Intel’s Cascade Lake-AP Xeon To Bring 48 Cores & 12 Channel Memory In A Multi-Chip Package

Intel has released tidbits of information regarding its next generation server architecture, Cascade Lake, for a few months. It builds off of the Xeon Scalable Skylake family, with most of its focus on security and compute. However, there are now going to be two Cascade Lake families, with the latest to be announced called Cascade-AP, meaning ‘advanced performance’.

Cascade-AP is strictly a dual socket solution at this time, with no word on single or quad socket implementations, which does seem a little odd for such an advanced architecture. The headline encompasses basically all that is known at this time, since Intel is being very secretive about certain specifics.

There is no flagship model number given, but we were told that there will be a 48-core CPU, with 12 channels of DDR4 memory – quite a step-up from the current flagship CPU, the 8180 with its 28 cores. We can probably assume that hyperthreading will be enabled for these, with a total of 96 threads available, but no confirmations yet.

What makes this particular chip interesting is Intel is not going with a monolithic CPU, but is instead opting for a multi-chip package (MCP). This isn’t the first time Intel has gone with an MCP design, since we can go all the way back to the days of the Pentium D for that. The irony is not lost on us here, as Intel called out AMD’s recent designs as being ‘glued together’. What is strange is that Intel did not go with its EMIB design, but instead will rely on the UPI interface for the two dies to communicate (the same interface used for its multi-socket interconnect solutions).

Beyond this though, there is very little information. The 12 channel memory is per socket, but the internal configuration of the dies is unknown. It’s probably safe to say that each die would be responsible for 6-channels and use the UPI interface to share the channels. What happens between sockets though, is unknown.

A big question is hanging over what socket would be used. The current 28-core 8180 makes use of 6-channel memory and the LGA3647 socket. Effectively doubling up the core count and channels would likely put a strain on the existing socket, unless some kind of compromise was made. This becomes more apparent when power is taken into consideration. The 8180 has a TDP of 205 Watts, and it seems unlikely that Intel would release a 400 Watt TDP processor, so clock speeds are likely to take a significant hit as well. Even with a mature 10nm manufacture process to reduce power, it would be difficult. There is a rumor of an LGA5903 socket for Cascade Lake, but unconfirmed at this time.

There was no word on the number of supported PCIe lanes, total memory capacity, or operating frequencies, but Intel did state some performance figures for the new Cascade-AP CPUs. Linpack was 3.4x faster and stream triad was up to 1.3x faster compared to an AMD EPYC 7601. There was mention of a large boost to deep learning capabilities with 17x the performance in image inferencing compared to the Skylake-based Xeons, mostly the result of a new instruction set.

Widespread availability of Cascade-AP won’t be until first half of next year, but Intel is already sampling the chips to select early access partners.