Intel Xeon 6+ Clearwater Forest Runs Wide, but the Fastest SKU Is Not Always the Best Value

Intel has expanded Intel Xeon 6+ Clearwater Forest availability after shipping the family for more than a quarter. Its four processors span 144 to 288 cores.

The widening rollout changes the decision facing server buyers. Clearwater Forest is no longer mainly a hyperscaler story or an Intel 18A technology demonstration. More system vendors can now offer the processors, forcing customers to choose among meaningfully different core, power, cache, and bandwidth profiles.

That choice is less obvious than the short product list suggests. The 288-core Xeon 6990E+ leads total throughput, while the 144-core Xeon 6960E+ gives each core more memory bandwidth and cache. AMD EPYC and custom Arm processors add pressure outside Intel's lineup.

Intel Xeon 6+ Clearwater Forest Reaches a Wider Market

Clearwater Forest has moved from an important manufacturing milestone into a processor family that more enterprise buyers can evaluate.

Intel launched the Xeon 6+ family in the second quarter of 2026. ServeTheHome reports that production has since increased, bringing chips to vendors beyond the earliest hyperscale customers.

That wider availability matters because hyperscalers often receive new server processors before the broader market. Their scale also lets them qualify specialized systems that smaller operators cannot easily adopt.

The new phase gives more original equipment manufacturers and infrastructure teams access to Intel's first data center processor built around Intel 18A compute tiles. Intel 18A is the company's advanced manufacturing process using RibbonFET transistors and backside power delivery.

According to Intel's Clearwater Forest catalog, the family contains four physical processor models:

• Xeon 6990E+ has 288 cores, 576 MB of cache, and a 450-watt listed thermal design power.

• Xeon 6980E+ has 264 cores, 528 MB of cache, and a 400-watt listed thermal design power.

• Xeon 6970E+ has 192 cores, 480 MB of cache, and a 400-watt listed thermal design power.

• Xeon 6960E+ has 144 cores, 432 MB of cache, and a 330-watt listed thermal design power.

Every model uses efficiency cores, commonly called E-cores. These compact cores prioritize throughput density and energy efficiency over the strongest possible single-thread performance.

Unlike some earlier Xeon families, Intel does not reserve major platform features for selected models. Every Clearwater Forest SKU supports two-socket configurations, 12 DDR5 memory channels, and 96 PCIe lanes.

The processors also support MRDIMMs, which combine data from two memory ranks to increase transfer rates. Intel lists memory speeds reaching DDR5-8000 with suitable configurations.

Application Energy Telemetry appears throughout the family. AET is hardware-based monitoring that attributes energy consumption to workloads, containers, virtual machines, applications, or individual software threads.

That common feature set simplifies the first stage of selection. Buyers do not need a flagship processor merely to unlock memory channels, socket support, accelerators, or energy reporting.

The real choices therefore involve throughput, power, clock speed, memory pressure, cache allocation, and software scalability. Those factors create sharply different winners for different workloads.

The lineup also carries strategic weight for Intel. Clearwater Forest exercises Intel 18A manufacturing and Foveros Direct, a packaging method that connects vertically stacked silicon with dense interconnects.

Intel uses 12 compute tiles, three active base tiles, two input-output tiles, and 12 embedded interconnect bridges. That design lets the company combine several manufacturing technologies within one processor package.

The production ramp should help Intel refine those systems before later server products arrive. It also gives customers real hardware on which to test Intel's renewed density argument.

Clearwater Forest therefore represents two launches at once. It is a high-core-count Xeon product, and it is a production test for Intel's manufacturing and packaging strategy.

Four SKUs Create More Choices Than the Model Count Suggests

The Intel Xeon 6+ SKU list is short, but fixed platform resources make each step down the stack behave differently.

At the top, the Xeon 6990E+ offers 288 cores and 576 MB of last-level cache. It targets highly parallel applications that can keep hundreds of independent cores busy.

The Xeon 6980E+ reduces the count to 264 cores while retaining the same 3.2 GHz maximum turbo frequency. Its cache falls proportionally to 528 MB.

The Xeon 6970E+ makes a larger step down to 192 cores. It carries 480 MB of cache and has a 2.3 GHz base frequency.

The Xeon 6960E+ returns to the 144-core count found in Intel's previous Sierra Forest flagship. However, it raises cache capacity to 432 MB and uses a 2.4 GHz base frequency.

Those basic specifications do not capture the complete Intel Xeon 6+ value analysis. The same 12 memory channels and 96 PCIe lanes serve every processor, regardless of core count.

Reducing the number of cores therefore increases the resources available to each remaining core. That relationship matters when software cannot scale cleanly across the entire flagship processor.

The 6990E+ provides 2 MB of last-level cache per core. The 6970E+ increases that allocation to 2.5 MB, while the 6960E+ reaches 3 MB per core.

Memory bandwidth follows an even steeper curve. With 12 DDR5-8000 MRDIMM channels, the flagship offers about 2.67 GB per second of theoretical bandwidth per core.

The 144-core model offers about 5.34 GB per second per core under the same assumption. That is twice the per-core bandwidth without changing the memory subsystem.

These figures do not guarantee application performance. Memory access patterns, compiler behavior, data placement, and non-uniform memory access can reshape actual results.

They still expose a key Clearwater Forest tradeoff. More cores increase aggregate compute capacity, but they also increase contention for shared memory bandwidth and cache.

The 6960E+ may therefore outperform expectations in memory-sensitive databases, analytics engines, network processing, and services with imperfect parallel scaling. It gives each thread more supporting resources.

The flagship makes more sense for stateless services, dense virtual-machine hosting, and large fleets of lightweight containers. These workloads can convert core density into usable throughput.

The 192-core 6970E+ occupies a practical middle position. It offers one-third more cores than the 6960E+ while preserving considerably more cache per core than the flagship.

The 264-core 6980E+ is harder to separate conceptually from the 6990E+. Its appeal depends on system power limits, availability, and workload behavior near the top of the range.

Buyers should also distinguish maximum turbo from sustained multicore frequency. Maximum turbo usually describes limited-core conditions, not the speed maintained across hundreds of busy cores.

ServeTheHome reports that the 144-core model can reach a 3.0 GHz all-core turbo at its listed power. That makes it the family's frequency-oriented option.

Its higher per-core clock can matter for workloads that need substantial parallelism but stop scaling before 288 cores. The smaller model is not simply a reduced flagship.

The four Intel Xeon 6+ Clearwater Forest processors instead form two broad camps. The upper pair maximizes density, while the lower pair gives each core more supporting resources.

Low-Power Modes Change the Efficiency Winner

Intel's selectable power profiles make the highest-core models unusually efficient per core, even when they surrender frequency.

Intel Speed Select Technology allows supported processors to operate within defined power and performance profiles. The 6990E+ and 6980E+ each support lower-power configurations.

The 6990E+ can operate at 450 watts or move into a 330-watt profile. ServeTheHome reports that its all-core turbo falls from 2.8 GHz to 2.4 GHz.

That is a 14 percent frequency reduction paired with a 27 percent decrease in thermal design power. The lower setting changes the processor's role from maximum throughput toward efficiency.

At 330 watts, the 288-core processor works out to roughly 1.15 watts of listed power per core. The low-power 6980E+ reaches approximately 1.14 watts per core.

The calculation does not measure whole-server consumption. Memory, storage, networking, cooling, voltage regulation, and idle behavior all contribute to actual energy use.

It also does not establish application efficiency. A slower job can consume more total energy when execution time rises enough to offset lower instantaneous power.

Still, selectable profiles give operators flexibility without requiring a different processor purchase. A platform can favor throughput today and reduce socket power when rack constraints change.

That flexibility matters in facilities where electrical capacity limits deployment before floor space does. A high-density processor becomes less useful when its rack cannot supply or remove the required heat.

The 6960E+ takes the opposite approach. It uses the same 330-watt envelope as the reduced-power flagship but allocates that power across half as many cores.

Each core consequently receives more frequency headroom. The processor favors stronger per-core performance instead of maximizing the number of concurrent tasks.

This difference illustrates why watts per core cannot choose a server CPU by itself. Low watts per core favor density, while higher clocks can finish poorly scaling work sooner.

Intel's own performance claims also require context. At launch, the company said the 6990E+ delivered 2.26 times the average performance of the 144-core Xeon 6780E.

The newer processor has twice as many cores and a substantially higher power ceiling. Much of the aggregate gain therefore comes from scale, not a proportional leap within each core.

Intel also claimed a 55 percent average performance-per-watt improvement over that previous-generation processor. These results come from Intel-selected workloads and configurations.

The company has reported up to a 38 percent reduction in rack power for selected telecommunications comparisons. It also cited more than 60 percent better performance per watt.

Those claims show the intended use case, but they do not replace independent measurements. Buyers need tests that reproduce their utilization, memory population, and software stack.

The launch analysis also notes that a 450-watt processor can remain efficient when divided across 288 productive cores. Utilization is the decisive qualification.

An idle or underused 288-core processor offers little benefit from its favorable per-core arithmetic. Scheduling must keep enough cores busy without overwhelming memory or storage.

For operators choosing between modes, the question is not simply how many watts the processor permits. It is how much completed work the entire server delivers per kilowatt-hour.

The Best Clearwater Forest Value Depends on the Bottleneck

No single Xeon 6+ model wins throughput, per-core speed, bandwidth, cache allocation, and acquisition efficiency simultaneously.

For maximum socket throughput, the Xeon 6990E+ is the clear starting point. Its 288 E-cores provide the largest pool for highly parallel services.

This processor suits cloud instances, web-serving fleets, content delivery, microservices, and dense virtualization. These environments can distribute work across many independent execution contexts.

The flagship can also reduce the number of physical servers required for a fixed throughput target. Consolidation may lower networking, management, and rack-space overhead.

However, consolidation raises failure-domain risk. One unavailable system removes more capacity, so operators need scheduling, redundancy, and recovery practices that match the denser node.

For efficiency under a socket power limit, the lower-power 6990E+ and 6980E+ profiles deserve attention. Their core density produces the lowest listed watts-per-core ratios.

These configurations work best when throughput remains strong at reduced frequency. They are less attractive when latency targets depend on sustained per-core speed.

For balanced scale, the Xeon 6970E+ offers 192 cores and 480 MB of cache. It avoids both the smallest model's throughput ceiling and the flagship's sharpest resource contention.

Its 400-watt rating remains substantial. Yet the additional cache per core can make the system easier to feed than the two larger processors.

For per-core resources, the Xeon 6960E+ stands apart. Its 144 cores receive the highest cache allocation, the highest theoretical memory bandwidth per core, and the strongest all-core frequency.

That combination should interest teams migrating software that scales across many cores but not across 288. It also provides more headroom for memory-intensive application mixes.

Commercial value is similarly workload-dependent. The 6960E+ carries the lowest processor acquisition cost and the strongest core value based on Intel's published list positioning.

Moving upward increases total throughput potential, but core cost rises more quickly near the flagship. Buyers pay a density premium for putting more compute into each socket.

That premium can still be rational. Processor spending is only one component of a deployed server, and memory frequently represents a large portion of system cost.

A server also requires a chassis, motherboard, network interfaces, storage, power supplies, cooling, software, support, and operational labor. More expensive processors can reduce those duplicated costs.

Software licensing complicates the calculation further. Products licensed by socket may reward maximum density, while per-core licensing can make unused cores financially painful.

Virtualization licenses and database contracts require individual review. A seemingly efficient processor can become the costliest choice when commercial software counts every enabled core.

Capacity planning should therefore begin with measured workload limits. Teams need to identify whether compute, memory bandwidth, memory capacity, storage, networking, or licensing constrains deployment.

A representative test should compare completed work at several utilization levels. It should also record tail latency, package power, memory power, and total server energy.

Performance at 40 percent utilization can matter more than peak benchmark output. Cloud systems often operate below full load to absorb bursts and maintain service-level objectives.

Application Energy Telemetry may help expose those operating patterns. Intel says AET can attribute energy use more precisely than conventional whole-server monitoring.

That information can support placement decisions across containers or virtual machines. It may also reveal workloads that waste power while waiting on memory or input-output operations.

The resulting value judgment is straightforward:

• Choose Xeon 6990E+ when maximum socket throughput and consolidation density dominate.

• Choose a reduced-power upper-tier profile when rack power limits matter more than peak frequency.

• Choose Xeon 6970E+ when workloads need substantial parallelism with more cache per core.

• Choose Xeon 6960E+ when memory bandwidth, cache allocation, frequency, or lower acquisition cost dominates.

These are starting assumptions, not universal purchasing rules. Clearwater Forest's wide internal tradeoffs make workload testing more important than its four-model lineup suggests.

AMD EPYC and Arm Make Intel's Benchmark Story Harder

Clearwater Forest strengthens Intel's density position, but comparisons with AMD and Arm remain sensitive to threads, workload selection, and system design.

Intel positions the Xeon 6990E+ against AMD's EPYC 9965. The AMD processor provides 192 Zen 5 cores and uses simultaneous multithreading to expose 384 software threads.

Clearwater Forest does not use simultaneous multithreading. Its 288 E-cores expose 288 threads, creating an unusual comparison between physical cores and logical execution contexts.

Intel claims 30 percent higher average performance per thread than EPYC 9965. It also claims a similar performance-per-thread-per-watt advantage.

Per-thread results can matter for licensing, virtual-machine sizing, and latency. However, they do not establish which socket completes more total work.

Tom's Hardware noted that Intel did not provide a broad aggregate socket-performance comparison for the two flagships. Its benchmark review identifies that omission as an important limitation.

Thread count also changes the interpretation. AMD's 192 cores can execute 384 threads, while Intel uses 288 individual cores for 288 threads.

Neither design wins every workload. Simultaneous multithreading can improve utilization, but competing threads also share core resources.

Intel's dedicated E-cores avoid that sharing relationship. They may provide more predictable execution for some cloud workloads, especially when each virtual CPU maps to a physical core.

AMD retains stronger compatibility with workloads needing wide vector instructions. Xeon 6+ supports AVX2 but does not support AVX-512 or AVX10.

That omission narrows Clearwater Forest's natural market. High-performance computing, scientific simulation, and selected media or analytics workloads may favor different processors.

Intel instead includes accelerators for compression, cryptography, data movement, load balancing, and in-memory analytics. Software must use those engines to capture their benefits.

The company has highlighted large gains in selected SHA and SM3 cryptographic workloads. Such results reflect accelerator support and should not be generalized to ordinary application code.

Arm creates another competitive front. AWS Graviton, Google Axion, Microsoft Cobalt, and merchant processors such as AmpereOne target cloud efficiency and density.

Intel designed Clearwater Forest partly to keep cloud-native workloads on x86. That preserves familiar software deployment while adopting a many-core efficiency strategy associated with Arm servers.

The advantage is compatibility with the vast x86 software base. The disadvantage is that cloud operators increasingly control enough software to support another instruction set.

Intel has said Xeon 6+ compares favorably with Arm alternatives, but its launch materials lacked broad direct benchmark coverage. That leaves an important verification gap.

Platform familiarity may still help Intel. Existing Xeon 6900P systems can reportedly support Xeon 6+ through firmware updates when vendors validate the configuration.

A lower-friction deployment path matters to enterprises with standardized server fleets. It reduces the operational work surrounding qualification, firmware, management, and spares.

Yet hyperscalers evaluate economics at a different scale. Their custom Arm processors can integrate tightly with internal software, procurement, and data-center design.

Intel must therefore prove more than competitive benchmark scores. It must show that x86 compatibility, accelerators, telemetry, and density produce lower operating costs together.

The Intel Xeon 6+ Clearwater Forest case remains credible, but not settled. Independent server testing will determine how its theoretical advantages survive real software and sustained utilization.

Three Signals Will Show Whether Clearwater Forest Delivers

The next stage depends on independent performance, broad system availability, and evidence that Intel can sustain 18A production.

The first signal is reproducible workload testing across all four processors. Reviews should report full-system power, memory configuration, frequency behavior, and sustained throughput.

Tests should include cloud services, databases, telecommunications, virtualization, storage, compression, and cryptography. They should also examine tail latency rather than publishing averages alone.

Results across multiple power profiles will be especially useful. The 6990E+ changes character when moving from its 450-watt configuration to the 330-watt profile.

If lower-power mode retains most application throughput, Intel's efficiency argument becomes stronger. A steep performance loss would weaken simple watts-per-core comparisons.

The second signal is broad system availability. More servers from major vendors would confirm that Clearwater Forest has moved beyond limited hyperscale deployments.

Platform diversity matters because cooling, firmware, memory topology, and power delivery affect performance. Buyers also need qualified systems matching their existing management environments.

Intel says the processors use the same LGA 7529 socket family as Xeon 6900P systems. Vendor support will show whether that compatibility produces practical deployment savings.

A larger installed base would also create better data about software scaling. Users can then identify which applications benefit from 288 physical cores and which prefer smaller SKUs.

The third signal is Intel's execution beyond this launch. Clearwater Forest is the first server product using Intel 18A compute tiles, making production quality strategically important.

Stable availability would support Intel's claim that its manufacturing recovery is reaching volume server products. Persistent shortages or narrow distribution would raise harder questions.

The production experience also feeds later Intel server designs. Clearwater Forest exercises manufacturing and advanced packaging before those products face broader performance expectations.

For buyers, the immediate decision should remain narrower. Start with the workload, determine its bottleneck, and test the smallest reasonable set of configurations.

Do not treat the flagship as the automatic answer. The 6990E+ wins aggregate core density, while the 6960E+ gives every core more bandwidth, cache, and frequency.

The 6970E+ may offer the cleanest compromise, and the 6980E+ may fit power-constrained high-density systems. Actual value emerges only after complete server costs enter the model.

Clearwater Forest runs wide because Intel uses one common platform to cover several optimization targets. That range is useful, but it transfers more responsibility to the buyer.

Which constraint governs your next deployment: socket count, rack power, memory bandwidth, licensed cores, or latency? Benchmark that constraint first before selecting an Intel Xeon 6+ SKU.