hey, progress. Though to be fair, most servers don't need a ton of PCI Express connectivity.
Intel actually did catch up in some ways. They offered up to 1.5 TB of memory per socket, while AMD offered 4 TB. Now Intel also offers support for 4 TB of memory per socket. AMD had 8 DDR4 memory channels, while Intel had six. Now Intel also has 8 DDR4 memory channels.
The upshot is that Ice Lake SP is competitive with EPYC Milan in single-threaded CPU performance, for those people who buy 40 core CPUs because they care primarily about single-threaded performance. It's really not competitive in highly threaded CPU performance, because 40 cores is still a lot fewer than 64. Oh, and it's also more expensive than EPYC Milan.
Intel still offers huge, monolithic dies, which makes it a lot more efficient to transfer massive amounts of data between random CPU cores than AMD's chiplet approach. So there's still that, even if Milan has greatly reduced the inter-die communication problems that EPYC Naples had. This may be a contributing factor to Ice Lake SP having much lower idle power consumption than EPYC Milan. Intel is going chiplets (well, Intel calls them "tiles"), too, but hasn't done so yet.
Even so, Intel still has most of the x86 server market, partially because of inertia. Additionally, AMD has focused on the higher end of the server market, and doesn't really have a competitive product for low end servers. You can get a lower end Intel server with 8 or 10 cores, suitably low price and power consumption, and common server features like ECC memory. AMD's approach with the huge I/O die just doesn't lend itself to that very well. But AMD is gaining on Intel, and the longer that AMD has clearly superior products, the more market share they're likely to claim.
Comments
I would say that a company that chooses a server brand on the basis of perceived reliability of client desktop processors is doing something severely wrong. When a desktop, laptop, or server fails, it's pretty rare for the CPU dying to be the culprit.