Two servers, listed part by part.
We run our own work on both of these and build the same designs for clients. PLATINUM-96 is one Blackwell card with 96 GB of VRAM. SPARK-1 is an NVIDIA DGX Spark, 128 GB shared by the CPU and GPU as one address space. Figures here are either measured on these machines or marked as NVIDIA’s rating.
Good for. A QLoRA run on a 70B base takes about 38 GB of the 96, so past 100B at 4-bit still fits. Full-weight training is a different budget altogether — every parameter gets an optimiser state, and the practical ceiling is around 8B with an 8-bit optimiser and gradient checkpointing. Big-batch serving through vLLM or TensorRT-LLM fits here. So does diffusion training and molecular dynamics. This is also the machine our own fine-tunes run on, so the figures on this page were read off it and not lifted from a datasheet.
Full specification
| GPU | 1× NVIDIA RTX PRO 6000 Blackwell Workstation Edition — 96 GB GDDR7 with ECC, 128 MB L2 cache, PCIe 5.0 ×16, 4× DisplayPort 2.1, 600 W maximum power. Full Blackwell SM architecture: 5th-gen Tensor Cores with FP4/FP6/FP8/FP16/BF16/FP32, 4th-gen RT Cores, 9th-gen NVENC and 6th-gen NVDEC. 24,064 CUDA cores, 188 RT cores, 752 Tensor cores. FP64 runs on the CUDA cores at 1:64 rate, not on the Tensor Cores. |
|---|---|
| Memory bandwidth | 1,792 GB/s (GDDR7, 512-bit bus) |
| CPU | AMD Ryzen 9 7900X — 12C/24T, 4.7 GHz base / 5.6 GHz boost, Zen 4, TSMC 5 nm, 76 MB total cache (12 MB L2 + 64 MB L3), 170 W TDP, socket AM5 |
| System memory | 64 GB DDR5-6000, dual-channel, CL30-38-38-96, 2×32 GB DIMMs, EXPO enabled |
| Storage | 2× 2 TB PCIe 5.0 NVMe M.2 (TLC NAND, DRAM-cached, Phison E26) in RAID-0 — ~14 GB/s sequential read, ~12 GB/s write, ~1.5M IOPS random read per drive. Aggregate is bounded by the AM5 lane budget: with the GPU at ×16, only one M.2 is CPU-attached at PCIe 5.0 ×4 and the second sits behind the chipset link. |
| Networking | Realtek 2.5 GbE (onboard RTL8125BG). Remote management is over SSH; there is no BMC on this board. |
| Power | Seasonic Vertex PX-1200 — 1,200 W 80+ Platinum, fully modular, Cybenetics Platinum (ETA-A), ATX 3.1 / PCIe 5.1 compliant with a native 12V-2×6 connector. Measured system draw under sustained load is roughly 950–1,000 W at the wall. |
| Cooling | Noctua NH-D15S chromax.black — dual-tower, 6× heatpipe, 140 mm NF-A15 PWM. Chassis: 3× Noctua NF-A12x25 PWM intake, 1× exhaust. GPU: NVIDIA's dual-fan double-flow-through cooler. |
| Chassis & power protection | Fractal Design Meshify 2 XL (E-ATX), rackmount conversion in a 20U open-frame StarTech rack. CyberPower CP1500PFCLCD 1,500 VA / 1,000 W line-interactive UPS, pure sine wave output, AVR boost/buck. At full GPU load the system sits close to that 1,000 W ceiling, so the UPS is sized for ride-through and an orderly shutdown. It will not run the machine for long. |
| Software stack | Ubuntu 24.04 LTS (kernel 6.8 HWE), CUDA 12.8.1, cuDNN 9.7, NCCL 2.25, NVIDIA driver 570.x, Docker CE 27.x with nvidia-container-toolkit, JupyterHub 5.x multi-user, Prometheus 3.x + Grafana 11.x (node_exporter and dcgm-exporter for GPU telemetry) |
| Typical workloads | LLM LoRA/QLoRA fine-tuning past 100B at 4-bit, full-weight fine-tuning to ~8B params (8-bit optimiser, gradient checkpointing), large-batch inference serving (vLLM, TensorRT-LLM), molecular dynamics (GROMACS, OpenMM), diffusion model training (SD3, FLUX.1), GNN training on large graphs |
Good for. The 128 GB is a single pool. CPU and GPU address the same flat space across a coherent ~600 GB/s link, so a kernel runs on a tensor the CPU allocated and nothing gets copied over PCIe first. NVIDIA rates it for fine-tuning up to 70B. That tracks with what we have seen: a 70B base at 4-bit is about 38 GB, which leaves room for adapters, optimiser state and a long context. It suits long tool-calling workflows and RAG pipelines, and anything whose working set is awkwardly bigger than a normal card’s VRAM.
Full specification
| Superchip | NVIDIA GB10 Grace-Blackwell — 20-core ARMv9.2 CPU (10× Arm Cortex-X925 + 10× Cortex-A725) plus a Blackwell-class GPU on NVLink-C2C. Multi-die package on TSMC 3 nm, CPU die co-designed with MediaTek. Shared 128 GB unified memory pool addressable as a flat space by both CPU and GPU. |
|---|---|
| Unified memory | 128 GB LPDDR5x, 256-bit bus, 273 GB/s bandwidth. All 128 GB directly addressable by the GPU — kernels operate on CPU-allocated tensors with no transfer. |
| GPU compute | Blackwell-class GPU IP, 5th-gen Tensor Cores, FP4/FP6/FP8/FP16/BF16/FP32, dedicated Transformer Engine with FP8/FP4 block quantisation. Up to 1 petaFLOP FP4 with sparsity (1,000 NVFP4 TOPS). 9th-gen NVENC (AV1/HEVC), NVDEC, OFA. |
| NVLink-C2C | ~600 GB/s coherent fabric between the CPU and GPU dies, roughly 5× PCIe Gen 5 ×16. Cache-coherent across both domains; no PCIe bus contention. |
| Networking | NVIDIA ConnectX-7 SmartNIC — 200 Gb/s RDMA over 2× QSFP56-DD; separate 10 GbE RJ-45; Wi-Fi 7 (802.11be, 4×4 MIMO, 320 MHz channels), Bluetooth 5.4 |
| Storage | 1 TB PCIe Gen4 NVMe M.2 (TLC, DRAM-cached) — ~7 GB/s sequential read, ~5.2 GB/s write |
| Operating system | DGX OS 7 — NVIDIA-optimised Ubuntu 24.04 LTS derivative, NVIDIA DGX software stack, NCCL, CUDA, PyTorch containers, NVIDIA AI Workbench, NeMo Framework and NeMo Curator preinstalled |
| Software stack | CUDA 12.8.1, PyTorch 2.7+, TensorRT-LLM 0.16+, vLLM 0.8+, NVIDIA AI Workbench, NeMo Curator 0.8, NeMo Framework 24.12, Docker 27.x with nvidia-container-toolkit, optional Triton Inference Server |
| Typical workloads | LoRA/QLoRA fine-tuning up to ~70B (NVIDIA’s rated figure for this machine) and full-weight fine-tuning to ~10B, multi-step tool-calling agents with long context, RAG pipelines (ColBERT, hybrid search, multi-vector indexing), an LLM served on your own hardware, ONNX/TRT-LLM optimised serving, GPU-accelerated Jupyter development |
Does your job fit in 96 GB?
Pick the model size, the job and the context length. It is an estimate, and we would not sign a purchase order against it. But it shows its working, so you can see which term is eating the memory and argue with that one.
—
Assumptions. Batch size 1, gradient checkpointing on, AdamW. The KV cache is calibrated against a 70B grouped-query model at fp16 and grows linearly with context. Real numbers move with architecture, with the attention implementation, and with how badly the allocator fragments. Treat the answer as the right order of magnitude, and as a fair yes-or-no on whether the attempt is worth making. SPARK-1’s 128 GB pool is more forgiving still.
We buy
parts.
▓███████████████████████▒
▓███████████████████████▒
░▒▓▓█░ ░▒▓████████████████▒ ░░░░░░░░░░░ ░░░░░ ░░░░░ ░░░░ ░░░░░░░░░ ░░░░ ░░░░░░
░▒▓██▓▒░░▓█▓▓▒░ ░▓█████████████▒ ░████████████▓▒░ ▒████▒ ░████▒ ░████▒ ▓██████████▓▓░ ░████▒ ▒██████▓
▒▓██▒░ ░▒▒▒▒▓▓██▓░ ▒███████████▒ ░███████████████▒ ▓████░ ▓███▓ ░████▒ ▓█████████████▓░ ░████▒ ░████████░
░▒███▒ ░▒██▓▓▒░ ▒▓█▓▒ ▒█████████▒ ░████▒ ░▒█████░ ░████▒ ░████░ ░████▒ ▓████░ ░▒████▓ ░████▒ ▓███░▒████
▓███░ ▓██▒░ ▓█▓░ ▒███░ ░▓████████▒ ░████▒ ▒████▒ ▓████ ▓███▓ ░████▒ ▓████ ▒████░ ░████▒ ▒███▓ ▓███▓
░▓██▓ ▒██░ ▓███░▒███▒ ▒██████████▒ ░████▒ ░████▒ ░████▒ ░████░ ░████▒ ▓████ ░████▒ ░████▒ ░████░ ░████░
░▓██▒ ▒██▒ ▓██████▒░ ▒███▓░░▒▓████▒ ░████▒ ░████▒ ▓████ ▓███▓ ░████▒ ▓████ ░████▒ ░████▒ ▓███▒ ▒████
▓██▓ ░▓██▒▒▓▓▓▒░ ░▓███▓░ ░▓██▒ ░████▒ ░████▒ ░████▒░████░ ░████▒ ▓████ ▓████░ ░████▒ ▒████▓▓▓▓▓▓████▓
▒▓██▒ ░▒▓▒░░░░▒▓███▓▒░ ░▓████▒ ░████▒ ░████▒ ▒████▓███▒ ░████▒ ▒████▒▒▒▒▓█████▓ ░████▒ ░████████████████░
▒▓██▓░░ ▓█████▓▒░ ░▒▓███████▒ ░████▒ ░████▒ ████████ ░████▒ ▓█████████████▒ ░████▒ ████▓ ▓████░
░▒▓███▒ ░░▒▓▓███████████▒ ░▓▓▓▓░ ▓▓▓▓░ ▒▓▓▓▓▓▓▒ ░▓▓▓▓░ ▒▓▓▓▓▓▓▓▓▓▒▒░ ░▓▓▓▓░ ░▓▓▓▓░ ░▓▓▓▓▒
░▒▓▓▓▓▓██████████████████▒
▓███████████████████████▒
The manufacturers whose parts are in the machines. We list them so you can check the bill of materials against a supplier of your own. We have no partnership with any of them.
Want a machine like one of these?
Tell us the model and the query load. We will say which of the two fits, or that something smaller will do the job for less money.