A platform for the project.

A manufacturer-defined Grace Blackwell Ultra desktop superchip system for local AI development and memory-intensive workloads. This is a Gold Series reference configuration, not a configurable eight-GPU server.

Host CPU

1× 72-core NVIDIA Grace ARM Neoverse V2 CPU

GPU architecture

1× Blackwell Ultra GPU on the GB300 desktop superchip; 252 GB HBM3e GPU memory

Host / GPU memory

496 GB LPDDR5X CPU memory + 252 GB HBM3e GPU memory = 748 GB coherent memory. These are fixed onboard memories, not 748 GB of GPU VRAM or user-selectable RDIMMs.

Storage

Gold reference: 2× 960 GB M.2 NVMe + 2× 1.92 TB M.2 NVMe

Networking

ConnectX-8 SuperNIC, 2× 400GbE QSFP; 10GbE RJ45 and dedicated management

Power & cooling

Direct-to-chip cold plate cooling; 1× 1600 W Titanium PSU. Validate the exact cooling assembly, electrical input and room conditions with the ordered SKU.

Confirm ARM64 application/container support and the exact manufacturer BOM. This is a published manufacturer reference; the actual offered SKU and configuration are confirmed in the written quotation.

A Xeon 6 platform explicitly listing RTX PRO 6000 Blackwell Server Edition and H200 NVL. Start a platform-aware request for either GPU family.

Host CPU

Dual Intel Xeon 6900-series P-core processors; up to 500 W TDP each

GPU architecture

Up to eight double-width GPUs; lower-count BOMs and supported bridge kits require qualification

Host / GPU memory

24 DIMM slots; up to 6 TB DDR5 RDIMM at up to 6400 MT/s. MRDIMM is a separate CPU-dependent option, not part of the generic picker.

Storage

8 front hot-swap E1.S NVMe bays; 2 onboard M.2 NVMe slots

Networking

Dedicated 1GbE management; application NICs selected separately. GPU and NIC PCIe resources must be planned together.

Power & cooling

10 fans; 4× 3200 W Titanium, 3+1 design. Full module rating requires 220–240 VAC; lower voltages derate output. GPU-specific ambient limits apply.

NVLink bridges apply to supported H200 NVL groups, not RTX PRO 6000. An eight-card NVL server is not an eight-way HGX NVSwitch system.

The Intel reference behind NexTec’s eight-GPU RTX PRO 6000 build, with separate host-memory, storage and fabric planning.

Host CPU

Dual 4th/5th Gen Intel Xeon Scalable; air-cooled CPU TDP up to 350 W each

GPU architecture

NexTec reference: eight RTX PRO 6000 Blackwell Server Edition GPUs, 96 GB each

Host / GPU memory

32 DDR5 slots; up to 4 TB at 1DPC / 5600 MT/s or 8 TB at 2DPC / 4400 MT/s, subject to CPU and DIMMs

Storage

Default 8 NVMe + 8 SATA 2.5-inch bays; dual M.2. Optional additional NVMe bays require hardware review.

Networking

Dual onboard 10GbE RJ45; AIOM / PCIe application-fabric options

Power & cooling

4× 2700 W Titanium, 200–240 VAC. Validate full-load redundancy and the exact GPU/CPU/ambient qualification; maximum component ratings are not a qualified combined load.

The current double-width GPU envelope and selected qualification govern the build, not older ten-GPU gallery labels.

An EPYC host alternative for an eight-GPU RTX PRO 6000 inquiry, retaining twelve memory channels per CPU and networking headroom.

Host CPU

Dual AMD EPYC 9005/9004; up to 500 W TDP each. Builder preferences use 9005.

GPU architecture

Eight RTX PRO 6000 Server Edition GPUs in the reference request; do not infer ten-card qualification from the chassis maximum

Host / GPU memory

24 DDR5 slots; up to 6 TB. 9005 platform ceiling up to 6400 MT/s; 9004 up to 4800 MT/s.

Storage

8 NVMe data bays + 2 SATA bays; only one onboard M.2 slot

Networking

Dual onboard 10GbE RJ45; dedicated BMC and AIOM options

Power & cooling

6× 2700 W Titanium, 200–240 VAC; redundancy depends on actual load. Eight-GPU RTX qualification carries a 30°C ambient condition.

A two-drive SATA boot mirror can be requested; a native dual-M.2 boot mirror is not offered. Exact matched DIMMs and parts require qualification.

A fixed eight-GPU HGX baseboard rather than a set of interchangeable PCIe cards. Host memory, storage and external fabric remain project choices.

Host CPU

Dual 4th/5th Gen Intel Xeon Scalable; up to 350 W TDP each

GPU architecture

Eight H200 SXM GPUs, 141 GB HBM3e each; 1,128 GB aggregate GPU memory

Host / GPU memory

32 DDR5 slots; up to 4 TB at 1DPC / 5600 MT/s or 8 TB at 2DPC / 4400 MT/s

Storage

12 default NVMe + 3 SATA bays; two M.2 boot slots. Optional extra bays need additional hardware.

Networking

GPU-affiliated PCIe NIC slots; external cluster fabric is distinct from the internal NVSwitch fabric

Power & cooling

6× 3000 W Titanium, 4+2 design, with input-voltage derating. Exact GPU cooling implementation and required onsite service must be confirmed.

No two- or four-GPU SXM configurations are offered. Do not substitute H200 NVL PCIe cards or assume every H200 thermal BOM is air-only.

A Supermicro HGX B200 platform for a Blackwell system inquiry. It is an OEM server, not an NVIDIA DGX B200 appliance.

Host CPU

Dual Intel Xeon 6700/6500-series P-core processors, up to 350 W TDP each

GPU architecture

Fixed eight HGX B200 SXM6 GPUs, 180 GB each; 1,440 GB aggregate, NVSwitch fabric

Host / GPU memory

32 DDR5 slots; up to 4 TB at 1DPC / 6400 MT/s or 8 TB at 2DPC / 5200 MT/s

Storage

8 E1.S NVMe bays; two PCIe 5.0 M.2 slots

Networking

Dual 10GbE management/application interfaces plus GPU-affiliated NIC options; scale-out fabric quoted separately

Power & cooling

Air-cooled reference, 12 fans and 6× 6600 W Titanium, 3+3 design. Full PSU nameplate requires the specified high-voltage input; lower input voltages derate output.

GPU population is fixed at eight. CPU, DIMMs, NICs, storage, complete sustained load, rack and facility requirements need a qualified written proposal.

A compact GPU-capable host for projects where rack space matters. GPU selection is a qualified platform request, not a promise of two maximum-power cards.

Host CPU

Dual Intel Xeon 6700/6500 P-core or 6700 E-core families; exact model and BIOS to confirm

GPU architecture

Chassis supports up to two double-width or four single-width GPUs; exact GPU edition and thermal configuration must be reviewed

Host / GPU memory

32 DDR5 slots; capacity and speed depend on CPU, DIMM type and population

Storage

8 default 2.5-inch NVMe/SAS/SATA bays; optional bay changes need controllers/cabling; two M.2 slots

Networking

AIOM application networking; dedicated BMC management

Power & cooling

4 fans; 2× 2000 W Titanium, 1+1 design. Combined GPU/CPU load must fit the cooling and surviving-PSU envelope.

Published qualification rows limit H200 NVL to CPU ≤270 W / ambient ≤30°C and RTX PRO 6000 Server Edition to CPU ≤300 W / ambient ≤25°C. No maximum-load or redundant-load guarantee is implied; confirm the exact BOM.