Intel’s Future Xeon Granite Rapids-SP, Granite Rapids-AP, Sierra Forest CPUs For Birch Stream Platform Detailed

Even more information regarding Intel's Xeon Granite Rapids-SP, Granite Rapids-AP, Sierra Forest CPUs for the Birch Stream platform has been leaked.

Over the last few weeks, we have been getting a serious dose of information on Intel's future Xeon CPU plans. We know for fact that after Emerald Rapids-SP which arrives later this year on the Eagle Stream platform, Chipzilla will be moving to its brand new Birch Stream platform that will offer CPU support for at least two families, the Granite Rapids and Sierra Forest. So let's start with the Granite Rapids family.

https://twitter.com/yuuki_ans/status/1639813047845011456

First up, we have the Intel Granite Rapids-SP Xeon CPUs which are targeting close to 40% higher aggregate performance than the Emerald Rapids-SP CPUs. The Granite Rapids-SP CPUs will utilize the Intel 3 process node and feature anywhere from 80 to 86 cores based on the Redwood Cove architecture. The Granite Rapids-SP CPUs can be configured into 1S, 2S, 4S & 8S configurations & even greater than 8S socket configs using xNC.

The Intel Granite Rapids-SP CPUs will incorporate next-gen Intel Quick Assist Technology, DLB, DSA, IAX, 5G ISA, BFloat16, and AMX-FP16 instruction sets. Each CPU will carry an 8-channel IMC that gets DDR5-6400 (1DPC) and DDR5-5200 (2PC) support. Each CPU will support up to 16 DIMMs per socket while offering up to 4 UPI 2.0 links per chip (x24 Width) with speeds rated at up to 24 GT/s.

In terms of PCIe lanes, each Intel Granite Rapids-SP CPU will get 88 PCIe Gen 5.0 lanes (56 lanes north and 32 lanes south). The company states higher PCIe P2P performance compared to the Sapphire Rapids chips with PCIe bifurcation of x16, x8, x4, and x2 (Gen 5). There will also be 64 CXL 2.0 lanes per CPU that can be used across any four of the five x16 ports. For security, the chips will get TDX (2048 Keys 256-bit keys) and Link protection (PCIe, CXL, UPI).

One significant detail mentioned here is that the Granite Rapids-SP CPUs will be supported on the Socket E2 which features the same pin count as the existing Eagle Stream platform of LGA 4677. It looks like this socket will be able to support up to 350W TDPs though recent info suggests that Intel may have moved their Granite Rapids-SP CPUs over to the LGA 7529 Socket BR too.

In addition to that, the Birch Stream platform will support the latest LAN controllers including Carlsville (2x10 GbE), Columbiaville (2x 100 GbE), Columbiaville SD (1x100 GbE), & Connorsville (200G). That plus there will be the DC-SCM 2.0 management engine along with Aspeed AST2600.

Moving over to the Intel Granite Rapids-AP family, these are the density & performance optimized SKUs for the higher-end Birch Stream platform. The CPUs are said to target 2x perf improvement over Intel's Emerald Rapids-SP (64-Core) CPUs. The chips will get anywhere from 120 to 136 cores based on the Redwood Cove core architecture and feature support on the massive LGA 7529 "BR" socket that we get to see yesterday in full detail.

The Granite Rapids-AP CPUs for the Birch Stream-AP platform will only come in 1S and 2S configs with up to 500W TDP SKUs. Each CPU will support the same accelerators as the Granite Rapids-SP with 2 additional UPI 2,0 links (6 vs 4). As for memory, the Granite Rapids-AP chips are going to get 12-Channel IMCs with 2 DIMMs per Channel for a total of 24 DIMMs per socket. We can see boards with up to 48 DIMMs. The memory speeds are rated at DDR5-6400 (1DPC) and DDR5-5200 (2DPC).

In terms of PCIe lanes, each Intel Granite Rapids-AP CPU will get 96 PCIe Gen 5.0 lanes (64 lanes north and 32 lanes south). The company states higher PCIe P2P performance compared to the Sapphire Rapids chips with PCIe bifurcation of x16, x8, x4, and x2 (Gen 5). There will also be 64 CXL 2.0 lanes per CPU that can be used across any four of the five x16 ports.

Lastly, we have the Intel Sierra Forest-SP & AP lineup which will utilize E-Cores to offer the highest performance per watt in its class of CPUs with a focus on single-threaded performance and higher aggregate throughput. The chips are designed on the same Intel 3 process node and will be aimed at scalar workloads. The CPU cores will support AVX2 (2 x 128) instructions and feature a similar die configuration with two high-speed I/O dies. The Sierra Forest-SP CPUs are also targeted at the Birch Stream-SP socket (LGA 4677) though once again, the latest reports point out support on the newer LGA 7529 socket.

The slides also mention the cache count for the E-Cores on Sierra Forest chips. Each cluster with 4 cores will carry 4 MB L2 and 3 MB L3 cache.

So in the case of 144 cores, we will be looking at up to 144 MB of L2 and 108 MB of L3 cache. But that's the Sierra Forest-SP, in the case of the AP chips, it looks like we are expecting either up to 334 or 512 cores as recently rumored. The 334 config may end up with 334 MB of L2 and 250 MB of L3 cache. The Sierra Forest-SP CPUs are listed with DDR5 8-channel memory support and up to 350W max TDPs. The main compute die would be connected to the High-speed I/O dies using an EMIB interconnect within the SRF-SP package.

The slides also list down a few platforms for the Birch Stream-SP and Birch Stream-AP lineups with different cooling solutions. We get anywhere from 2U HP heatsink designs to 2D Vapor Chamber & Cold plate cooling systems. We first have the Granite Rapids-SP 2S reference validation platform as Beechnut City which features 32 DDR5 DIMM slots and then we have the Avenue City RVP for Intel's Granite Rapids-SP CPUs with 24 DDR5 DIMM slots. We detailed the Avenue City RVP a few days ago here.

In terms of die configurations, the Birch Stream-AP platform will get the GRR-AP UCC die which is the flagship die with three compute dies and two I/O dies, making up total of five dies in total. For the Birch Stream-SP platform, you get Granite Rapids-SP XCC (2 Compute + 2 IO Die), the HCC (1 Compute Die + 2 IO Die), and LCC (1 Compute Die + 2 IO Die).

Once again, it's important to note that the Intel 7nm process node which is now known as Intel 4, is mentioned here for the Granite Rapids Compute die. This isn't the case no more since these slides are old and we know that the compute tiles are fabricated on the newer Intel 3 process node. The same is the case for Sierra Forest. The Granite Rapids IO Die is mentioned as Intel 10 (Intel 7) but we cannot confirm that either.

The thing is, these slides are a few years old and while they give us more information about the future Xeon products, some claims and specs mentioned here may be out of date. Especially the performance claims, core counts, and TDP figures since they were made with the older Intel 4 and Intel 7 processes in mind.

News Source: YuuKi_AnS