GX10 or ET900N G3: how to choose your ASUS machine for local AI

For learning, developing a personal assistant or testing a clearly defined service, we recommend starting with the 64 GB GX10 if your software and model fit comfortably in memory. The 128 GB GX10 is worth considering when models grow, context gets longer or several AI components need to coexist. Be sure to check the size of the models you want to use simultaneously. The ET900N G3 addresses a different category of needs: HBM3e GPU memory, much greater coherent memory capacity, remote management and computing infrastructure.

These recommendations are not benchmark results. Here, we compare documented specifications and their practical implications. Without measurements on your model, we cannot promise a tokens-per-second rate or a specific number of users. The aim is to select the right category of machine before testing a complete configuration.

An initial estimate starts with the model weights: number of parameters × bits per weight ÷ 8. For a dense model with 32 billion parameters, that is around 64 GB at 16 bits, 32 GB at 8 bits or 16 GB at 4 bits. This calculation uses decimal gigabytes and covers only theoretical weights, without quantisation metadata or runtime engine allocations.

You then need to add the KV cache, which stores information related to the conversation context, the buffers used for computation, the operating system and other services. Longer context and several active conversations can greatly increase memory usage. A 32 GB model file therefore does not mean a 32 GB machine is enough to run it comfortably.

Applying the same calculation to 70 billion parameters gives around 140 GB at 16 bits or 35 GB at 4 bits for the weights alone. This shows why quantisation changes sizing requirements, without guaranteeing that a given model will work well on a GX10. You need to check format support, output quality and response time: fitting in memory is a necessary condition, not a performance verdict.

Enlarge

The 64 GB GX10 may suit a developer who already knows their model, accepts suitable quantisation and wants a dedicated work machine. A small internal assistant, a document retrieval prototype or a learning environment can fit this approach. ASUS targets models with around 30 to 35 billion parameters, with capabilities depending on precision and context.

Its appeal lies in retaining the GB10 platform: a 20-core Arm CPU, an integrated Blackwell GPU and a stated memory bandwidth of 273 GB/s. Reducing capacity to 64 GB therefore does not automatically halve all performance. However, reaching the memory limit may require a smaller model, reduced context or a different service setup.

The GX10-GG0062BN model, also identified as 90MS0371-M001Y0, pairs this memory with a 1 TB PCIe 4.0 NVMe SSD. Choose it for a clearly defined need, not on the assumption that you can add a RAM stick later: this platform's unified memory is not offered as add-in DIMMs.

The 128 GB GX10 mainly offers more freedom to experiment: comparing larger models, retaining more context or running a generation model alongside supporting services. For a document assistant, these may include an embedding model, a result reranker, an index and document preparation processes. Their memory and computing requirements also need to be taken into account.

This extra capacity may avoid the need to replace the machine as the project evolves. It does not, however, automatically make the same requests faster: both GX10 models share the GB10 family and the stated bandwidth. If the model already fits comfortably within 64 GB and responds quickly enough, the main benefit of 128 GB is headroom.

The three listed models are GX10-GG0003BN with 1 TB, GX10-GG0026BN with 2 TB and GX10-GG0027BN with 4 TB. According to the specifications consulted, the first two use a PCIe 4.0 SSD and the third uses PCIe 5.0. For a lab storing several datasets and multiple quantised versions, 2 or 4 TB may be more practical; this choice does not change the 128 GB of memory available for computation.

The ExpertCenter Pro ET900N G3 combines a 72-core Arm NVIDIA Grace CPU with a Blackwell Ultra GPU. Its 748 GB coherent memory pool comprises 496 GB of LPDDR5X on the CPU side and 252 GB of HBM3e on the GPU side, connected through NVLink-C2C. This is therefore not 748 GB of HBM3e. ASUS states bandwidths of 396 GB/s and 7.1 TB/s respectively for the two memory types: their location and characteristics still matter.

This architecture is worth considering when a project exceeds a GX10's capacity or makes use of very fast GPU memory, with software capable of using the platform properly. It may appeal to research teams, developers of large models or those running a demanding shared service. It is not automatically the best choice for a small assistant used only occasionally.

Some variants add a separate RTX PRO card. The ET900N G3-8T-07 variant, for example, includes an RTX PRO 6000 Blackwell Max-Q with 96 GB of GDDR7 ECC memory. This memory does not automatically join the GB300 pool to form a uniform 844 GB. You need to identify the card's role: professional display output, another compute workload or a pipeline explicitly designed for multiple GPUs. Buying this option without software that uses it may unnecessarily increase the quote.

Enlarge

Inference means running an existing model. A RAG approach adds document retrieval to inform the model's answers: storage, indexing and data quality then become just as important as the model. A large document collection does not necessarily require a very large generation model; selecting the right passages may be more useful than simply increasing capacity.

Adapting a model with LoRA or QLoRA and performing full training each require a separate estimate. Activations, gradients and optimiser states can consume far more memory than the weights alone needed for inference. A model that can answer on a machine therefore cannot necessarily be trained on it with the same settings. Ask for validation of your adaptation method, batch size and sequence length.

The GX10 and ET900N are Arm-based. Check the versions of Python, PyTorch and CUDA, the native extensions, inference engines and container images you actually use. A container built only for x86-64 does not become compatible just because the machine has an NVIDIA GPU. The GX10 is advertised with NVIDIA DGX OS; ASUS documents Ubuntu and its NVIDIA tools for the ET900N, but the operating system supplied must be confirmed for each SKU.

You can keep a Windows laptop or a Mac as your client computer and access the workstation over the network. This does not remove the need to check server-side software, but avoids replacing your entire desktop environment. For a service handling confidential information, also check remote APIs, downloads and telemetry: running the model locally does not prove that the entire application stays out of the cloud.

Enlarge

On the GX10, the single M.2 slot means you need to size the SSD from the outset. Leave room for models, caches, environments and temporary data. An external backup is still necessary. On the ET900N, the listed configurations provide 4, 6 or 8 TB of raw capacity across two to four 2 TB SSDs. RAID and formatting affect usable capacity; mirroring protects against a drive failure, not accidental deletion.

The GX10 models have 10 GbE, Wi-Fi 7 and ConnectX-7. Connecting two machines for a distributed workload requires a compatible engine and an appropriate setup: two 64 GB GX10s do not behave like a single 128 GB GX10 in every application. Network communication and model partitioning can affect performance.

The ET900N adds two ConnectX-8 ports rated at up to 400 Gb/s each and BMC management with a dedicated Ethernet port. These links require the right modules, cables and network equipment; some SKUs include QSFP modules, while others do not mention them. Its stated weight of around 27 kg and 1,600 W power supply also require you to plan the location and room cooling. The GX10 power adapter's 240 W and the tower's 1,600 W are rated power figures, not measurements of power consumption or noise.

Enlarge

Prepare a representative set of requests using the exact model, its quantisation, the maximum expected context and the number of concurrent conversations. Measure model loading, time to first token and generation throughput separately. Repeat with long documents and a sustained workload: a good result on a short question does not describe the whole service.

Then compare answer quality and remaining memory. If the 64 GB GX10 meets the scenario with headroom for context and supporting services, it is a reasonable starting point in this range. If capacity is the limit, the 128 GB model becomes relevant. If throughput is the limit, do not assume doubling RAM will solve it: test another configuration or engine.

For the ET900N, require a quote that specifies the SKU, any RTX card, the operating system, licences, network modules, usable storage and support. Add electricity costs estimated from measured consumption, administration time and backups to the hardware cost. The commercial offer consulted is sold on a project basis subject to NVIDIA's approval; terms and lead times must be confirmed before planning deployment.

ASUS Ascent GX10 64 GB: who is this new configuration for?

ASUS Ascent GX10 128 GB: a compact machine for local AI development

ASUS ExpertCenter Pro ET900N G3: a GB300 workstation for large-scale AI projects

Local AI: ASUS expands its range, from the 64 GB GX10 to GB300 workstations