Lenovo ThinkStation PGX AI Developer Workstation NVIDIA GB10 - 128GB / 4TB
PRODUCT DESCRIPTION
Lenovo ThinkStation PGX Compact AI Developer Workstation, NVIDIA GB10 Grace Blackwell Superchip, 20-Core Arm CPU, 128GB Unified Memory, 4TB Self-Encrypting NVMe SSD, ConnectX-7 200GbE, 10GbE, Wi-Fi 7 (30KL000GSG) - Singapore Local Warranty
Lenovo ThinkStation PGX 4TB — Lenovo's First GB10 Grace Blackwell AI Workstation, 1.13 Litres on the Desk
The Lenovo ThinkStation PGX (part number 30KL000GSG, machine type 30KL) is the first Lenovo workstation accelerated by the NVIDIA GB10 Grace Blackwell Superchip and Lenovo's build of the NVIDIA DGX Spark platform. It pairs a 20-core Arm CPU with an integrated NVIDIA Blackwell GPU and 128 GB of coherent unified LPDDR5X memory in a 150 × 150 × 50.5 mm, 1.2 kg chassis, and Lenovo rates it at 1,000 TOPS, or 1 petaFLOP of FP4 AI performance with sparsity. Lenovo describes the PGX as designed solely for AI development: a single unit supports AI models of up to 200 billion parameters, and two PGX nodes linked over the integrated NVIDIA ConnectX-7 handle models of up to 405 billion parameters. This configuration carries the largest drive Lenovo offers on the platform: a 4 TB M.2 NVMe SSD with self-encryption.
For Singapore AI developers, data scientists, university research groups and enterprise AI teams, the 4TB PGX is the top tier of the range: the same GB10 compute, memory and networking as the 1TB unit, with a drive sized to keep several full-precision checkpoints, RAG indexes, datasets and an entire model zoo resident on one unit. SourceIT, an authorised Lenovo reseller in Singapore, supplies it with a full GST invoice, Singapore local warranty and enterprise, education and research-grant procurement support.
NVIDIA GB10 Grace Blackwell Superchip — 20 Arm Cores, Blackwell GPU, 5th-Generation Tensor Cores
The GB10 Superchip combines an NVIDIA Grace 20-core Arm CPU — ten Cortex-X925 performance cores and ten Cortex-A725 efficiency cores — with an NVIDIA Blackwell architecture GPU on one package, joined by NVLink-C2C so that CPU and GPU address one coherent memory space rather than copying data across a PCIe bus. The GPU carries Blackwell-generation CUDA cores, fifth-generation Tensor Cores with native FP4 precision, fourth-generation RT Cores for rendering and simulation, and one NVENC encoder and one NVDEC decoder for vision and media pipelines. Lenovo publishes the AI figure as 1000 TOPS, 1 PFLOP (FP4, sparsity), the same theoretical peak NVIDIA quotes for DGX Spark.
128 GB Unified Memory at 273 GB/s — 200B-Parameter Models on One Unit
Discrete workstation GPUs with 16–32 GB of VRAM stop well before a 70B-parameter model loads; the PGX's 128 GB of LPDDR5X sits on a 256-bit interface at 273 GB/s and is shared by the CPU and GPU as one pool, so Llama, Qwen, DeepSeek, Gemma, Mistral, gpt-oss and Nemotron-class models of up to 200 billion parameters run in full on a single unit, and NVIDIA's platform figures add fine-tuning of models up to 70 billion parameters. Memory is 128 GB on every PGX; Lenovo publishes no other capacity and no upgrade option.
4 TB Self-Encrypting NVMe SSD, ConnectX-7 200 Gbps, 10GbE RJ-45, Wi-Fi 7
Storage on 30KL000GSG is a 4 TB M.2 NVMe SSD with self-encryption (AES SED), the maximum Lenovo publishes for the platform. Networking is what sets a GB10 workstation apart from an ordinary mini PC: an integrated 200 Gbps NVIDIA ConnectX-7 NIC exposes two QSFP ports at 200 Gbps each for direct connection between PGX units, while a separate 10GbE RJ-45 port serves the office or lab network and is the right port for pulling model weights. Wi-Fi 7 (802.11be) and Bluetooth 5.3 cover wireless.
Four USB-C 20 Gbps Ports with DisplayPort 2.1, HDMI 2.1a, USB PD 3.1 Power Input
The rear panel carries four USB-C ports at 20 Gbps. One takes USB Power Delivery 3.1 input from the supplied 240 W power supply and also carries DisplayPort 2.1; the other three are data ports, each with DisplayPort 2.1, so a monitor and peripherals share a single cable. A dedicated HDMI 2.1a port with multichannel audio output adds a further display, alongside the 10GbE RJ-45 and the two QSFP ports.
1.13-Litre Chassis, 1.2 kg, 240 W USB-C Power Supply, TPM 2.0 Security
Lenovo lists the PGX as a small form factor system of 1.13 litres: 150 × 150 × 50.50 mm (5.91 × 5.91 × 1.99 inches), from 1.2 kg (2.65 lb). It runs from an ordinary wall socket through the included 240 W USB-C adapter; NVIDIA rates the GB10 Superchip at a 140 W TDP for CPU and GPU combined. Lenovo lists the security features as a self-encrypting NVMe drive (AES SED), TPM 2.0, the NVLink-C2C enclave, NVIDIA firmware recovery, AMI setup password and UEFI Secure Boot. Lenovo publishes no acoustic rating, operating-temperature range or certification list for the PGX, so none is claimed here.
NVIDIA DGX OS, CUDA 13, AI Workbench, GB10 Dashboard — and the NemoClaw Agentic Stack
The PGX ships preloaded with NVIDIA DGX OS, which Lenovo describes as Ubuntu Linux Pro with NVIDIA Base OS, together with the NVIDIA AI software stack, CUDA 13, the GB10 Dashboard, NVIDIA AI Workbench, PyTorch and Jupyter Notebooks. Because it is the same DGX software foundation as the NVIDIA DGX Spark, DGX Station and DGX Cloud, models prototyped here move to larger NVIDIA infrastructure with, in Lenovo's words, virtually no code changes. As a GB10 system it also runs the platform software NVIDIA publishes for DGX Spark: NemoClaw, NVIDIA's open-source reference stack that wraps OpenClaw agents with OpenShell policy-based privacy and security guardrails; NVIDIA AI Enterprise for DGX Spark; NIM microservices and Blueprints; and the DGX Spark Playbooks. Lenovo lists the suitable workloads as classic statistics, machine learning, computer vision, generative AI and LLMs, robotics and agentic design.
Scaling — Two or Three PGX Units Direct, More Through a QSFP Switch
Lenovo's specification sheet states that the 200 Gbps ConnectX-7 NIC supports direct connection between two or three ThinkStation PGX systems, and that connecting more than three requires a compatible QSFP switch; ConnectX-7 cables are required for every configuration and sold separately, with Lenovo listing the optional cable part 4X91U42988. Lenovo's published headline is two nodes for models up to 405 billion parameters; NVIDIA's platform figures go further, with up to four GB10 systems supporting models of up to 700 billion parameters. The 10GbE RJ-45 port stays free for site networking throughout.
Available in Singapore Through SourceIT
SourceIT is an authorised Lenovo reseller in Singapore and supplies the ThinkStation PGX 4TB (30KL000GSG) with a full GST invoice, Singapore local warranty and enterprise, education and research-grant procurement assistance. Lenovo's published base warranty for the ThinkStation PGX is 1-year limited onsite service; Lenovo Premier Support and Keep Your Drive are available as optional services and can be quoted on request. The PGX is also available from SourceIT in 1TB (30KLS00L00) on the same GB10 platform. Browse every GB10 and GB300 system we stock in the NVIDIA DGX Spark & DGX Station collection and the wider AI workstations and AI servers collection.
Lenovo ThinkStation PGX 4TB (30KL000GSG) Datasheet — Technical Specifications
General
| Product Name | Lenovo ThinkStation PGX Compact AI Developer Workstation — 4TB |
| SKU / Part Number | 30KL000GSG (machine type 30KL) |
| Category | Desktop AI development workstation / personal AI supercomputer (NVIDIA DGX Spark GB10 platform) |
| Form Factor | Small form factor, 1.13 litres; Lenovo PSREF lists the form factor as mini-tower |
| Colour / Finish | Not published by Lenovo in PSREF or the Lenovo Press product guide |
| Manufacturer Base Warranty | 1-year limited onsite service, per Lenovo PSREF; Premier Support and Keep Your Drive available as optional services |
| Local Warranty (Singapore) | Singapore Local Warranty — supplied by SourceIT with GST invoice and local support |
Processor & GPU — NVIDIA GB10 Grace Blackwell Superchip
| Architecture | NVIDIA Grace Blackwell — CPU + GPU on one superchip, NVLink-C2C interconnect |
| CPU | NVIDIA Grace 20-core Arm: 10 × Cortex-X925 performance cores + 10 × Cortex-A725 efficiency cores |
| GPU | NVIDIA Blackwell architecture GPU (integrated), Blackwell-generation CUDA cores |
| Tensor Cores | 5th generation (FP4 / FP8) |
| RT Cores | 4th generation |
| Media Engines | 1 × NVENC video encoder, 1 × NVDEC video decoder |
AI Performance
| AI Performance | 1000 TOPS, 1 PFLOP (FP4, with sparsity), per Lenovo |
| Model Support (single unit) | AI models up to 200 billion parameters, per Lenovo; fine-tuning up to 70 billion parameters per NVIDIA GB10 platform figures |
| Model Support (linked) | Two PGX nodes linked via integrated ConnectX-7: models up to 405 billion parameters, per Lenovo; up to four GB10 systems: models up to 700 billion parameters, per NVIDIA |
Memory
| System Memory | 128 GB LPDDR5X unified system memory, shared by CPU and GPU |
| Memory Interface / Bandwidth | 256-bit, 273 GB/s |
| Capacity Options | 128 GB only; no memory upgrade option published by Lenovo |
Storage
| SSD | 4 TB M.2 NVMe SSD with self-encryption (AES SED) |
| Platform Range | 1 TB or 4 TB M.2 NVMe with self-encryption, per Lenovo PSREF and Lenovo Press; PCIe generation not published by Lenovo |
Networking
| High-Speed NIC | Integrated NVIDIA ConnectX-7, 200 Gbps; 2 × QSFP ports (200 Gbps each) for connecting ThinkStation PGX units |
| Ethernet | 1 × 10GbE RJ-45 |
| Wi-Fi | Wi-Fi 7, 802.11be |
| Bluetooth | Bluetooth 5.3 LE |
| Clustering Cable | Sold separately; Lenovo lists optional ConnectX-7 cable 4X91U42988 |
I/O & Display
| USB (power in) | 1 × USB-C 20 Gbps (USB 3.2 Gen 2x2) with USB PD 3.1 input and DisplayPort 2.1 |
| USB (data / display) | 3 × USB-C 20 Gbps (USB 3.2 Gen 2x2) with DisplayPort 2.1 |
| Video | 1 × HDMI 2.1a; DisplayPort over USB-C |
| Audio | HDMI multichannel audio output |
| Front Panel | Power button |
Software & Security
| Operating System | NVIDIA DGX OS — Ubuntu Linux Pro with NVIDIA Base OS, arm64 |
| Preinstalled Software | NVIDIA AI Software Stack, CUDA 13, GB10 Dashboard, NVIDIA AI Workbench, PyTorch, Jupyter Notebooks |
| Security | Self-encrypting NVMe (AES SED), TPM 2.0, NVLink-C2C enclave, NVIDIA FW Recovery, AMI Setup Password, UEFI Secure Boot, administrator password, boot sequence control |
Power, Physical, Acoustics & Certifications
| Power Supply | 240 W USB-C power supply, included |
| Superchip TDP | 140 W (CPU + GPU), per NVIDIA GB10 platform specification |
| Dimensions (W × D × H) | 150 × 150 × 50.50 mm (5.91 × 5.91 × 1.99 in) |
| Volume | 1.13 litres |
| Weight | From 1.2 kg (2.65 lb) |
| Acoustics | Not published by Lenovo |
| Certifications | Not published by Lenovo in PSREF; Lenovo Press lists RoHS compliance and 90% recycled or FSC-certified carton content |
| In the Box | ThinkStation PGX unit, 240 W USB-C power supply, documentation |
Compatibility & Software Support
What Runs on the ThinkStation PGX
Anything in the NVIDIA and open-source AI ecosystem published for arm64: Hugging Face Transformers and Diffusers, vLLM, Ollama, llama.cpp, LangChain and LlamaIndex RAG pipelines, NVIDIA NIM and Blueprints, NeMo, TensorRT-LLM, ComfyUI, RAPIDS, and the NVIDIA Isaac, Metropolis and Holoscan frameworks. NVIDIA's DGX Spark Playbooks apply directly because the PGX is the same GB10 platform.
Arm64 Linux — Plan Your Software Accordingly
The GB10's Grace CPU is Arm-based and the PGX ships with NVIDIA DGX OS, so the operating system, drivers and container images are arm64 (aarch64) builds. Every major AI framework and NVIDIA NGC container is published for arm64, but x86-only proprietary desktop software will not run natively and Windows is not supported on this machine. Most teams drive the PGX from a laptop over SSH, Jupyter or VS Code Remote and keep Windows applications there. For a Windows 11 Arm desktop in the same collection, see the ASUS Ascent QN10; for a Windows x86 large-memory alternative, see the GMKtec EVO-X2.
Linking Two or Three PGX Units
Two PGX units connect back-to-back over their ConnectX-7 QSFP ports with a direct-attach cable; Lenovo also allows three units connected directly, with a compatible QSFP switch beyond that. Cables are sold separately: Lenovo lists its optional ConnectX-7 cable 4X91U42988, and SourceIT stocks the NVIDIA QSFP112 400G DAC cable, 40 cm (X0101G00400A) used across the DGX Spark family. Two linked units pool 256 GB of unified memory for models up to 405 billion parameters.
Where the PGX Sits in the GB10 Family
The PGX is one of the GB10 desktops SourceIT stocks alongside the NVIDIA DGX Spark, with ASUS, Dell, HP, GIGABYTE, MSI, Acer and xFusion builds. All share the GB10 Superchip, 128 GB unified memory and 1 PFLOP FP4; the PGX is distinguished by Lenovo's ThinkStation support path, its TPM 2.0 and firmware-recovery security list, preinstalled CUDA 13, AI Workbench and GB10 Dashboard, and a published three-unit direct-connect option.
Package Contents & Ordering Information
What Ships in the Box
- Lenovo ThinkStation PGX (30KL000GSG) with NVIDIA GB10 Grace Blackwell Superchip, 128 GB unified memory and 4 TB self-encrypting M.2 NVMe SSD
- 240 W USB-C power supply
- NVIDIA DGX OS preinstalled with the NVIDIA AI software stack
- Lenovo documentation
Monitor, keyboard and mouse are not included; the QSFP cable is sold separately.
Lenovo ThinkStation PGX Range at SourceIT Singapore
- Lenovo ThinkStation PGX 1TB — 30KLS00L00, 128 GB unified memory, 1 TB self-encrypting NVMe
- Lenovo ThinkStation PGX 4TB — 30KL000GSG (this listing), 128 GB unified memory, 4 TB self-encrypting NVMe
- NVIDIA QSFP112 400G DAC Cable, 40 cm — X0101G00400A (links two units)
Both tiers share the same GB10 Superchip, memory, networking, ports and chassis; only SSD capacity differs.
Other NVIDIA GB10 and DGX Station Systems at SourceIT
- NVIDIA DGX Spark 4TB — NVIDIA's own first-party GB10 unit
- ASUS Ascent GX10 — 1TB / 2TB / 4TB
- Dell Pro Max with GB10 — 2TB / 4TB
- HP ZGX Nano G1n — 4TB
- GIGABYTE AI TOP ATOM — 1TB / 4TB
- MSI EdgeXpert MS-C931 — 4TB
- Acer Altos BrainSphere GB10 F1 — 4TB
- xFusion FusionXpark GB10 — 1TB / 2TB / 4TB
- HP ZGX Fury G1n (NVIDIA DGX Station GB300) — step up to 748 GB coherent memory and 20 PFLOPS for trillion-parameter models
Browse the full NVIDIA DGX Spark & DGX Station collection and the AI workstations and AI servers collection. For volume, project, research-grant or education pricing, GST invoicing and current lead times, email sales@sourceit.com.sg quoting 30KL000GSG.
Datasheet, Downloads & Setup Guide
Official Lenovo and NVIDIA documentation for the ThinkStation PGX and the GB10 platform:
- Lenovo ThinkStation PGX — product page
- Lenovo PSREF specification sheet — ThinkStation PGX (PDF)
- Lenovo Press LP2321 — ThinkStation PGX Product Guide
- NVIDIA DGX Spark (GB10) — official platform page (Singapore)
- NVIDIA DGX Spark Playbooks — step-by-step recipes for LLMs, fine-tuning, RAG and agents
- NVIDIA DGX Spark Setup and Enterprise Deployment Guide — SourceIT (applies to the PGX)
For a Lenovo configuration sheet or quotation, email sales@sourceit.com.sg quoting 30KL000GSG.
Frequently Asked Questions — Lenovo ThinkStation PGX 4TB
What is the Lenovo ThinkStation PGX?
Lenovo's compact AI developer workstation on the NVIDIA DGX Spark GB10 platform: a 1.13-litre, 1.2 kg desktop with the NVIDIA GB10 Grace Blackwell Superchip, 128 GB of unified memory and 1 PFLOP FP4 with sparsity, running NVIDIA DGX OS for local LLM inference, fine-tuning, data science, vision, robotics and agentic AI.
What is the difference between the ThinkStation PGX and the NVIDIA DGX Spark?
Same NVIDIA GB10 Superchip, 128 GB unified memory, 1 PFLOP FP4 and ConnectX-7 networking, so AI performance is identical. The PGX is Lenovo's build with ThinkStation support, TPM 2.0 and firmware-recovery security, CUDA 13, AI Workbench and the GB10 Dashboard preinstalled, and a choice of 1TB or 4TB; the DGX Spark is NVIDIA's first-party unit in 4TB only.
Should I choose the 4TB or the 1TB ThinkStation PGX?
Compute, memory and networking are identical. A 70B-parameter model at 4-bit is roughly 40 GB and at 16-bit roughly 140 GB, so 4TB keeps several full-precision checkpoints, RAG indexes and an entire model zoo on one unit, while the 1TB (30KLS00L00) holds a working library of quantised models plus datasets and containers. Lenovo publishes no storage upgrade path, so choose the tier at purchase.
What AI models can the ThinkStation PGX run?
Lenovo states support for AI models up to 200 billion parameters on one unit, and NVIDIA's platform figures add fine-tuning up to 70 billion — Llama, Qwen, DeepSeek, Gemma, Mistral, gpt-oss and Nemotron-class models in practice. Two linked units handle up to 405 billion parameters per Lenovo; four GB10 units reach 700 billion per NVIDIA.
Can I link two ThinkStation PGX units together?
Yes. Lenovo's specification sheet allows direct connection of two or three PGX systems over the ConnectX-7 QSFP ports, with a compatible QSFP switch required for more. Cables are sold separately — Lenovo lists part 4X91U42988, and SourceIT stocks the NVIDIA QSFP112 400G DAC 40 cm cable.
Does the ThinkStation PGX run Windows or x86 software?
No. It ships with NVIDIA DGX OS (Ubuntu Linux Pro with NVIDIA Base OS, arm64) and is a Linux AI development system. Mainstream AI frameworks and NVIDIA NGC containers are published for arm64; proprietary x86-only desktop applications will not run natively.
What is NemoClaw and does the PGX support it?
NemoClaw is NVIDIA's open-source reference stack that adds OpenShell policy-based security and privacy guardrails around OpenClaw agents and runs open models locally. NVIDIA publishes it for the DGX Spark GB10 platform and DGX OS includes a streamlined NemoClaw installation, so it applies to the PGX.
Does it need special power or cooling?
No. It runs from a standard socket through the included 240 W USB-C power supply; the GB10 Superchip is rated by NVIDIA at 140 W TDP. Lenovo publishes no acoustic figure for the PGX.
Can I upgrade the memory or storage?
Memory is 128 GB on every PGX and Lenovo publishes no upgrade option. Storage is offered as 1 TB or 4 TB self-encrypting NVMe and Lenovo publishes no storage upgrade path, so treat the configuration you buy as the configuration you keep.
Who is the ThinkStation PGX for?
Lenovo names AI developers, data scientists, students, researchers and AI enthusiasts prototyping, fine-tuning and running inference locally before deploying to the data centre or cloud; in Singapore that includes university labs, robotics and computer-vision teams, and enterprises in finance, healthcare, legal and government that must keep models and data on-premise.
What warranty does the ThinkStation PGX come with in Singapore?
This listing carries a Singapore Local Warranty supplied through SourceIT. Lenovo's published base warranty for the PGX is 1-year limited onsite service; Lenovo Premier Support and Keep Your Drive are optional services we can quote. Every unit comes with a GST invoice and local support.
How do I buy the Lenovo ThinkStation PGX in Singapore?
Order on this page or email sales@sourceit.com.sg quoting 30KL000GSG. SourceIT is an authorised Lenovo reseller in Singapore and supplies the PGX with GST invoice, local warranty and the QSFP clustering cable.
