Ryzen AI Max+ PRO 495: 12% Faster, More AI Memory

Ryzen AI Max+ PRO 495: 12% Faster, More AI Memory

The AMD Ryzen AI Max+ PRO 495 delivers an 11.8% performance improvement over the previous-generation Ryzen AI Max+ 395 in 3DMark Time Spy, according to early testing published by YouTube creator ETA Prime. The comparison used two GMKtec mini PCs running at a sustained 120W power setting: the EVO-X5 Pro with the new processor and the EVO-X2 with the older chip.

The results indicate that AMD’s new Gorgon Halo platform offers measurable performance gains, particularly in CPU and graphics benchmarks. However, the Ryzen AI Max+ PRO 495 is a refresh rather than a fundamental architectural redesign. Both processors use 16 Zen 5 CPU cores and 40 RDNA 3.5 graphics compute units, with relatively small increases in clock speed.

The more consequential change is memory capacity. The new platform supports up to 192GB of LPDDR5X memory, with up to 160GB configurable for the integrated GPU, compared with 128GB of system memory and a maximum 96GB GPU allocation on the tested Ryzen AI Max+ 395 system.

For developers running large language models (LLMs) locally, building AI applications, or experimenting with multiple models on a compact workstation, that additional memory capacity may be more valuable than the approximately 10% improvement seen in several conventional benchmarks.

One qualification matters when interpreting the results: the EVO-X5 Pro and EVO-X2 have different cooling systems and memory configurations. Although both systems operated at 120W, the comparison measures complete-system performance rather than an isolated difference between the processors.

📊 Synthetic Benchmark Results: Up to 14.4% Higher CPU Performance

#

The initial results show a consistent improvement in several synthetic benchmarks, but the size of the gain varies by workload. CPU-intensive testing and 3D graphics benchmarks show stronger improvements than the OpenCL test of integrated graphics.

3DMark Time Spy comparison

#

Source: ETA Prime’s comparison of the GMKtec EVO-X2 and EVO-X5 Pro.

The Ryzen AI Max+ PRO 495 scored 11,743 overall, compared with 10,499 for the Ryzen AI Max+ 395. Its CPU sub-score increased by 14.4%, while the graphics sub-score improved by 11.3%.

These results are notable because the newer chip retains essentially the same CPU and GPU architecture. The performance gains therefore reflect a combination of higher operating frequencies, faster memory, and system-level differences rather than a completely new core design.

However, 3DMark Time Spy is a synthetic benchmark. Its scores are useful for comparing performance within a defined test, but they do not directly translate into equivalent percentage improvements in application performance, gaming frame rates, or AI inference throughput.

Additional CPU and graphics benchmarks

#

The other synthetic results reinforce the same general pattern.

The 495’s CPU gains are relatively consistent across single-core and multi-core workloads. Its graphics results, however, depend more heavily on the benchmark.

The 8.6% improvement in Steel Nomad contrasts with the 3.1% increase in Geekbench OpenCL. This discrepancy illustrates why a single graphics benchmark cannot fully characterize an integrated GPU’s performance.

For developers and workstation users, the practical outcome will depend on the application’s compute patterns, memory access behavior, software optimization, and power management.

The importance of the testing conditions

#

Both systems were configured to operate at a sustained 120W, but the comparison is not a perfectly controlled processor-only experiment.

The EVO-X2 has 128GB of memory, while the EVO-X5 Pro has 192GB of LPDDR5X. The systems also use different cooling designs, which can affect thermal behavior and sustained clock speeds.

Consequently, the results establish that the tested EVO-X5 Pro system outperformed the EVO-X2 system under the reported conditions. They do not establish that every Ryzen AI Max+ PRO 495 implementation will outperform every Ryzen AI Max+ 395 system by the same percentages.

That distinction is particularly important when comparing compact computers, where firmware, thermal limits, memory configuration, and chassis design can influence performance substantially.

⚙️ AMD Gorgon Halo: A Refresh of Strix Halo

#

The Ryzen AI Max+ PRO 495 belongs to AMD’s Gorgon Halo platform, which builds on the previous Strix Halo design. Its core architecture remains unchanged, with improvements concentrated in clock speeds, memory support, and AI processing capability.

Ryzen AI Max+ PRO 495 specifications

#

Compared with the Ryzen AI Max+ 395, the new processor raises its maximum CPU boost frequency from 5.1GHz to 5.2GHz and the graphics clock from 2.9GHz to 3.0GHz.

The NPU also increases from 50 TOPS to 55 TOPS. TOPS measures theoretical AI operation throughput under specified conditions, rather than directly measuring application-level inference speed.

The underlying architecture remains the same: 16 Zen 5 CPU cores and a 40-compute-unit RDNA 3.5 GPU. There is no fundamental change in core count or graphics architecture.

For conventional workloads, this makes Gorgon Halo an incremental refresh rather than a generational leap. Its broader value proposition comes from combining modest performance improvements with greater memory capacity for AI workloads.

🧠 Memory Capacity Is the Real Advantage for Local AI

#

The most important distinction between the two platforms is their memory configuration.

The tested EVO-X2 provides 128GB of system memory, while the EVO-X5 Pro offers 192GB of LPDDR5X. The newer platform also increases theoretical memory bandwidth from 256GB/s to 273GB/s.

The GPU memory figures refer to configurable allocations from shared system memory, not dedicated graphics memory physically installed on a discrete graphics card.

This distinction matters because the CPU and integrated GPU share the system’s memory resources. Allocating more memory to the GPU makes a larger working set available to graphics and AI applications, but it does not create additional physical memory or eliminate contention for bandwidth.

Why unified memory matters for LLM inference

#

Running a large language model locally requires memory for model weights, intermediate activations, the key-value (KV) cache, runtime overhead, and other working data.

Model weights typically account for a substantial portion of the memory requirement. Quantization reduces the storage needed per parameter, but the total requirement still depends on model size, quantization format, context length, and inference implementation.

For example, a model with approximately 100 billion parameters requires roughly 50GB for its weights at an idealized four bits per parameter, before accounting for metadata, runtime overhead, the KV cache, and other allocations.

A model with 120 billion parameters would require roughly 60GB under the same simplified assumption. Actual memory requirements vary with the quantization format and software implementation.

The available memory can therefore determine whether a model fits on a machine at all, rather than merely influencing how quickly it runs.

The 495 platform’s 192GB memory pool gives developers more room to experiment with larger models, longer context windows, and more demanding local AI workflows without immediately moving to a multi-GPU workstation.

For additional context on memory bandwidth and local inference, see KAD’s analysis of the Ryzen AI Max+ 395 and its role in local AI computing.

Memory capacity and bandwidth solve different problems

#

Capacity and bandwidth are related, but they address different performance constraints.

Memory capacity determines how much model data and working state can remain in memory. If a model exceeds the available allocation, the application may have to use a smaller model, reduce the context length, offload work to another device, or rely on slower storage.

Memory bandwidth determines how quickly data can move between memory and the processors. During LLM token generation, model weights must be accessed repeatedly, making memory bandwidth an important factor in decoding performance.

The increase from 256GB/s to 273GB/s represents approximately 6.6% more theoretical bandwidth. The maximum memory capacity, meanwhile, rises by 50% from 128GB to 192GB.

For workloads that already fit comfortably within 128GB, the extra capacity may offer limited direct performance benefits. For workloads that require more than the previous system can allocate to the GPU, it can change which models and configurations are practical.

This is why the 495’s strongest argument is not simply that it is faster. It provides a larger memory envelope while retaining an integrated CPU and GPU in a compact system.

🤖 Local AI Benchmarks: Modest Gains Across Tested Models

#

ETA Prime’s local AI tests show improvements in the single-digit percentage range, broadly consistent with the processor’s incremental hardware changes.

An important distinction is that the reported Geekbench AI 1.7 tests ran on the integrated GPU through ONNX DirectML, rather than on the dedicated NPU. The results therefore measure GPU-based AI execution, not the performance of AMD’s XDNA 2 NPU.

Geekbench AI 1.7 results

#

The results show that the 495 improves GPU-based AI benchmark performance, with the largest increase occurring in single-precision testing.

Half-precision and quantized workloads improve by smaller amounts. This variation is expected because different numerical formats and execution paths place different demands on compute throughput, memory bandwidth, and software optimization.

These results should not be interpreted as a direct comparison of NPU performance. Measuring the XDNA 2 NPU would require a workload that specifically targets that processor and an appropriate software runtime.

Local LLM generation speed

#

The comparison also tested several local language models.

Source: ETA Prime’s published comparison. Both systems operated at 120W, with 128GB and 192GB of system memory, respectively.

The 495’s generation-speed advantage ranges from 4.7% to 6.7% across the tested models. These are useful improvements, but they do not indicate a major leap in inference speed.

More importantly, the result depends on the model and test configuration. Quantization, runtime implementation, context length, batch size, and memory allocation can all influence generation throughput.

The added capacity may be more valuable than the small speed increase when an application requires larger models or longer contexts. However, more available memory does not automatically make a model run faster. If both systems can already hold the same model and working data, the higher clock speeds and additional bandwidth are likely to have a more direct effect on throughput.

What the results mean for AI developers

#

The practical appeal of the Ryzen AI Max+ PRO 495 is its ability to combine a general-purpose CPU, a capable integrated GPU, and a large shared memory pool in one compact system.

Potential use cases include:

• Local LLM development: Testing and serving larger models without relying exclusively on cloud inference.
• AI application prototyping: Running model inference alongside API services, data preprocessing, and development tools.
• Multimodel experimentation: Keeping multiple model components or workloads resident in memory, subject to software and memory requirements.
• Private inference environments: Processing supported workloads locally when data-handling requirements favor keeping information on-device.
• Compact AI workstations: Running CPU and GPU workloads in a small desktop system without installing a separate discrete GPU.

The platform’s suitability depends on the software ecosystem as well as the hardware. Framework support, GPU backends, quantization libraries, memory management, and operating-system compatibility determine how effectively applications can use the available resources.

For developers choosing a local AI platform, the AMD Ryzen AI Halo development platform provides a useful reference point for understanding how AMD’s unified-memory approach targets local inference workloads.

🖥️ Product Positioning: A Specialized Mini PC for AI and Workstation Tasks

#

The Ryzen AI Max+ PRO 495 is not designed solely for benchmark performance. Its combination of 16 CPU cores, integrated graphics, an NPU, and support for up to 192GB of unified memory gives system builders flexibility when designing compact AI workstations.

The GMKtec EVO-X5 Pro illustrates this positioning by pairing the processor with a small desktop chassis and a large LPDDR5X memory configuration.

That combination is attractive for developers who need substantial local memory but do not want a conventional tower workstation with a discrete GPU. It can also reduce the complexity of building a compact system, although the integrated design does not eliminate thermal, power, or software limitations.

Who should consider the Ryzen AI Max+ PRO 495?

#

The 495 is most compelling for users whose workloads benefit from a larger memory pool or who need a balanced CPU-and-GPU platform in a small form factor.

AI developers working with large models may benefit from the extra memory allocation. General workstation users may appreciate the higher CPU scores and integrated graphics improvements. Developers running smaller models that already fit comfortably in 128GB, however, may see less value from the upgrade.

For gaming-focused users, the 11.3% Time Spy graphics improvement is promising but does not guarantee the same uplift in every game. Real-world performance depends on resolution, graphics settings, memory pressure, drivers, and the application’s workload.

The system should therefore be evaluated against the intended use case rather than treated as a universally faster replacement for the previous generation.

🔬 Benchmark Limitations and What to Watch Next

#

The first comparison establishes a useful performance baseline, but several factors limit how broadly the results can be applied.

First, the EVO-X2 and EVO-X5 Pro differ in cooling design and memory capacity. The common 120W operating condition controls one major variable but does not eliminate system-level differences.

Second, synthetic benchmarks measure defined workloads. Their results cannot fully predict performance in every game, developer tool, or AI framework.

Third, LLM inference depends on more than processor specifications. The selected model, quantization scheme, inference backend, context length, and allocation of memory to the GPU can materially change performance.

Future testing should compare systems with closely matched cooling and memory configurations where possible, while also examining power consumption, acoustic behavior, sustained performance, and performance per watt.

AI tests should additionally distinguish GPU-based inference from NPU execution and report the software stack used. That would help clarify which workloads benefit most from the architectural changes and which are primarily limited by memory capacity or software support.

✅ Conclusion: More Memory Matters More Than a Small Clock Boost

#

The AMD Ryzen AI Max+ PRO 495 delivers a meaningful but incremental improvement over the Ryzen AI Max+ 395 in early benchmarks. The tested GMKtec EVO-X5 Pro scored 11.8% higher in 3DMark Time Spy, with a 14.4% CPU sub-score improvement and an 11.3% graphics improvement.

The more important change is the expansion from 128GB to 192GB of system memory, alongside an increase in maximum GPU memory allocation from 96GB to 160GB. For local AI workloads, that larger memory pool can make models and configurations practical that would otherwise exceed the previous system’s available allocation.

The distinction is between performance improvements and capability improvements. Higher clock speeds and modest bandwidth gains can accelerate workloads that already fit within the system’s limits. Greater memory capacity can expand the range of models and applications that the system can accommodate.

For developers running local LLMs, experimenting with AI applications, or building compact workstations, the Ryzen AI Max+ PRO 495 is therefore more than a slightly faster version of its predecessor. Its strongest value proposition is the combination of a modest performance uplift and significantly greater memory capacity in a single integrated platform.

Whether it is worth upgrading will ultimately depend on workload size, software compatibility, system pricing, and how often the additional memory is needed. For users constrained by model size rather than raw inference speed, that additional capacity may prove to be the most important improvement of all.