LaunchedXeon 6 (Granite Rapids‑AP)

Intel · Xeon 6900P Series

Intel Xeon 6980P

128 P‑cores and 256 threads with 12‑channel DDR5/MRDIMM and 96 PCIe 5.0 lanes for HPC and AI servers.

HPC SimulationsAI Inference & TrainingIn‑Memory DatabasesVirtualized InfrastructureDense Multi‑Tenant Cloud

Cores / Threads

128/ 256

Base / Boost

2/ 3.9 GHz

PCIe Lanes

96

L3 Cache

504MB

TDP

500W

Socket

FCLGA7529

Verdict

8.8/ 10

88

Quick Verdict

A flagship Xeon 6 P‑core SKU that restores Intel’s competitiveness at the top of the server stack, with huge core counts, strong AI and HPC performance, and mature software support, though at very high platform cost and power.

Best for:HPC SimulationsAI Inference & TrainingIn‑Memory DatabasesVirtualized InfrastructureDense Multi‑Tenant Cloud

Overview

Launch

2024

Status

Launched

Generation

Xeon 6 (Granite Rapids‑AP)

Market

2S Data Center / HPC / AI

About this CPU

The Intel Xeon 6980P is a 128‑core, 256‑thread data center processor based on the Granite Rapids‑AP P‑core architecture, designed for dual‑socket HPC, AI, and scale‑out cloud workloads with 12 channels of DDR5/MRDIMM memory and 96 PCIe 5.0 lanes per socket.

Intel’s Xeon 6980P pushes Xeon back into the high‑core‑count fight with 128 P‑cores, 256 threads, and a 500 W TDP in the LGA7529 socket. It supports 12 channels of DDR5‑6400 or MRDIMM‑8800 memory and 96 PCIe 5.0 lanes per CPU, with integrated accelerators like AMX, QAT, DLB, DSA and IAA.

Intel’s own benchmarks show large gains over prior Xeon Platinum 8592+ and competitive wins versus AMD EPYC 9755/9654 in AI, HPC and database workloads, making it a strong fit for HPC clusters, AI inference farms and dense virtualized infrastructure—provided you can accommodate its power, cooling and price.

Specifications

ArchitectureGranite Rapids‑AP (Redwood Cove P‑cores)
Manufacturing ProcessCompute tiles: Intel 3; I/O tiles: Intel 7
Cores / Threads128 / 256
Base Clock2 GHz
Boost Clock3.9 GHz
L3 Cache504 MB
TDP500 W
SocketFCLGA7529
Memory TypeDDR5 / MRDIMM
Memory SpeedDDR5‑6400; MRDIMM‑8800
Memory Channels12×-Channel (12)
Max Memory3072 GB
PCIe Version / Lanes5.0 × 96
Integrated GraphicsNone
12×-Channel96 PCIe Lanes
Target Audience
GamersStreamersContent CreatorsDevelopersWorkstation UsersOffice UsersStudents

Performance

Productivity
N/A

Intel’s benchmarks show roughly 2× integer/FP throughput and ~2× memory bandwidth vs Xeon Platinum 8592+, and ~1.3–1.4× vs AMD EPYC 9654 in general compute workloads.

Virtualization
N/A

Strong VM density due to 128 cores and 12‑channel memory; Intel reports high Java, NGINX and database throughput, and up to ~1.9× perf/watt vs 5th‑gen Xeon at typical utilization.

Gaming
N/A

Not targeted at gaming; single‑thread performance is competitive with other server CPUs but not optimized for low‑latency game workloads.

Efficiency
N/A

500 W TDP is high; Intel’s data shows up to ~1.9× perf/watt vs Xeon 8592+ at 40% utilization, but absolute power and cooling requirements are substantial.

GamingNot applicable
  • Server‑oriented CPU with no integrated graphics and no gaming‑specific tuning.
  • Single‑thread performance is adequate for light game server workloads but not a design target.
CreatorExcellent
BlenderV‑RayAdobe Premiere Pro / Media EncoderDaVinci ResolveFFmpeg / SVT‑AV1 / SVT‑HEVC transcoding
AI / MLExcellent
  • Intel benchmarks show up to ~2.2× ResNet‑50, ~1.9× BERT‑Large, and up to ~2.5× DLRM inference vs Xeon 8592+ with MRDIMM.
  • Up to ~3.7× AI inference vs AMD EPYC 9654 in some Intel‑published comparisons.
  • AMX and AVX‑512‑FP16 accelerate int8/bf16 inference; software stack (oneAPI, OpenVINO) is mature on Linux.
Industry Impact
Gaming
Low
Workstations
Moderate
Content Creation
Moderate
Virtualization
High

Architecture

Compute tiles: Intel 3; I/O tiles: Intel 7

Process Node

Granite Rapids‑AP

Codename

128C / 256T

Core Config

504 MB

L3 Cache

500 W

TDP

Architecture Overview

Granite Rapids‑AP is a chiplet‑based server CPU with up to three Redwood Cove compute tiles and two I/O tiles connected via EMIB, delivering a logically monolithic mesh of up to 128 P‑cores with 12‑channel DDR5/MRDIMM and 96 PCIe 5.0 lanes.

CPU Design

128 Redwood Cove P‑cores with 2‑way SMT (256 threads), 504 MB shared L3, and per‑core 112 KB L1 and 2 MB L2; compute tiles built on Intel 3, I/O tiles on Intel 7.

Memory Subsystem

12‑channel DDR5‑6400 with MRDIMM‑8800 support; compute tiles host DDR5 controllers and a mesh interconnect, providing up to ~614 GB/s peak bandwidth in Granite Rapids‑AP configurations.

PCIe & I/O

96 PCIe 5.0 lanes per CPU with CXL 2.0 Type‑3 support; I/O tiles also provide up to 6 UPI 24 GT/s links for 2S coherence.

Overclocking

No unlocked multiplier; frequency is managed by Intel Turbo Boost and Speed Select Technology with configurable power/performance profiles.

Generation Comparison
Intel Xeon Platinum 8592+ (Emerald Rapids)Intel Xeon 6980P
  • 128 P‑cores vs 64 in 8592+, roughly 2× core count.
  • 12‑channel DDR5/MRDIMM vs 8‑channel DDR5‑5600.
  • 96 PCIe 5.0 lanes vs 80, plus CXL 2.0.
  • Integrated AMX, DLB, DSA, IAA, QAT accelerators vs more limited acceleration in prior gen.

Key Highlights

128 P‑cores / 256 threads
Full‑density Granite Rapids‑AP compute tile with 128 Redwood Cove P‑cores, maximizing throughput for parallel HPC and AI workloads.
12‑channel DDR5/MRDIMM up to 8800 MT/s
12‑channel memory subsystem with DDR5‑6400 and MRDIMM‑8800 support, delivering very high memory bandwidth for streaming HPC and vector workloads.
96 PCIe 5.0 lanes with CXL 2.0
96 PCIe 5.0 lanes per CPU with CXL 2.0 support, enabling large GPU/NVMe fabrics and CXL memory expansion in 2S servers.
Integrated accelerators (AMX, QAT, DLB, DSA, IAA)
On‑package accelerators for AI (AMX), crypto/QAT, load balancing (DLB), streaming (DSA) and in‑memory analytics (IAA), reducing need for discrete PCIe cards.
Strengths
  • 128 P‑cores / 256 threads for massive parallel throughput
  • 12‑channel DDR5‑6400 and MRDIMM‑8800 memory bandwidth
  • 96 PCIe 5.0 lanes with CXL 2.0 per socket
  • Strong AI/HPC performance with AMX and AVX‑512‑FP16
  • Mature Linux and compiler support (GCC/LLVM ‑march=graniterapids)
  • Integrated accelerators reduce need for discrete PCIe cards
Weaknesses
  • 500 W TDP demands high‑end cooling and power design
  • Very high CPU and platform cost compared to EPYC alternatives
  • 96 PCIe lanes trail AMD’s 128‑lane EPYC offerings
  • No integrated graphics; not suitable for graphical workloads
  • New LGA7529 platform with limited motherboard ecosystem initially

History

Launch Date
2024
Status
Launched
Generation
Xeon 6 (Granite Rapids‑AP)
Market
2S Data Center / HPC / AI
The Story

The Xeon 6980P represents Intel’s effort to reclaim leadership in the high‑core‑count server CPU segment. After AMD EPYC pushed past Intel’s 28–40‑core Xeon Scalable generations, Intel responded with the chiplet‑based Granite Rapids architecture. The 6900P series, led by the 6980P, is the first Xeon to reach 128 full P‑cores in a single socket, matching AMD’s core counts and restoring competition at the top of the HPC and AI server market.

</br>Intel’s own benchmarks and independent reviews show Granite Rapids delivering substantial performance lifts over the prior Emerald Rapids generation, especially when paired with MRDIMM memory and AMX‑enabled software. The platform also introduces CXL 2.0 and a large set of integrated accelerators, signaling Intel’s strategy to differentiate via on‑package specialization rather than raw core count alone.

However, the high 500 W TDP and premium pricing have kept the focus on large data center deployments where the throughput and feature set justify the cost.

Improvements over Previous Generation

  • 128 P‑cores vs 64 in 8592+, roughly 2× core count.
  • 12‑channel DDR5/MRDIMM vs 8‑channel DDR5‑5600.
  • 96 PCIe 5.0 lanes vs 80, plus CXL 2.0.
  • Integrated AMX, DLB, DSA, IAA, QAT accelerators vs more limited acceleration in prior gen.

Alternatives & Competitors

AMD EPYC 9755
Same 128‑core class with 500 W TDP and 128 PCIe 5.0 lanes; often better raw throughput per dollar and more PCIe lanes.
AMD EPYC 9654
96‑core alternative with lower TDP and good memory bandwidth; strong where 128 cores are underutilized.
Intel Xeon Platinum 8592+
Lower‑cost, lower‑power 64‑core option if you don’t need 128 cores or 12‑channel memory.
Intel Xeon w9‑3595X
Single‑socket workstation‑class CPU with high clocks and moderate core count for non‑HA workloads.
Intel Xeon 6 E‑core (Sierra Forest) SKUs
Better perf/watt and density for scale‑out cloud workloads that don’t require P‑core frequency.
Direct Competitors
AMD EPYC 9755AMD EPYC 9654Intel Xeon Platinum 8592+Intel Xeon w9‑3595XAMD EPYC 9575F

Should You Buy It?

Recommended for the right buyer

2S HPC or AI clusters where per‑socket throughput, memory bandwidth, and PCIe connectivity are critical, and where software is optimized for AMX/AVX‑512.

Avoid if…

  • Budget‑sensitive builds where EPYC 9654/9755 offers similar performance at lower cost.
  • Facilities that cannot supply 500 W per socket and robust cooling.
  • Workloads that don’t scale beyond 64–96 cores or can’t exploit the 12‑channel memory and 96 PCIe lanes.

Use Cases

HPC Simulations (CFD, CAE, Weather)
Excellent
AI Inference & Training (LLMs, Vision, Recommenders)
Excellent
In‑Memory Databases (MongoDB, Redis, Cassandra)
Excellent
Virtualized / Cloud Infrastructure
Excellent
General Purpose Business Workloads
Very Good

Interesting Facts

Xeon 6980P is the first Xeon to reach 128 full P‑cores in a single socket, matching AMD’s EPYC core counts at the top end for the first time in years.

It uses a 3‑compute‑tile, 2‑I/O‑tile chiplet design connected by EMIB, with compute tiles on Intel 3 and I/O tiles on Intel 7.

Intel’s performance index shows up to ~2.1× general compute performance and ~2.3× LINPACK vs Xeon Platinum 8592+, and up to ~1.85× BERT‑Large inference vs AMD EPYC 9755.

The 12‑channel MRDIMM‑8800 configuration can deliver memory bandwidth above 600 GB/s per socket in real benchmarks.

CXL 2.0 support enables memory expansion and pooling use cases, with Intel reporting up to ~16% higher Redis Vector Search performance when adding CXL memory.

Intel’s SGX implementation on Xeon 69xx shows <10% overhead for ResNet‑50 with Gramine, making confidential AI more practical.

The CPU uses Intel Speed Select Technology to define high‑ and low‑priority cores with different frequency targets, allowing admins to tune for latency vs throughput.

Early retail listings showed street prices dropping well below Intel’s original MSRP as 2025 progressed.

It is paired with the Avenue City platform (LGA7529), which supports up to 192 PCIe 5.0 lanes in a 2S system.

Despite the 500 W TDP, Intel reports up to ~1.9× perf/watt vs 5th‑gen Xeon at typical server utilization levels.

People Also Ask

Is Intel Xeon 6980P good for AI inference?

Yes. Intel benchmarks show up to ~2.5× DLRM, ~2.2× ResNet‑50, and ~1.9× BERT‑Large inference vs prior Xeon 8592+, and strong leads vs EPYC 9654/9755 when using AMX and MRDIMM.

How much memory does Xeon 6980P support?

Up to 3 TB of DDR5/MRDIMM memory across 12 channels, using DDR5‑6400 or MRDIMM‑8800 DIMMs.

How many PCIe lanes does Xeon 6980P have?

96 PCIe 5.0 lanes per CPU with CXL 2.0 support; 192 lanes in a dual‑socket system.

What socket does Xeon 6980P use?

Intel LGA7529 (FCLGA7529) for the Granite Rapids‑AP Avenue City platform.

Is Xeon 6980P better than EPYC 9755?

It depends on workload and pricing. Intel shows leads in some AI/HPC benchmarks, while EPYC 9755 offers 128 PCIe lanes and often better performance per dollar; real choice depends on software stack and TCO.

What process node is Xeon 6980P built on?

Compute tiles use Intel 3; I/O tiles use Intel 7, in a chiplet configuration.

Does Xeon 6980P have integrated graphics?

No. It has no integrated GPU and is intended for headless server and compute workloads.

What is the TDP of Xeon 6980P?

500 W. This is the default and max TDP; Intel’s Speed Select Technology allows configuring power profiles within that envelope.

Is Xeon 6980P suitable for gaming?

Not really. It can run game servers, but it’s optimized for server and HPC workloads, not for low‑latency client gaming.

What are the key accelerators in Xeon 6980P?

Intel AMX for AI, QAT for crypto/compression, DLB for load balancing, DSA for data streaming, and IAA for in‑memory analytics.

Frequently Asked Questions

What is the Xeon 6980P’s launch date?

Intel lists Q3 2024 as the launch quarter; Xeon 6900P series launched September 24, 2024.

How does Xeon 6980P compare to Xeon Platinum 8592+?

6980P roughly doubles core count (128 vs 64), adds 4 more memory channels, MRDIMM support, more PCIe lanes, and integrated accelerators, with Intel reporting ~2× throughput in many workloads.

Can Xeon 6980P run DDR4 memory?

No. It supports only DDR5 and MRDIMM DDR5 memory types.

What operating systems support Xeon 6980P?

Linux support is mature from day one (GCC/LLVM with ‑march=graniterapids); Windows Server and major hypervisors are also supported with appropriate drivers.

What is the max memory speed on Xeon 6980P?

8800 MT/s when using MRDIMM; DDR5‑6400 is the baseline speed.

Does Xeon 6980P support CXL?

Yes. It supports CXL 2.0 Type‑3 over the PCIe 5.0 lanes, enabling CXL memory expansion and pooling.

Is Xeon 6980P unlocked for overclocking?

No. The multiplier is locked; frequency is managed by Turbo Boost and Speed Select Technology profiles.

What cooling is recommended for Xeon 6980P?

High‑end air or liquid cooling designed for 500 W+ CPUs; 2U servers typically require heavy‑duty heatsinks or AIO liquid coolers.

How many Xeon 6980P CPUs can be used together?

Two in a dual‑socket (2S) configuration; Intel’s specification lists 2S scalability.

What chipsets work with Xeon 6980P?

It uses the Intel Avenue City platform (LGA7529) with associated server chipsets; there is no backwards compatibility with older LGA4189/3647 platforms.