Quick Verdict
A balanced Granite Rapids-SP SKU that pairs 36 P-cores with strong I/O and accelerators, ideal for consolidating older 2S clusters or building new general-purpose + AI inference nodes.
Overview
Launch
2025
Status
LaunchedGeneration
6th Gen Xeon Scalable (Granite Rapids-SP)
Market
Server / Enterprise
The Intel Xeon 6736P is a 36-core, 72-thread server processor based on the Granite Rapids-SP architecture, built on Intel’s 3 process and targeted at dual-socket enterprise, cloud, and AI-adjacent workloads requiring high memory bandwidth and strong per-core performance.
Intel’s Xeon 6736P sits in the midrange of the 6700P Granite Rapids-SP lineup, offering 36 P-cores, 72 threads, 8-channel DDR5-6400, and 88 PCIe 5.0 lanes in a 205W TDP envelope. It is designed for dual-socket servers running virtualization, databases, analytics, and AI inference, with features like AMX and QAT accelerators, high-priority/low-priority core tuning via SST-BF, and robust security including TDX and SGX.
It is not aimed at gaming or client workloads but excels at dense, data-center-oriented tasks where per-core performance and I/O matter more than raw core count.
Specifications
Performance
Strong multi-threaded and memory-bandwidth-heavy productivity workloads; SPEC CPU2017 integer rate scores around 75.7k base on dual-socket systems show robust throughput.
Excellent for VM density and I/O-heavy virtualized environments, with 8-channel DDR5, 88 PCIe 5.0 lanes, and VT-x/VT-d/EPT support.
Not intended for gaming; lack of integrated graphics and optimized client drivers means gaming is a secondary use case at best.
Intel 3 and SST-PP/SST-BF tuning help, but 36 cores at 2–4.1 GHz and 205W TDP still favor performance-per-socket over absolute efficiency.
- •No integrated graphics; requires discrete GPU.
- •Platform optimized for server workloads, not client gaming.
- •Latency and driver stack not tuned for gaming.
- •Single-thread performance is good, but not competitive with best gaming CPUs.
- •AMX and DL Boost accelerate matrix operations for inference.
- •Best suited for CPU-hosted inference models or pre-/post-processing alongside discrete accelerators.
- •Not a replacement for high-end GPUs or specialized AI accelerators for training.
Architecture
Intel 3 (~3nm-class)
Process Node
Granite Rapids-SP
Codename
36C / 72T
Core Config
144 MB
L3 Cache
205 W
TDP
Architecture Overview
Xeon 6736P is a Granite Rapids-SP processor built from compute chiplets on Intel’s 3 process and I/O chiplets that handle DDR5, PCIe, and accelerators. This chiplet approach lets Intel mix and match compute and I/O dies for different SKUs while keeping a logical mesh interconnect across the package.
CPU Design
The 6736P uses 36 Redwood Cove P-cores without E-cores, organized into compute complexes. Each core supports AVX-512 and AMX, with two 512-bit FMA units for vector and matrix work. Intel exposes a subset of cores as high-priority (12 at 2.1 GHz base) and low-priority (24 at 1.9 GHz base) via SST-BF, allowing administrators to tune clocks for latency-sensitive services without exceeding the 205W TDP.
Memory Subsystem
Eight DDR5-6400 memory controllers are distributed across the compute dies, providing up to 4 TB per socket and very high aggregate bandwidth. The mesh interconnect and CHA caching home agents maintain coherence and low latency across dies, with SNC (sub-NUMA clustering) modes to optimize local memory access when desired.
PCIe & I/O
Up to 88 PCIe 5.0 lanes are routed from the I/O dies, supporting multiple high-speed NVMe drives, GPUs, and SmartNICs in dual-socket servers. The I/O dies also host the QAT, DLB, DSA, and IAA accelerators, offloading crypto, compression, load balancing, and streaming workloads from the cores.
Overclocking
The 6736P does not support an unlocked multiplier for client-style overclocking. Instead, Intel exposes SST-PP and SST-BF profiles, letting you select different base and turbo regimes to favor either maximum core count or higher per-core clocks within the TDP limit.
- Intel 3 process vs 10nm Enhanced on Sapphire Rapids, improving power and density.
- Higher DDR5 speeds (6400 vs 4800 MT/s) and more memory channels (8 vs 8, but at higher data rates).
- Integrated accelerators (QAT/DLB/DSA/IAA/AMX) more tightly coupled and widely available.
- Higher maximum core counts and more flexible SST-PP/SST-BF tuning options.
Key Highlights
- 36 high-efficiency P-cores with 72 threads for dense server workloads.
- 8-channel DDR5-6400 with up to 4 TB per socket and high bandwidth.
- 88 PCIe 5.0 lanes for GPUs, NVMe, and SmartNICs.
- Integrated accelerators (AMX, QAT, DLB, DSA, IAA) for AI, crypto, and data processing.
- Granular SST-PP and SST-BF tuning for per-core clock and TDP optimization.
- Strong security feature set including TDX, SGX, and MK-TME for confidential computing.
- No integrated graphics; requires discrete GPU for any display output.
- Not optimized for gaming or client workloads.
- Platform is server-only; LGA4710 motherboards are not desktop boards.
- Higher platform cost compared to older Sapphire Rapids systems.
- Core count is modest versus top Granite Rapids-SP SKUs that reach 86+ cores.
History
The Xeon 6736P emerged as part of Intel’s Granite Rapids-SP generation, which launched in late 2024 and early 2025 to counter AMD’s EPYC 9004 and 8004 series. Granite Rapids-SP introduced Intel’s 3 process for compute dies and a more aggressive chiplet strategy, separating cores and memory controllers from I/O and accelerators. The 6700P series, including the 6736P, targeted mainstream enterprise and cloud deployments with core counts in the 16–86 range and a strong emphasis on AI accelerators and security.
The 6736P’s 36-core configuration filled a sweet spot for organizations wanting more cores than earlier 32–36 core Sapphire Rapids parts but without moving to the highest-density SKUs. Over time, it has become a common choice for dual-socket servers running virtualization, databases, and mixed AI inference workloads.
Improvements over Previous Generation
- Intel 3 process vs 10nm Enhanced on Sapphire Rapids, improving power and density.
- Higher DDR5 speeds (6400 vs 4800 MT/s) and more memory channels (8 vs 8, but at higher data rates).
- Integrated accelerators (QAT/DLB/DSA/IAA/AMX) more tightly coupled and widely available.
- Higher maximum core counts and more flexible SST-PP/SST-BF tuning options.
Alternatives & Competitors
Should You Buy It?
Recommended for the right buyer
New or refreshed dual-socket servers for virtualization, databases, and mixed enterprise + AI inference workloads where you want strong per-core performance, high memory bandwidth, and integrated accelerators without moving to the highest core-count SKUs.
Avoid if…
- Building a gaming PC or client desktop.
- You need maximum core count (>36 cores) for heavily parallel workloads.
- Your workloads are heavily GPU-accelerated and CPU is secondary.
- You want a clear upgrade path within the same platform (LGA4710) beyond Granite Rapids-SP.
Use Cases
Interesting Facts
Granite Rapids-SP is the first Xeon generation to use Intel’s 3 process for compute dies, roughly comparable to TSMC’s N3 node in density and performance.
The 6736P’s high-priority/low-priority core split is implemented via SST-BF, not E-cores; all 36 cores are P-cores, just with different base clocks.
SPEC CPU2017 results for dual-socket 6736P systems show integer rate scores above 75k, indicating strong throughput for enterprise workloads.
Intel’s Xeon 6 chiplet architecture moves I/O and accelerators onto separate dies, leaving compute dies focused on cores and memory controllers.
AMX support on Xeon 6 brings the same matrix instruction extensions used in Meteor Lake to the server line, targeting AI inference and HPC.
The 6736P supports Intel Speed Select Technology Performance Profile (SST-PP), allowing you to choose between more cores at lower clocks or fewer cores at higher clocks.
TDX and SGX enable confidential computing and secure enclaves at scale, important for multi-tenant cloud deployments.
Even though the 6736P is a 36-core part, its I/O and memory subsystem are shared with higher-core-count Granite Rapids-SP SKUs, giving it a relatively strong I/O and memory system for its core count.
Intel positions the 6700P series as the mainstream Granite Rapids-SP line for enterprise, versus the 6500P series for lower-cost segments.
Dual-socket 6736P servers can deliver 72 cores and 144 threads per node, offering a dense consolidation target for older 2S clusters.
People Also Ask
Is Intel Xeon 6736P good for gaming?
Not really. It lacks integrated graphics and is optimized for server workloads, not client gaming. A mainstream desktop CPU would be a better choice for gaming.
What socket does the Xeon 6736P use?
It uses the Intel FCLGA4710 socket (also referred to as Socket 4710), part of the Xeon 6 Granite Rapids-SP platform.
How much memory does the Xeon 6736P support?
It supports up to 4 TB of DDR5-6400 memory across 8 channels, depending on the memory type and DIMM population.
Does the Xeon 6736P have integrated graphics?
No, it does not have integrated graphics. You need a discrete GPU for any display output.
What is the difference between high-priority and low-priority cores on the Xeon 6736P?
The 6736P has 12 high-priority cores with a 2.1 GHz base and 24 low-priority cores with a 1.9 GHz base. Intel’s SST-BF feature lets you boost clocks on some cores for latency-sensitive workloads while lowering others to stay within TDP.
Is the Xeon 6736P unlocked for overclocking?
No. It does not support an unlocked multiplier for client-style overclocking, but it does support Intel SST-PP and SST-BF for controlled frequency and TDP profiles.
What accelerators are included in the Xeon 6736P?
It includes Intel QAT (crypto/compression), DLB (load balancing), DSA (data streaming), IAA (in-memory analytics), and AMX (matrix math for AI).
Can the Xeon 6736P be used in workstations?
Yes, it can be used in workstations for compute-heavy tasks, but it is primarily a server processor. Workstation variants (Granite Rapids-WS) may offer different feature sets and clock speeds.
How does the Xeon 6736P compare to AMD EPYC 9334?
The EPYC 9334 has 32 Zen 4 cores, 128 MB L3, and 128 PCIe 5.0 lanes, while the 6736P has 36 P-cores, 144 MB L3, and 88 PCIe 5.0 lanes. Both are strong server CPUs; choice depends on platform preference, pricing, and specific workload behavior.
What process node does the Xeon 6736P use?
The compute dies are manufactured on Intel’s 3 process (roughly equivalent to a 3 nm-class node from TSMC), while the I/O dies use a different, less advanced node.
Frequently Asked Questions
Does the Xeon 6736P support AVX-512?
Yes, it supports AVX-512 with two 512-bit FMA units per core, suitable for HPC and AI workloads.
Does the Xeon 6736P support AMX?
Yes, Intel Advanced Matrix Extensions (AMX) are present to accelerate matrix operations for AI inference and some HPC routines.
Can the Xeon 6736P run in single-socket servers?
Intel lists 2S scalability, but many vendors support 1S designs as well. Check your server vendor’s documentation for 1S support and any limitations.
What is the maximum turbo frequency on the Xeon 6736P?
The maximum turbo frequency is 4.1 GHz, with an all-core turbo of 3.4 GHz when many cores are active.
Does the Xeon 6736P support ECC memory?
Yes, it supports DDR5 ECC memory, which is standard for server platforms.
How many UPI links does the Xeon 6736P have?
It supports up to 4 UPI links at 24 GT/s for coherent inter-socket connectivity in dual-socket systems.
What security features does the Xeon 6736P include?
It includes Intel TDX, SGX, MK-TME, total memory encryption, and a full set of virtualization and root-of-trust technologies for secure cloud deployments.
Is the Xeon 6736P suitable for AI workloads?
It is suitable for CPU-based AI inference and pre-/post-processing, especially with AMX and DL Boost, but for large-scale model training you would typically pair it with GPUs or dedicated AI accelerators.
What is the Xeon 6736P’s TDP?
The default TDP is 205 W. Intel also lists SST-PP profiles with the same TDP but different base frequencies.
Does the Xeon 6736P have any E-cores?
No, the 6700P series uses only P-cores (Redwood Cove). The high/low priority split is a frequency and power tuning feature, not a heterogeneous core design.