The LGA 1954 socket is a major shift in Intel’s desktop strategy, according to leaked info. The new platform is expected to move away from the short-lived sockets to establish a high-performance, long-lasting ecosystem. Scheduled for a late 2026 launch, this platform introduces the Nova Lake-S (Core Ultra Series 4) processor architecture. It’s designed to compete with AMD through massive CPU core count increases and innovative cache solutions.
1. Nova Lake-S Architecture: Core Ultra Series 4 CPUs
Nova Lake-S utilizes a disaggregated tiled architecture with compute tiles primarily manufactured on TSMC’s N2P (2nm) process, except entry-level models based on Intel 18A.
- P-Cores & E-Cores: Features Coyote Cove P-cores for performance and Arctic Wolf E-cores for efficiency.
- Low-Power Island: A dedicated Hub Die includes 4 Low-Power Efficient (LPE) cores to handle background tasks and provide extreme idle efficiency.
Compared to the current Arrow Lake counterparts, Nova Lake-S is targeted for a 16% to 20% single-threaded uplift and a 12% to 23% multi-threaded boost for standard models. The 52-core flagship may reach an 80% multi-threaded boost, according to rumors.
Processor SKU Hierarchy
Here’s a Core Ultra 4 Series processor list based on unofficial info we were able to find online, primarily this Twitter account. We can’t confirm accuracy of available data about the CPUs (and other parts of the platform as well), so please take this list with a grain of salt.
| Intel Core Ultra Series Tier | Core Configuration (P+E+LPE) | Total Cores | Cache (L3/bLLC) | Mfg. Node |
|---|---|---|---|---|
| Core Ultra 9 (Flagship) | 16 + 32 + 4 | 52 Cores | 288MB (bLLC) | TSMC N2P |
| Core Ultra 9 (Sub-Flagship) | 16 + 24 + 4 | 44 Cores | 288MB (bLLC) | TSMC N2P |
| Core Ultra 7 (Premium Gaming) | 8 + 16 + 4 | 28 Cores | 144MB (bLLC) | TSMC N2P |
| Core Ultra 7 or 5 (Sub-Premium) | 8 + 12 + 4 | 24 Cores | 144MB (bLLC) | TSMC N2P |
| Core Ultra 5 (Mainstream) | 8 + 16 + 4 | 28 Cores | 36MB (Standard) | TSMC N2P |
| Core Ultra 3 (Budget) | 4 + 8 + 4 | 16 Cores | 18MB (Standard) | TSMC N2P |
| Core Ultra 3 (Entry) | 4 + 0 + 4 | 8 Cores | 12MB (Standard) | Intel 18A |
Processors for Every User: Gamers, Enthusiasts, Creators, Budget and Business
The Core Ultra Series 4 (Nova Lake-S) lineup is segmented to meet the specific needs of different user tiers, from budget-conscious builders to extreme workstation professionals.
- The Ultimate Enthusiast & Creator (Core Ultra 9 – Flagships): This top-tier SKU is designed for those who demand maximum multi-threaded performance. Featuring a massive 52 cores (16 P-cores, 32 E-cores, and 4 LPE cores) and 288MB of bLLC cache, it is estimated to deliver up to an 80% multi-threaded boost over previous generations. With an NPU capable of 74 TOPS, it is the premier choice for heavy AI workloads, 3D rendering, and professional video production. Intel reportedly also prepares a 44-core sub-flagship model with the same 288MB cache.
- The High-End Gamer (Core Ultra 7 – Premium Gaming): Specifically optimized for high-frame-rate gaming, this 28-core processor includes 144MB of bLLC cache. When combined with the APO+ program, which utilizes specialized instruction swaps to boost gaming performance by 15% to 25%, this CPU is expected to be 30% to 45% faster in games compared to non-bLLC models. Another potential option with 144MB bLLC cache is a 24-core Core 7 or 5 variant.
- The Balanced Mainstream User (Core Ultra 5 – Mainstream): Providing a middle ground for gaming and productivity, the Core Ultra 5 offers the same 28-core count as the Ultra 7 (8 P-cores, 16 E-cores, and 4 LPE cores) but utilizes a standard 36MB L3 cache instead of the massive bLLC. This makes it an efficient, high-performance option for general enthusiasts who do not require extreme cache levels.
- The Budget Gamer & Productivity User (Core Ultra 3 – Budget): This tier brings high core counts to the entry-level market with a 16-core configuration (4 P-cores, 8 E-cores, and 4 LPE cores). It provides a significant uplift for budget builds, offering modern features like AVX10.2 and APX support at a more accessible price point.
- The Entry-Level and Business Users (Core Ultra 3 – Entry): For those focused on extreme efficiency and basic computing, this 8-core model (4 P-cores and 4 LPE cores) is unique for being manufactured on the Intel 18A node. It provides modern architecture and 12MB of standard cache for light workloads and energy-efficient systems. The Core Ultra 3 is also suitable business environments where energy efficiency and stable “performance mainstream” operation are key.
The Cache Revolution: bLLC (Big Last Level Cache)
To compete with AMD’s 3D V-Cache high-cache architecture, Intel introduced bLLC, a specialized on-die L3 cache in the middle of the ringbus. It’s available on the high-end Core Ultra 7 and 9 CPUs.
- Capacity: Core Ultra 7 and 9 feature 144MB of bLLC per compute tile.
- Flagship Total: The 52- and 44-core Core Ultra 9 flagships use two tiles for a total of 288MB of L3 cache.
- APO+ Program: This specialized optimization program targets a 15% to 25% boost in gaming performance through instruction swaps.
Advanced Instruction Sets (FRED, AVX10.2, APX)
Nova Lake-S is the first consumer platform to introduce several next-generation instruction sets that optimize core processing logic and system efficiency.
- FRED (Flexible Return and Event Delivery): FRED is a significant architectural enhancement designed to improve how the processor handles low-level transitions, such as system calls and interrupts. By streamlining these events, FRED increases overall system efficiency and security.
- AVX10.2 & APX (Advanced Performance Extensions): These extensions work in tandem with the core architecture to enhance data processing and software performance across the entire system.
- System-Wide Benefits: Unlike dedicated hardware, these instruction sets provide a foundational boost to general system stability, security, and computational efficiency for all types of software.
Artificial Intelligence & AI Acceleration
The LGA 1954 platform features dedicated hardware and memory support specifically engineered for the increasing demands of local machine learning and AI workloads.
- High-Throughput NPU: The processors feature an integrated Neural Processing Unit (NPU) that delivers up to 74 TOPS of AI throughput, allowing for intensive AI tasks to be processed locally on the chip. According to leaks, the 74-TOPS NPU will be called NPU6.
- AI-Specific Memory (MRDIMM): For high-end and server-centric AI environments, the platform supports the future MRDIMM standard with extreme bandwidth to feed the data requirements of large-scale AI models.
2. Physical and Mechanical Specifications: LGA 1954
The new socket increases the contact count to 1,954 active landing pads.
- Physical Size: The package remains 45mm x 37.5mm, the same as LGA 1700/1851.
- Cooler Compatibility: Because dimensions and mounting patterns are identical, users can reuse CPU coolers from the LGA 1700 and 1851 generations.
- 2L-ILM Mechanism: On select higher-end designs where higher IHS flatness is crucial for good cooling, the socket will feature optional two-lever independent loading mechanism (2L-ILM).
- Power Delivery: The extra pins support massive power requirements, with peak draws potentially reaching 700W in specific states.
3. Memory Revolution: CUDIMM and DDR5-8000
To meet the massive bandwidth demands of its high-core-count processors, the LGA 1954 platform uses new CUDIMM (Clocked Unbuffered DIMM) DDR5 RAM as standard. This overhaul ensures that data throughput can keep pace with up to 52 cores by utilizing specialized hardware to maintain signal stability at the extreme speeds required by the Nova Lake-S architecture.
- Mandatory CKD Chip: To achieve stable operation at 6400 MT/s and higher, CUDIMM modules must include a Client Clock Driver (CKD) chip. This chip redrives the clock signal from the CPU to the DRAM, ensuring a clean and stable signal path that prevents performance degradation at high frequencies.
- Native DDR5-8000 Support: By adopting this new standard, the platform provides native support for DDR5-8000 (JEDEC). This transition offers a 25% increase in raw bandwidth over the previous DDR5-6400 baseline, effectively eliminating potential memory bottlenecks for the Core Ultra Series 4.
- CAMM2 Integration: While CUDIMM is the primary standard, some ultra-high-end overclocking motherboards may employ the horizontal CAMM2 form factor. These horizontal modules are designed to further improve signal integrity, catering specifically to the needs of bleeding-edge performance enthusiasts.
- MRDIMM (Multi-Ranked DIMM): This is a future memory standard supported by the LGA 1954 platform. It is specifically engineered for high-performance servers and AI data centers, targeting extreme data rates between 8800 and 12800 MT/s to meet the massive bandwidth requirements of next-generation workloads.
Memory Performance Sweet Spots
| Profile | Speed (MT/s) | Timings (CL) | True Latency | Impact |
|---|---|---|---|---|
| JEDEC Baseline | 8000 | CL52 | 13.0 ns | 30% slower than XMP |
| Sweet Spot | 8000 | CL38 | 9.5 ns | Standard for high-end gaming |
| Bleeding Edge | 9600 | CL46 | 9.58 ns | 20% more bandwidth than sweet spot |
| World Record | 10000+ | CL48 | ~9.6 ns | Requires OC-focused 2-DIMM boards |
Besides CUDIMM, the LGA 1954 platform will work with standard DDR5 UDIMM RAM as well, but that may degrade performance and stability significantly, especially when paired with higher-end Nova Lake-S CPUs.
4. Intel 900-Series Chipset Ecosystem
The 900-series chipset family (B960, Z970, Z990, Q970, and W980) introduces a major architectural shift by moving to the DMI 5.0 bus, which doubles the link speed to the CPU and effectively eliminates the I/O bottlenecks found in previous generations. This improvement is designed to support the massive data throughput required by modern gaming and professional workflows.
Chipset Comparison Table
| Feature | B960 | Z970 | Z990 | Q970 | W980 |
|---|---|---|---|---|---|
| Estimated Longevity | 4 Gens | 4 Gens | 4 Gens | 4 Gens | 4 Gens |
| IA (Core) Overclocking | No | Yes | Yes | No | No |
| BCLK Overclocking | No | No | Yes | No | No |
| Memory Overclocking | Yes | Yes | Yes | No | Yes |
| Integrated Thunderbolt 4 | 1 Port | 1 Port | 2 Ports | 2 Ports | 2 Ports |
| High-Speed USB 3.2 (20G) | ~2 | ~2 | ~5 | ~4 | ~5 |
| Primary GPU Slot (5.0) | 1×16 | 1×16 | 1×16/2×8/4×4 | 1×16/2×8/4×4 | 1×16/2×8/4×4 |
| Chipset PCIe 5.0/4.0 Lanes | 0 / 14 | 0 / 14 | 12 / 12 | 8 / 12 | 12 / 12 |
| CPU PCIe 5.0 NVMe Lanes | 1×4 | 1×4 | 1×8 or 2×4 | 1×8 or 2×4 | 1×8 or 2×4 |
| PCIe RAID Support | No | No | Yes | Yes | Yes |
| ECC Memory Support | No | No | No | No | Yes |
| Intel vPro Support | No | No | No | Yes | Yes |
Chipsets for Every User
- For the Average Gamer (B960): The B960 is positioned as the “standard” for most gamers. It offers a balance of performance and value. While it does not support IA (core) overclocking, it allows for full memory overclocking, which is critical for reaching the DDR5-8000+ sweet spot. The B960 provides a solid foundation with 34 PCIe lanes for high-speed connectivity.
- For the Budget Overclocker (Z970): A new category in the lineup, the Z970 is designed to bring IA (core) overclocking to more affordable desktop builds. It supports both core and memory tuning but features fewer high-speed I/O ports compared to the flagship enthusiast model.
- For the Hardcore Enthusiast (Z990): The Z990 is the exclusive choice for those requiring maximum control. It is the only chipset that supports BCLK (Base Clock) overclocking in addition to core and memory tuning. With 48 total PCIe lanes, five high-speed USB 3.2 (20G) ports, and dual integrated Thunderbolt 4 ports, it is built for the most complex builds.
- For the Enterprise & Business Users (Q970): Specifically designed for the “performance mainstream” corporate environment, the Q970 chipset prioritizes security and manageability over performance tuning. It includes Intel vPro support and PCIe RAID (0, 1, 5, 10) for data redundancy, but intentionally excludes IA and memory overclocking to ensure maximum system stability across large-scale deployments.
- For the Professional Creator (W980): The W980 is the dedicated workstation solution, providing true end-to-end ECC memory support to ensure data integrity during intensive creative tasks. Like the Z990, it offers 48 PCIe lanes. This allows creators to run multiple high-speed storage drives and expansion cards simultaneously.
“No Compromise” I/O for Modern Hardware
The platform features 24 CPU-direct PCIe 5.0 lanes, solving a common frustration for enthusiasts. This configuration allows a flagship PCIe 5.0 x16 GPU and two PCIe 5.0 x4 M.2 SSDs to run at their maximum rated bandwidth simultaneously without any reduction in speed (bifurcation). Additionally, the platform is future-proofed with integrated support for Wi-Fi 7 and Thunderbolt 5.
5. Professional & Enterprise Features
- W980 Platform: Provides true end-to-end ECC when paired with Xeon-W “Nova Lake-W” processors.
- Workstation Memory: Users can choose between ECC UDIMMs, ECC RDIMMs (for 512GB+ capacity), or the future MRDIMM standard (targeting 8800 to 12800 MT/s).
- Advanced I/O: The platform provides 24 CPU-direct PCIe 5.0 lanes, allowing a PCIe 5.0 x16 GPU and two PCIe 5.0 x4 SSDs to run simultaneously without bandwidth reduction. It also supports Wi-Fi 7 and Thunderbolt 5.
6. Platform Longevity
Intel is reportedly planning a four-generation lifespan for the LGA 1954 socket, supporting the following CPUs:
- Nova Lake (2026)
- Razer Lake
- Titan Lake
- Hammer Lake
This strategy aims to provide an upgrade path lasting four to six years.