NVIDIA DGX Spark (MediaTek AHJ11488B)

NVIDIA DGX Spark (MediaTek AHJ11488B) is a 20-core, 20-thread mobile processor built on TSMC’s 3 nm process node, with a base clock of 2.80 GHz and a boost clock of 4.00 GHz, a 140 W TDP, and a die size of 208 mm². It launched on 2025-10-14 with a launch MSRP of $3999, and the benchmark data positions it at the 50th percentile among all CPUs, indicating a mid-pack standing in the broader performance landscape.

Benchmark Performance

The benchmark results for the NVIDIA DGX Spark show a score that places it exactly at the 50th percentile compared to all CPUs in the database. This is a median position, meaning half of all recorded processors outperform it and half underperform it. For a 20-core part, this central ranking suggests that raw throughput is competitive but not class-leading, and the absence of any nearestRivals data means there are no direct percentage deltas to cite for head-to-head comparisons. The avgBenchmarkScore of 0 further indicates that no specific multi-threaded or single-threaded aggregate score has been recorded, so the percentile ranking is the primary quantitative reference point.

Given the 20-core/20-thread configuration, the lack of hyper-threading means each core handles exactly one thread. This design choice typically favors workloads with high per-core efficiency rather than those that benefit from massive thread oversubscription. In synthetic benchmarks, a 20-thread processor at the 50th percentile would likely deliver consistent mid-tier results in heavily parallel tasks like video encoding or 3D rendering, but it would not challenge top-tier 24-core or 32-core parts that occupy higher percentiles. The 4.00 GHz boost clock provides a solid frequency ceiling for bursty single-threaded tasks, though the 2.80 GHz base clock indicates that sustained all-core loads will operate at a more moderate pace.

The median percentile also implies that in mixed workload suites, the DGX Spark will neither embarrass nor impress relative to the competition. Users comparing against higher-percentile CPUs should expect a noticeable gap in multi-core throughput, while those coming from lower-percentile parts will find a clear step up. Without explicit rival scores, the data alone supports a verdict of "competent mid-pack" rather than "performance leader."

Power and Thermals

The NVIDIA DGX Spark carries a TDP of 140 W, which classifies it as a high-power mobile processor. This is not an ultra-low-power chip; it requires a robust cooling solution capable of dissipating that thermal load continuously. In a mobile context, 140 W typically demands a substantial vapor chamber or a dual-fan cooling module, and it precludes thin-and-light chassis designs. The 3 nm process node from TSMC helps mitigate power density, but the sheer TDP figure dominates the thermal story.

For cooling tier inference, a 140 W TDP places the DGX Spark in the same category as high-end laptop CPUs that necessitate gaming-laptop-style cooling—thick heatsinks, high-static-pressure fans, and possibly a shared heat pipe system with the integrated GPU. The "GB20B Blackwell" integrated graphics also contributes to the total thermal envelope, though the databook does not specify its individual power draw. Users should expect significant fan noise under sustained load and a chassis that prioritizes airflow over portability.

The lack of ECC memory support (eccMemory: false) is worth noting for thermal and power planning, as error-correcting memory typically adds a small power overhead. Here, the absence means slightly lower memory-related power draw, but the overall 140 W TDP remains the dominant constraint. For battery-powered operation, this processor would drain cells quickly; the design clearly targets performance-per-watt over longevity, with the 3 nm node doing the heavy lifting to keep efficiency reasonable at that power level.

How It Compares

The nearestRivals field in the FACT PACK is empty, so no direct competitor names, scores, or deltaPct values are available for comparison. This absence means the analysis must rely on the percentileVsAllCpus value of 50 to contextualize the DGX Spark. In the absence of specific rivals, the processor’s position can only be described in aggregate: it sits at the median of all tested CPUs.

If one were to infer rivals from typical 20-core mobile parts, the DGX Spark would likely face off against other high-core-count laptop processors. However, without deltaPct data, any such comparison would violate the rule against citing rival specs. Therefore, the honest statement is that the benchmark database currently provides no head-to-head entries for this chip, leaving the 50th percentile as the sole positional reference. This is a limitation of the data, not a reflection of the processor’s capability.

In practical terms, a user moving from a processor at the 30th percentile would find the DGX Spark a meaningful upgrade, while one from a 70th-percentile chip would see a downgrade. The 20-core count suggests it will hold its own in multi-threaded productivity tasks, but the median percentile warns against expecting top-tier frame rates or render times. The lack of rival data means the "How It Compares" section is necessarily sparse, but the percentile anchor provides a baseline for expectations.

Who Should Consider It

For gaming, the DGX Spark’s 20 cores and 4.00 GHz boost clock are sufficient for modern titles, but the 50th percentile ranking implies it is not a gaming flagship. The integrated GB20B "Blackwell" graphics handles display output, but the CPU’s mid-pack standing suggests that pairing it with a discrete GPU would be necessary for high-refresh-rate gaming. The high TDP of 140 W also means it is most viable in a large laptop with dedicated graphics, where the CPU can feed frames without bottlenecking mid-range GPUs. Casual gamers at 1080p with medium settings would find it adequate, but enthusiasts should look elsewhere.

For content creation, the 20 threads and 273.1 GB/s memory bandwidth (from LPDDR5X quad-channel support) make the DGX Spark a capable choice for video editing, 3D modeling, and software compilation. The median percentile indicates that render times will be average—not blazing fast, but not frustratingly slow. The 2 MB L2 cache per core and 8 MB shared L3 cache provide enough on-die storage for working sets in these workloads. Users who render 4K video or compile large codebases will see acceptable throughput, though they would benefit from a higher-percentile CPU if time is money.

For office productivity and general desktop use, the DGX Spark is overkill. Web browsing, spreadsheets, and document editing rarely stress 20 cores, and the 140 W TDP means the laptop will run hot and drain battery unnecessarily. The 50th percentile is more than sufficient for these tasks, but the power draw makes it a poor fit for all-day battery life. This is a processor for heavy lifting, not light errands. The 3 nm node keeps idle power reasonable, but sustained office workloads would still trigger the cooling fans.

Platform and Compatibility

The NVIDIA DGX Spark uses LPDDR5X memory with a quad-channel bus, delivering a memory bandwidth of 273.1 GB/s. This is a high-bandwidth configuration typical of integrated-memory designs, which bodes well for memory-intensive applications like data analytics or in-memory databases. The quad-channel layout means the memory controller is not a bottleneck for the 20 cores, especially in multi-threaded scenarios where many cores compete for data.

The socket field is listed as null, indicating a soldered or proprietary mounting solution rather than a user-replaceable socket. The PCIe field is "N/A", which suggests that the processor does not expose standard PCIe lanes for expansion, likely relying on the integrated GB20B "Blackwell" graphics and possibly vendor-specific interconnects. This limits upgradeability—users cannot add a discrete GPU via PCIe in the traditional sense, though the integrated GPU handles display duties.

Memory support is exclusive to LPDDR5X, which is soldered on most platforms, meaning RAM is not user-upgradeable. ECC memory is not supported (eccMemory: false), which is a consideration for workstation users who require error correction for long compute jobs. The production status is "Active", and the release date of 2025-10-14 means it is a current product. The part number DSE1-275-A1 identifies the specific SKU, but the lack of a socket and PCIe details points to a closed platform—likely an embedded or laptop design where the CPU, memory, and GPU are integrated into a single board.

Single-Thread vs Multi-Thread Behavior

The DGX Spark has 20 cores and 20 threads, which is a pure multi-threaded configuration without simultaneous multi-threading (SMT). This means each core executes one thread, and the processor’s multi-threaded performance scales linearly with core count only if the workload has sufficient parallelism. The 4.00 GHz boost clock is the key single-thread metric, and the 2.80 GHz base clock indicates the sustained all-core frequency. In single-threaded tasks, the boost clock allows for responsive performance in applications like web browsing or legacy software that cannot utilize multiple cores.

The 50th percentile ranking reflects the balance of these two behaviors. For multi-threaded workloads (e.g., video encoding, physics simulations, or batch photo processing), the 20 cores provide strong parallel throughput, but the median percentile suggests that competing 24-core or 32-core parts with higher clocks or SMT would outperform it. The 2 MB L2 cache per core is generous, reducing cache misses in single-threaded loops, while the 8 MB shared L3 cache is modest for 20 cores, potentially causing contention in data-heavy multi-threaded tasks.

The single-thread to multi-thread split means that the DGX Spark excels in mixed workloads that combine a few heavy single-threaded tasks with many light parallel ones—such as a developer running a compiler while streaming and browsing. However, in pure single-threaded benchmarks, it would likely sit near the 50th percentile as well, given the 4.00 GHz boost is not exceptionally high. In pure multi-threaded benchmarks, the 20 cores help, but the lack of SMT and the moderate frequency cap its ceiling. The memory bandwidth of 273.1 GB/s is a strong asset for multi-threaded workloads that stream data, mitigating some of the L3 cache limitations.

FAQ

Q: What is the processor core and thread count?

A: The NVIDIA DGX Spark has 20 cores and 20 threads, with no hyper-threading, so each core handles exactly one thread.

Q: What is the launch MSRP?

A: The launch MSRP is $3999, as stated in the FACT PACK.

Q: Does the processor support ECC memory?

A: No, ECC memory is not supported (eccMemory: false).

Q: What is the process node and foundry?

A: It is built on a 3 nm process node by TSMC.

Q: What is the memory bandwidth and type?

A: It supports LPDDR5X memory with a quad-channel bus, providing a memory bandwidth of 273.1 GB/s.

Q: What is the TDP and what cooling does it require?

A: The TDP is 140 W, which requires a robust cooling solution typical of high-performance mobile processors, such as a large vapor chamber or dual-fan setup.

Architecture and Design

The NVIDIA DGX Spark is based on the GB10 codename, part of the "Spark (GB10)" generation. It is fabricated on a 3 nm process node at TSMC, with a die size of 208 mm². The architecture is not explicitly named in the FACT PACK, but the integrated graphics is the GB20B "Blackwell", indicating a unified design where the CPU and GPU share the same die or package. The transistor count is not listed, so the density cannot be quantified, but the 3 nm node implies a high transistor density for the 208 mm² area.

The cache hierarchy consists of a 128 KB L1 cache per core, a 2 MB L2 cache per core, and an 8 MB shared L3 cache. This per-core L2 allocation is generous, aiding single-threaded performance by reducing reliance on the shared L3. The 8 MB L3 is relatively small for a 20-core processor, which could lead to cache thrashing in workloads with large working sets, but the high memory bandwidth of 273.1 GB/s partially compensates. The memory controller supports LPDDR5X in a quad-channel configuration, and the lack of PCIe expansion (pcie: N/A) suggests that all I/O is handled through the integrated GB20B GPU or proprietary interfaces.

The multiplier is locked (multiplierUnlocked: false), meaning overclocking is not supported. The production status is "Active", and the release date is 2025-10-14. The part number DSE1-275-A1 identifies the specific silicon. The market segment is "Mobile", which aligns with the 140 W TDP and LPDDR5X memory, indicating a high-end laptop or portable workstation design. The architecture is a single, monolithic die with 20 cores, no SMT, and a balanced cache structure, optimized for throughput in a power-constrained mobile envelope, leveraging the 3 nm process to keep thermals manageable at 140 W.