Every megawatt
has a fingerprint.

GB300 NVL72

≈ $108,345/yr to power one unit · ×72 GPU

Cooling · Direct liquid cooling at the rack/compute-tray level

Memory / DIMMs · Grace uses soldered LPDDR5; 1 TB aggregate CPU main memory per compute tray, not field-replaceable DIMMs.

Network · Derived rack total: 72 x ConnectX-8 HCAs and 18 x BlueField-3 DPUs from the per-tray specification.

142 kW is a published maximum for the full rack, not a measured workload average. Memory uses decimal GB. Storage count excludes management nodes and external storage. Installed PSU nameplate capacity must not be treated as rack draw.

GB300 NVL72 compute tray

Cooling · Direct liquid cooling

Memory / DIMMs · 1 TB aggregate LPDDR5 CPU main memory across two Grace CPUs; no replaceable DIMM population is specified.

Network · 4 x ConnectX-8 HCAs and 1 x BlueField-3 DPU per compute tray.

A tray is a component of the NVL72 system rather than a standalone server bill of materials. The five-device count is 1 M.2 OS device plus 4 E1.S NVMe cache devices and does not express capacity.

Recommended x86 control node for GB300 NVL72 AI Factory

Cooling · Not prescribed; depends on the selected OEM control node

Memory / DIMMs · 8 x 64 GB DDR5 DIMMs = 512 GB.

Network · 4 x ConnectX-7 200 Gb/s adapters.

The architecture recommends 12 management/control nodes, but this profile describes one node. CPU wording follows the reference configuration; it is not an endorsement of a particular CPU vendor or SKU.

Instinct MI355X accelerator

≈ $1,068/yr to power one unit · ×1 GPU

Cooling · Platform implementation may use passive or active cooling

Memory / DIMMs · 288 GB HBM3E accelerator memory; no host DIMM requirement is defined by the accelerator page.

Network · PCIe 5.0 x16 host interface; fabric and NIC design are platform-specific.

Do not use 1.4 kW as server or rack input power. AMD also reports 8 TB/s peak memory bandwidth and 185 billion transistors; neither is a facility-load metric.

8 x Instinct MI355X accelerator platform

≈ $8,546/yr to power one unit · ×8 GPU

Cooling · OEM/platform-specific; both passive and active accelerator implementations are supported

Memory / DIMMs · 2,304 GB aggregate HBM3E is derived; host CPU memory and DIMM population are unspecified.

Network · Host, scale-up fabric, NIC, DPU, and switch requirements are not included.

This is transparent screening arithmetic, not an AMD server BOM. It omits CPUs, DIMMs, storage, networking, cooling auxiliaries, PSUs, conversion losses, and redundancy; it cannot be compared directly with the complete 142 kW NVL72 rack rating.

NVIDIA DGX GB200 NVL72 (rack system)

≈ $91,560/yr to power one unit · ×72 GPU

Cooling · Hybrid cooling: GPUs and CPUs are direct liquid-cooled via cold plates fed by rack manifolds; networking and storage devices are air-cooled (per the same user guide).

Memory / DIMMs · Not stated in the reviewed NVIDIA GB200 documentation at rack-spec level (no CPU DIMM/memory-channel table published); accelerator memory is separately stated as 'Up to 13.4 TB HBM3e | 576 TB/s' on NVIDIA's DGX GB200 product page.

Network · Per compute tray (18 trays/rack): 4x NVIDIA ConnectX-7 single-port 400G OSFP NIC for cross-rack compute networking; 2x NVIDIA BlueField-3 DPU, dual-port 400G InfiniBand/Ethernet, for storage/management. Out-of-band: 1x 1GbE RJ45 from the compute tray BMC plus 2x 1GbE RJ45 from the BlueField-3 BMC interface. Rack also includes 9 NVLink switch trays providing the 72-GPU NVLink domain.

Covers only the DGX GB200 NVL72 rack itself (compute trays, NVLink switch trays, power shelves). Does not include external InfiniBand/Ethernet leaf-spine switches, external high-performance storage appliances, management nodes, or facility cooling (CDU/pump) overhead -- these are shown as separate racks in NVIDIA's own SuperPOD Scalable Unit diagrams and are folded into the separate, not-directly-comparable 1.2MW per-SU TDP figure.

NVIDIA DGX B200

≈ $10,911/yr to power one unit · ×8 GPU

Cooling · Air-cooled, 10 RU rack-mount chassis (per the datasheet's Rack Units spec).

Memory / DIMMs · 2TB system memory, configurable to 4TB, per the official datasheet ('System Memory: 2TB, configurable to 4TB'); specific DIMM count/type is not published in the datasheet.

Network · 4x OSFP ports serving 8x single-port NVIDIA ConnectX-7 VPI (up to 400Gb/s NVIDIA InfiniBand/Ethernet each); 2x dual-port QSFP112 NVIDIA BlueField-3 DPU (up to 400Gb/s InfiniBand/Ethernet); management network: 10Gb/s onboard NIC with RJ45, 100Gb/s dual-port Ethernet NIC, and host BMC with RJ45.

Power figure covers the DGX B200 server chassis only -- it does not include top-of-rack network switches, external storage, or facility-level PUE overhead. NVIDIA's DGX SuperPOD B200 guidance elsewhere discusses deploying 2 systems per 42U rack, which would roughly double this per-system figure, but no NVIDIA-published dual-system rack total was located for this candidate.

NVIDIA HGX B200 (8-GPU baseboard)

≈ $6,104/yr to power one unit · ×8 GPU

Cooling · Not specified for the baseboard alone. The datasheet states the baseboard 'can be paired with advanced air or liquid cooling solutions and additional server components' -- cooling design is left to the OEM system integrator.

Memory / DIMMs · Not applicable / not stated -- the baseboard itself carries no system DIMMs; host CPU memory is provided by the OEM server board, which is out of scope for this NVIDIA baseboard datasheet.

Network · Not stated in this document. The baseboard integrates fifth-generation NVLink and fifth-generation NVSwitch for GPU-to-GPU interconnect, with 14.4 TB/s total NVLink bandwidth. Host-facing networking (e.g., ConnectX/BlueField NICs) is determined by the OEM system and is not part of the baseboard reference.

Not suitable for facility-level PUE math -- this is a GPU-only TBP sum for a bare 8-GPU OEM baseboard, not a measured or vendor-stated system power draw. Real HGX B200-based servers from OEMs publish their own higher system-level power specs (including CPUs, DIMMs, storage, fans, PSU overhead) that were out of scope for this NVIDIA-primary-source-only candidate.

AMD Instinct MI300X Platform (Universal Baseboard, 8x OAM)

≈ $4,578/yr to power one unit · ×8 GPU

Cooling · Not stated in the platform data sheet — no air/liquid designation is given for the 8x OAM UBB baseboard itself.

Memory / DIMMs · Not specified — this data sheet documents the 8x OAM GPU baseboard only; no host CPU/DIMM memory configuration is given.

Network · 8x PCIe Gen 5 x16 (128 GB/s) host I/O, one per GPU; intra-baseboard AMD Infinity Fabric mesh at 7x 128 GB/s per GPU, 896 GB/s peak aggregate bidirectional ring bandwidth.

This is a GPU-baseboard-only figure (8x OAM accelerators on a UBB 2.0), not a full-server or full-rack power draw. It excludes host CPU(s), system memory, storage, networking, fans, PSUs, and AC/DC conversion losses, so it is not suitable for facility-level PUE math on its own — only for screening accelerator-silicon load.

AMD Instinct MI325X Platform (Universal Baseboard, 8x OAM)

≈ $6,104/yr to power one unit · ×8 GPU

Cooling · Not stated in the platform data sheet — no air/liquid designation is given for the 8x OAM UBB baseboard itself.

Memory / DIMMs · Not specified — this data sheet documents the 8x OAM GPU baseboard only; no host CPU/DIMM memory configuration is given.

Network · 8x PCIe Gen 5 x16 (128 GB/s) host I/O, one per GPU; intra-baseboard AMD Infinity Fabric mesh at 7x 128 GB/s per GPU, 896 GB/s peak aggregate bidirectional ring bandwidth.

This is a GPU-baseboard-only figure (8x OAM accelerators on a UBB 2.0), not a full-server or full-rack power draw. It excludes host CPU(s), system memory, storage, networking, fans, PSUs, and AC/DC conversion losses, so it is not suitable for facility-level PUE math on its own — only for screening accelerator-silicon load.

AMD Instinct MI300A APU (illustrative 4-APU node architecture)

Cooling · Air or liquid capable up to 550W Max TDP per socket; liquid-cooling-only above that, up to 760W Max TDP per socket.

Memory / DIMMs · 128 GB HBM3 unified memory per APU package, coherently shared between the integrated CPU cores and GPU compute units (not conventional host DIMMs). No separate host-memory/DIMM configuration is given since the data sheet covers only the APU package.

Network · Per APU: 8x 128 GB/s AMD Infinity Fabric interfaces (4 dedicated Infinity Fabric, 4 assignable to Infinity Fabric or PCIe Gen 5 x16); scale-out listed as 400 Gbps Ethernet or InfiniBand; two x4 NVMe storage interfaces and two USB interfaces per APU (device counts/models not stated).

No AMD-published platform/system total power figure exists for any multi-APU MI300A configuration in this primary source — only per-socket TDP is stated, and the document's own '4-APU' reference architecture carries no accompanying wattage. Third-party OEM systems (e.g., Supermicro 4-way MI300A servers, HPE Cray EX blades used in El Capitan) do publish their own node-level power figures, but those are vendor-integrator sources, not AMD primary sources, and are out of scope for this candidate. This entry is included to document that the search was made and no genuine AMD platform-power figure was found, per the 'unknowns stay unknown' policy.

Cloud TPU v4

≈ $130/yr to power one unit · ×1 GPU

Cooling · Not stated on this page (no cooling method specified for this measured-power figure).

Memory / DIMMs · Not stated on this page. HBM2 is on-package (32 GiB, 1200 GBps); no separate DIMM/memory-channel detail given.

Network · Six inter-chip interconnect (ICI) links per chip enabling a 3D mesh/torus pod topology; x16 PCIe gen3 interface to host (direct connect), per the same document.

Chip-only figure, not a system/rack/pod total. Does not cover host servers, networking, or cooling. Not directly comparable to full-server or full-rack power figures used elsewhere in this dataset. Google has not published an equivalent figure for TPU v5e, v5p, v6e (Trillium), or TPU7x (Ironwood) in any official documentation or blog post found during this research.

Cloud TPU v3

≈ $168/yr to power one unit · ×1 GPU

Cooling · Not stated on this page (no cooling method specified for this measured-power figure).

Memory / DIMMs · Not stated on this page. HBM2 is on-package (32 GiB, 900 GBps); no separate DIMM/memory-channel detail given.

Network · 2D torus pod interconnect topology per the same document; per-chip host-connect interface not detailed on this page.

Chip-only figure, not a system/rack/pod total. Does not cover host servers, networking, or cooling. Not directly comparable to full-server or full-rack power figures used elsewhere in this dataset. Google has not published an equivalent figure for TPU v5e, v5p, v6e (Trillium), or TPU7x (Ironwood) in any official documentation or blog post found during this research.

Amazon EC2 Trn2 UltraServer

Cooling · Not stated by AWS -- no cooling method or thermal design is published for the Trn2 UltraServer on the reviewed pages.

Memory / DIMMs · Not stated by AWS -- no host CPU/DIMM memory configuration is published for the Trn2 UltraServer on the reviewed pages.

Network · 64 Trainium2 chips connected via NeuronLink, delivering up to 83.2 FP8 petaflops of compute, 6 TB of total HBM3 with 185 TBps of total memory bandwidth, 12.8 Tbps of EFAv3 networking, and support for up to 32 TB of local NVMe storage.

Captures only the compute/memory/network specs AWS publishes for the Trn2 UltraServer. No power, thermal, cooling, DIMM, or physical host-CPU details are published at chip, instance, or aggregate UltraServer level. Local NVMe capacity (up to 32 TB) is stated but device count is not. Not usable for PUE or facility power planning without an independently sourced power figure.

Amazon EC2 Trn2 Instance (trn2.48xlarge)

Cooling · Not stated by AWS -- no cooling method or thermal design is published for trn2.48xlarge on the reviewed pages.

Memory / DIMMs · AWS states 2 TB of total instance memory for trn2.48xlarge (see cpu_memory_gb) but does not publish DIMM count, type, or channel configuration.

Network · 16 Trainium2 chips deliver up to 20.8 FP8 petaflops of compute, 1.5 TB HBM3 total with 46 TBps memory bandwidth, and 3.2 Tbps of EFA networking; supports up to 8 TB local NVMe storage. The trn2.48xlarge instance has 192 vCPUs and 2 TB of instance memory.

No power, thermal, cooling, DIMM, or physical host-CPU (as opposed to vCPU) details are published by AWS for the Trn2 instance. Not usable for PUE or facility power planning without an independently sourced power figure.

Amazon EC2 Inf2 Instance (inf2.48xlarge)

Cooling · Not stated by AWS -- no cooling method or thermal design is published for inf2.48xlarge on the reviewed pages.

Memory / DIMMs · AWS states 768 GiB of total instance memory for inf2.48xlarge (see cpu_memory_gb) but does not publish DIMM count, type, or channel configuration.

Network · inf2.48xlarge: 12 Inferentia2 chips, 384 GB total accelerator memory (32 GB per chip), 9.8 TB/s of total memory bandwidth, 192 GB/s NeuronLink interconnect between chips, 192 vCPUs, 768 GiB instance memory, and up to 100 Gbps network bandwidth.

No power, thermal, cooling, DIMM, or physical host-CPU (as opposed to vCPU) details are published by AWS for Inf2 instances. Not usable for PUE or facility power planning without an independently sourced power figure.

Intel Gaudi 3 AI Accelerator HL-325L OAM Mezzanine Card

≈ $687/yr to power one unit · ×1 GPU

Cooling · Not specified in this product brief (no air vs. liquid cooling designation is given alongside the 900W figure).

Memory / DIMMs · Not applicable / not stated — this is an accelerator mezzanine card, no DIMM slots are documented on it.

Network · 9.6 Tbps bi-directional networking capacity via 24x200 GbE RoCE v2 RDMA ports integrated on-die (native on-chip RDMA over converged Ethernet).

Single-accelerator component TBP only, not a system/server/rack figure. Cooling method not stated. Not itself an 8-accelerator platform figure — see the companion HLB-325 baseboard candidate for that.

Intel Gaudi 3 AI Accelerator HLB-325 Baseboard (8x HL-325L OAM accelerators)

≈ $5,799/yr to power one unit · ×8 GPU

Cooling · Not specified (no air vs. liquid cooling designation given). Brief states the baseboard 'derives the majority of its power from 54V power input, only requiring a separate 12V input for standby.'

Memory / DIMMs · Not applicable / not stated — the baseboard carries 8 accelerator OAM cards, not CPU DIMMs.

Network · Baseboard provides 4.2 TB/s bi-directional all-to-all bandwidth directly between the 8 onboard accelerators (no separate switching IC required), plus an additional 1.2 TB/s bi-directional scale-out bandwidth via 6 OSFP connectors (24x200 Gbps RoCE v2 RDMA).

This is a baseboard/accelerator-tray-only power figure explicitly excluding the host CPU subsystem, system memory, storage, and chassis cooling — it is not a complete server or rack figure and should not be used directly for facility-level PUE math without adding host-server overhead. No cooling method (air/liquid) is stated.

Intel HLS-Gaudi2 AI Accelerator Server

≈ $5,722/yr to power one unit · ×8 GPU

Cooling · Not explicitly stated in the datasheet (no air vs. liquid cooling designation given); described only as a 19-inch rack-mount server chassis.

Memory / DIMMs · 1TB of host system memory stated for the dual-socket Xeon 8380 subsystem ('Dual socket Intel 3rd Generation Xeon 8380 CPUs with 1TB memory'); DIMM count/type not specified. Separately, 768GB of HBM2E is on-package accelerator memory across the 8 Gaudi 2 cards (96GB each, stated directly in the spec table).

Network · 24x100GbE RoCE v2 RDMA scale-out via 6x QSFP-DD, plus internal non-blocking all-to-all interconnect (21x100GbE RoCE ports from each accelerator to the other 7), plus two 2x100GbE PCIe host NICs for host connectivity (per block diagram description).

Power figure is Intel's stated 'Max Power Usage,' likely a design/nameplate ceiling rather than a continuously-measured typical draw; cooling method, PSU redundancy, and AC/DC conversion overhead are not stated. This document describes Gaudi 2 (not Gaudi 3) and is dated October 2023 per its filename/copyright; hosted on habana.ai, Intel's wholly-owned AI-accelerator subsidiary domain, and is Intel-branded throughout ('Intel® HLS-Gaudi®2'). An Intel-hosted landing page for this same datasheet exists at intel.com/content-details/784779 but returned HTTP 403 when fetched directly.

CS-3 (Cerebras Wafer-Scale System, 1x WSE-3)

≈ $20,601/yr to power one unit · ×1 GPU

Cooling · Liquid cooling, direct-to-chip. Datasheet specifies: Water Filtering 50 micron; Water Chemistry PG25; Water Flow Rate 100 L/min/system; Water Temperature 20 ± 2°C.

Memory / DIMMs · Not stated. The datasheet lists 44 GB / 21 PB/s on-chip SRAM as processor memory; no host DIMM/DRAM configuration is given for the CS-3 system itself.

Network · 1.2 Tb/s I/O bandwidth; management: 1x 1GbE management port (RJ45), 1x console port (RJ45), 1x PoE flow controller interface.

Single-system power draw only; excludes any external MemoryX/SwarmX or networking cabinets used in multi-system training/inference clusters (the datasheet's separate 'Inference Cluster Configurations' table gives multi-cabinet provisioned heat loads instead, from ~240 kW for an 8-node/5-cabinet cluster up to multi-MW for 64-88 node clusters). No PUE/facility overhead included. CPU/host memory/storage not specified by this datasheet.

CS-3 Rack (2x Cerebras CS-3 systems)

≈ $41,202/yr to power one unit · ×2 GPU

Cooling · Combined liquid + air cooling. Datasheet states 'Water Cooling Load per Rack: 50.4 kW' and 'Air Cooling Load per Rack: 3.6 kW' (sums to the 54 kW total). Liquid cooling water coupling: 1.5" sanitary fittings, 2 pairs supply/return. Power feed: 2+1 redundant, 240V/415Y 3-phase, 60A, 5-wire, IEC 60309 plug.

Memory / DIMMs · Not stated.

Network · 8x QSFP-DD transceiver, 4x100G; management network: 1x 1GbE management port (RJ45), 1x console port (RJ45), 1x PoE flow controller interface.

Rack-level total for exactly 2x CS-3 systems only; excludes any additional WSE NET/interconnect cabinets used in larger multi-rack clusters (the datasheet's separate cluster-configuration table shows those as additional cabinets with their own heat loads, e.g. ~240 kW provisioned total for an 8-node/5-cabinet cluster). No PUE/facility overhead included.

GroqCard Accelerator (GC1-010B)

≈ $286/yr to power one unit · ×1 GPU

Cooling · Not stated in this spec sheet. GroqCard is a PCIe Gen4 x16 add-in card with no dedicated cooling spec published here; it relies on host-server chassis airflow.

Memory / DIMMs · Not applicable / not stated. GroqCard has no DIMM slots — its 230 MB of memory is on-die SRAM, not DRAM.

Network · Up to 11 RealScale chip-to-chip connectors for multi-server/multi-rack scalability without external switches; PCIe Gen4 x16 host interface, up to 31.5 GB/s bi-directional bandwidth.

Component-level (single PCIe accelerator card) power figure only — not a system, node, or rack total. Groq's GroqNode server datasheet states it packages up to 8x this card plus dual host CPUs (2x AMD EPYC 7313) and 1TB DRAM in a 4U chassis with '4 x 2000W (220-240VAC)' power supplies, but that is a PSU nameplate-capacity figure (likely N+redundant), not a stated power-draw number, so no full-node or full-rack GroqNode/GroqRack power figure is included here. We also could not locate a working GroqRack spec PDF with any published rack-level power figure as of this check.

Atlas 800T A2 Training Server

≈ $4,425/yr to power one unit

Cooling · Air-cooled; 8 hot-swappable fan modules in N+1 redundancy mode (per Hardware Description, Maintenance and Service Guide 05).

Memory / DIMMs · Up to 32 DDR4 DIMM slots (RDIMM), max speed 3200 Mbit/s, ECC/SEC-DED/SDDC + patrol scrubbing protected, 16 GB/32 GB/64 GB per module (all installed modules must share the same part number).

Network · CPU mainboard supports one FlexIO card providing 4x 25GE/10GE optical ports (PXE-capable); NPU carrier board supports up to 4 parameter-plane interface cards, each with 2x 200GE optical ports (fiber or copper).

Does not capture NPU (Ascend 910) accelerator count or exact model suffix (e.g. 910B) -- Huawei's own Overview page names the processor only generically as 'Ascend 910 AI Processors' without a quantity on the pages retrieved, so gpu_qty is left null and gpu_model reflects only the generic name actually published. Power figure is a manufacturer-stated maximum across configurations, not a measured or typical operating draw. Storage device count and accelerator HBM capacity are not stated on the retrieved pages.

Atlas 800I A3

≈ $11,750/yr to power one unit · ×8 GPU

Cooling · Air-cooled; deployable in standard air-cooled equipment rooms, using integrated fan modules plus Huawei-developed liquid-assisted air cooling (LAAC) modules. Each drawer integrates 5 fan modules with 4+1 hot-swap redundancy (single-fan failover time <= 2 minutes).

Memory / DIMMs · Not stated with specific capacity or DIMM type on the pages retrieved.

Network · Not stated with explicit port-count/speed figures on the pages retrieved. The server integrates separate 'LingQu UB' interconnect boards, documented by Huawei in a distinct 'LingQu UB Device Hardware' / 'LingQu UB Device Specifications' section, for supernode-to-supernode fabric connectivity.

Does not capture accelerator HBM memory capacity or interconnect bandwidth -- figures such as '128 GB HBM per NPU' or specific 400GE/UB link counts are widely reported by industry/trade press but were not found stated in these terms on the specific Huawei pages retrieved, so accelerator_memory_gb and detailed network specs are left null. The exact Ascend NPU variant (e.g. '910C') is not named as such in the retrieved Huawei text, which refers to it generically as an 'Ascend 910 AI module' (HiAM module). Power figure is a manufacturer-stated maximum across configurations, not a measured or typical operating draw, and excludes the separate LingQu UB Device module's own power draw.

NVIDIA DGX H100

≈ $7,783/yr to power one unit · ×8 GPU

Cooling · Air-cooled — front-to-back airflow; heat output ~38,557 BTU/hr; operating temp 5–30 °C.

Memory / DIMMs · 2 TB total system memory across 32 DIMMs.

Network · 4× OSFP serving 8× single-port ConnectX-7 InfiniBand (up to 400 Gb/s each); plus dual-port ConnectX-7 Ethernet for storage/in-band mgmt, and 10GbE + BMC out-of-band.

10.2 kW is the vendor's stated system MAXIMUM, not a typical/average draw; some third-party pages cite ~8.5 kW typical, which is not in any primary NVIDIA source and is not used here. Air-cooled in general practice; the extracted primary spec tables state airflow/heat output rather than the word 'air-cooled' explicitly.

NVIDIA DGX H200

≈ $7,783/yr to power one unit · ×8 GPU

Cooling · Air-cooled (standard configuration).

Memory / DIMMs · 2 TB total system memory (datasheet); DIMM count not on the datasheet — the DGX H100/H200 User Guide lists 32 DIMMs.

Network · 4× OSFP serving 8× single-port ConnectX-7 (up to 400 Gb/s InfiniBand/Ethernet); 2× dual-port ConnectX-7; management NICs + BMC.

10.2 kW is the standard air-cooled maximum. NVIDIA's datasheet footnote notes a separate Custom Thermal Solution (CTS) variant rated up to 14.3 kW, so a CTS-equipped unit would exceed this — the standard-config figure is used here.

NVIDIA DGX A100 (640GB)

≈ $4,959/yr to power one unit · ×8 GPU

Cooling · Not stated in the primary source.

Memory / DIMMs · Not stated in the primary source.

Network · 8× single-port ConnectX-6/7 VPI 200 Gb/s InfiniBand (clustering) + up to 2× dual-port ConnectX-6/7 VPI 10/25/50/100/200 Gb/s Ethernet (storage/networking).

6.5 kW is the datasheet maximum for the 640GB variant. The earlier 320GB DGX A100 (MAY20 datasheet) shipped with 1 TB system memory; this profile is the 640GB/2TB variant. Cooling method is not explicitly stated in the datasheet (only a 5–30 °C operating range), so it is left null.

NVIDIA HGX H100 (8-GPU SXM5 baseboard)

≈ $4,273/yr to power one unit · ×8 GPU

Cooling · Air- or liquid-cooled depending on the OEM system integration (the baseboard itself is not a complete server).

Memory / DIMMs · Not stated in the primary source.

Network · NVLink/NVSwitch on-baseboard fabric; host and cluster NICs are provided by the integrating server, not the baseboard.

5.6 kW is NVIDIA's stated TYPICAL baseboard power, not a maximum, and covers only the 8 GPUs + NVSwitch. It is not a facility or complete-server figure; a real HGX H100 server (host CPUs, memory, networking, PSU losses) draws materially more — see the DGX H100 profile (10.2 kW) for a complete-system comparison.

Cisco UCS C885A M8 Rack Server

≈ $9,537/yr to power one unit · ×8 GPU

Cooling · Air-cooled; 8RU rack server; operating 5–35 °C.

Memory / DIMMs · 24× 128 GB DDR5-6400 RDIMMs (alternate fixed config: 24× 96 GB).

Network · East-west: 8× ConnectX-7 (1×400G) or 8× AMD Pensando Pollara 400 per GPU; north-south: ConnectX-7 (2×200G / 1×400G / 4×25G); 1× OCP 3.0 Intel X710-T2L 2×10G.

12.5 kW is Cisco's stated maximum system power. The C885A M8 ships in multiple GPU configurations (NVIDIA HGX H100/H200/B200 or AMD MI350X OAM), so accelerator memory is not a single value and is left null; the 12.5 kW maximum is the whole-chassis envelope across supported configs.