GPUs & AI compute
NVIDIA and AMD Instinct data-center GPU specs and rack-scale AI server comparisons — GB300 / GB200 NVL72, HGX B300 / B200 / H200 / H100, AMD Instinct MI300X–MI355X, plus PCIe and L40S inference nodes. Compare memory, NVLink and Infinity Fabric bandwidth, power draw, weight, and reference architecture across Supermicro, Dell, HPE, Lenovo, and Gigabyte. Pricing on request.
NVIDIA
NVIDIA GB300 NVL72 (Supermicro MGX)
NVIDIA GB300 NVL72 rack-scale GPU system — 72 Blackwell Ultra GPUs and 36 Grace CPUs in a single, fully-integrated, liquid-cooled 19-inch Supermicro MGX rack.
NVIDIA
NVIDIA GB300 NVL72 (Lenovo 48U MGX)
NVIDIA GB300 NVL72 rack-scale GPU system — 72 Blackwell Ultra GPUs and 36 Grace CPUs on the Lenovo 48U MGX reference rack, factory-integrated with direct water cooling.
NVIDIA
NVIDIA GB300 NVL72 (HPE 48U MGX)
NVIDIA GB300 NVL72 rack-scale GPU system — 72 Blackwell Ultra GPUs and 36 Grace CPUs factory-integrated by HPE in a 48U liquid-cooled MGX-compliant rack.
NVIDIA
NVIDIA GB300 NVL72 (Pegatron)
NVIDIA GB300 NVL72 rack-scale GPU system — 72 Blackwell Ultra GPUs and 36 Grace CPUs in a single, fully-integrated, liquid-cooled Pegatron NVL72 rack.
NVIDIA
NVIDIA HGX B300 (Supermicro 8-GPU)
NVIDIA HGX B300 8-GPU Blackwell Ultra server — 2,304 GB HBM3e per node on a Supermicro platform, offered as a 4U liquid-cooled NVL8 or an 8U air-cooled build.
NVIDIA
NVIDIA HGX B300 (HPE ProLiant XD690)
NVIDIA HGX B300 8-GPU Blackwell Ultra server — 2,304 GB HBM3e per node on the air-cooled HPE ProLiant Compute XD690 with up to 4 TB of system memory.
NVIDIA
NVIDIA HGX B300 (Dell PowerEdge)
NVIDIA HGX B300 8-GPU Blackwell Ultra server — 2,304 GB HBM3e per node on a direct-liquid-cooled Dell PowerEdge platform.
NVIDIA
NVIDIA HGX B300 (Lenovo ThinkSystem SR680a V4)
NVIDIA HGX B300 8-GPU Blackwell Ultra server — 2,304 GB HBM3e per node on the Lenovo ThinkSystem SR680a V4 (8U) with ConnectX-8 800 Gb/s scale-out networking.
NVIDIA
NVIDIA HGX B300 (Gigabyte G894-SD3-AAX7)
NVIDIA HGX B300 8-GPU Blackwell Ultra server — 2,304 GB HBM3e per node on the Gigabyte G894-SD3-AAX7 (8U air-cooled), with an onboard ConnectX-8 8-port 800G SuperNIC and 3 TB of DDR5-6400.
NVIDIA
NVIDIA HGX B200 (Lenovo ThinkSystem SR780a V3)
NVIDIA HGX B200 8-GPU Blackwell server — 1,440 GB HBM3e per node on the Lenovo ThinkSystem SR780a V3 with Lenovo Neptune liquid cooling.
NVIDIA
NVIDIA HGX B200 (Supermicro 8-GPU)
NVIDIA HGX B200 8-GPU Blackwell server — 1,440 GB HBM3e per node on Supermicro’s 8U air-cooled platform, with a direct-liquid-cooled 4U build also available.
NVIDIA
NVIDIA HGX B200 (Gigabyte 8-GPU)
NVIDIA HGX B200 8-GPU Blackwell server — Gigabyte’s air-cooled 8-GPU node with 1,440 GB HBM3e and 64 TB/s of aggregate bandwidth.
NVIDIA
NVIDIA HGX B200 (Dell PowerEdge XE9680 / XE9685L)
NVIDIA HGX B200 8-GPU Blackwell server — 1,440 GB HBM3e per node on the air-cooled Dell PowerEdge XE9680, with a direct-liquid-cooled XE9685L build also available.
NVIDIA
NVIDIA HGX H200 (Dell PowerEdge XE9680)
NVIDIA HGX H200 8-GPU Hopper server — 1,128 GB HBM3e per node on the air-cooled Dell PowerEdge XE9680 with 4th-gen NVLink and ConnectX-7 NDR400 InfiniBand.
NVIDIA
NVIDIA HGX H200 (Supermicro 6U 8-GPU)
NVIDIA HGX H200 8-GPU Hopper server — 1,128 GB HBM3e per node on the air-cooled Supermicro 6U SXM platform, offered in networking-optimized (1:1 GPU:NIC) and compute-optimized (1:2) builds.
NVIDIA
NVIDIA HGX H200 (Gigabyte 8-GPU)
NVIDIA HGX H200 8-GPU Hopper server — 1,128 GB HBM3e per node on the air-cooled Gigabyte platform with dual AMD EPYC Genoa CPUs and ConnectX-7 NDR400 InfiniBand.
NVIDIA
NVIDIA HGX H200 (HPE Cray XD670)
NVIDIA HGX H200 8-GPU Hopper server — 1,128 GB HBM3e per node on the HPE Cray XD670 (5U) with 4th-gen NVLink and ConnectX-7 NDR400 InfiniBand.
NVIDIA
NVIDIA HGX H100 80GB — Liquid-Cooled (Supermicro 4U H13)
NVIDIA HGX H100 8-GPU server — liquid-cooled on the Supermicro 4U H13 with 8× H100 80GB SXM5, AMD EPYC Genoa, and ConnectX-7 NDR400 networking.
NVIDIA
NVIDIA HGX H100 80GB — Air-Cooled (Supermicro 8U H13)
NVIDIA HGX H100 8-GPU server — air-cooled on the Supermicro 8U H13 with 8× H100 80GB SXM5 and AMD EPYC Genoa, for standard air-cooled facilities.
NVIDIA
NVIDIA HGX H100 80GB — HP AI Server (8-GPU, Intel Sapphire Rapids)
HP 8-GPU NVIDIA HGX H100 AI server — 8× H100 80GB SXM (640 GB HBM3), dual Intel Xeon Platinum Sapphire Rapids, 2 TB DDR5, and NDR200 InfiniBand, with a 3-year warranty.
NVIDIA
NVIDIA HGX H100 (Dell PowerEdge XE9680)
NVIDIA HGX H100 8-GPU Hopper server — 640 GB HBM3 per node on the air-cooled Dell PowerEdge XE9680 with 4th-gen NVLink and ConnectX-7 NDR400 InfiniBand.
NVIDIA
NVIDIA HGX H100 (Gigabyte G593 8-GPU)
NVIDIA HGX H100 8-GPU Hopper server — 640 GB HBM3 per node on the air-cooled Gigabyte G593 platform with dual AMD EPYC Genoa CPUs and ConnectX-7 NDR400 InfiniBand.
NVIDIA
NVIDIA GB200 NVL72 (Dell PowerEdge XE9712)
NVIDIA GB200 NVL72 rack-scale GPU system — 72 Blackwell GPUs and 36 Grace CPUs in a single liquid-cooled Dell PowerEdge XE9712 (IR7000) rack.
NVIDIA
NVIDIA GB200 NVL72 (Supermicro SRS-GB200-NVL72)
NVIDIA GB200 NVL72 rack-scale GPU system — 72 Blackwell GPUs and 36 Grace CPUs in a single liquid-cooled 19-inch Supermicro NVL72 rack.
NVIDIA
NVIDIA GB200 NVL72 (HPE)
NVIDIA GB200 NVL72 rack-scale GPU system — 72 Blackwell GPUs and 36 Grace CPUs in a single liquid-cooled NVL72 rack integrated by HPE.
AMD
AMD Instinct MI300X (Dell PowerEdge XE9680)
AMD Instinct MI300X 8-GPU server — 1,536 GB HBM3 per node on the air-cooled Dell PowerEdge XE9680, with 192 GB per GPU for large memory-bound inference.
AMD
AMD Instinct MI300X (Supermicro AS-8125GS-TNMR2)
AMD Instinct MI300X 8-GPU server — 1,536 GB HBM3 per node on the Supermicro AS-8125GS-TNMR2 (8U air-cooled) with dual AMD EPYC CPUs.
AMD
AMD Instinct MI325X (HPE ProLiant Compute XD685)
AMD Instinct MI325X 8-GPU server — 2,048 GB HBM3E per node on the HPE ProLiant Compute XD685 (5U), 256 GB per GPU for the largest memory-bound models.
AMD
AMD Instinct MI355X (Supermicro 4U Liquid-Cooled)
AMD Instinct MI355X 8-GPU server — 2,304 GB HBM3E per node on Supermicro’s 4U direct-liquid-cooled platform, the CDNA 4 flagship with 288 GB per GPU.
NVIDIA
NVIDIA HGX A100 8-GPU (Supermicro AS-4124GO-NART)
NVIDIA HGX A100 8-GPU server — 640 GB HBM2e per node on the Supermicro AS-4124GO-NART, the deep secondary-market workhorse for training and inference.
NVIDIA
NVIDIA DGX A100 (640 GB)
NVIDIA DGX A100 — the 8× A100 SXM4 turnkey appliance with 640 GB HBM2e, a brand-recognized building block for AI training now trading on the secondary market.
NVIDIA
NVIDIA H100 NVL (PCIe Server)
NVIDIA H100 NVL PCIe server — 94 GB HBM3 per card with NVLink bridges, Hopper compute for facilities that want PCIe rather than SXM.
NVIDIA
NVIDIA H200 NVL (PCIe Server)
NVIDIA H200 NVL PCIe server — 141 GB HBM3e per card with up to 4-way NVLink, the highest-memory PCIe Hopper option for air-cooled racks.
NVIDIA
NVIDIA RTX PRO 6000 Blackwell Server Edition
NVIDIA RTX PRO 6000 Blackwell Server Edition — 96 GB GDDR7 passive PCIe cards, the mainstream air-cooled platform for enterprise inference and graphics.
NVIDIA
NVIDIA L40S (2U/4U PCIe Server)
NVIDIA L40S PCIe server — 48 GB GDDR6 Ada Lovelace cards for cost-efficient inference, fine-tuning, and graphics/VDI at volume.
TORmem
TORmem TOR-AI-Edge-2G
TORmem TOR-AI-Edge-2G — an enterprise AI edge server with dual NVIDIA RTX 5090 32GB GPUs and a 32-core AMD Threadripper PRO 9975WX, built for on-prem AI inference and LLM deployment.
Which one do I need?
Training frontier-scale models as one memory domain
GB300 NVL72 — 72 Blackwell Ultra GPUs in a single liquid-cooled rack, one NVLink domain end to end.
GB300 NVL72 specs→Scaling out in nodes that fit the racks you already have
HGX B300 or B200 — 8-GPU Blackwell nodes, air- or liquid-cooled, no rack-scale retrofit.
HGX B300 specs→Proven silicon, shorter lead time, lower power per node
HGX H200 or H100 — 8-GPU Hopper nodes still in volume supply, at roughly half the per-GPU draw of Blackwell Ultra.
HGX H200 specs→Inference, fine-tuning, and mixed workloads on PCIe
H100 NVL, H200 NVL, L40S, and RTX PRO 6000 cards in standard servers — no HGX baseboard, no liquid loop.
GPU vs server vs rack→The most HBM per GPU, on an AMD stack
Instinct MI355X and MI325X — Infinity Fabric instead of NVLink, in 8-GPU OAM nodes.
Instinct MI355X specs→NVIDIA & AMD GPU specs, power & datasheet comparisons
What are the key specs of a GB300 NVL72 or HGX H200 system?
A GB300 NVL72 rack packs 72 Blackwell Ultra GPUs (288 GB HBM3e each, roughly 20.7 TB per rack) on fifth-generation NVLink and draws about 132–140 kW liquid-cooled; an 8-GPU HGX H200 node carries 141 GB HBM3e per GPU (1,128 GB per node) at around 700 W per GPU. Every SKU page lists memory, interconnect, power, weight, and form factor.
How do I compare NVIDIA HGX B200 vs H200 vs AMD Instinct MI300X?
Compare on memory per GPU (B200 180 GB, H200 141 GB, MI300X 192 GB), interconnect (NVLink versus AMD Infinity Fabric), per-GPU power (B200 ~1,000 W, H200 ~700 W, MI300X ~750 W), and cooling (air versus direct-liquid). The GPU pages and the linked /learn comparisons break down each dimension.
What are the power and cooling requirements for these GPU servers?
Per-GPU board power runs from ~350 W (H100 NVL PCIe) to ~1,400 W (MI355X and GB300 NVL72). Air-cooled 8-GPU nodes typically need 5–14 kW; NVL72 racks are liquid-cooled at about 120 kW (GB200) to 140 kW (GB300). Each page lists the board or rack power and the cooling method.
How much does an NVIDIA H100 or GB300 server cost?
Pricing is not listed — GPU systems are quoted per configuration on request, since the figure depends on model, integrator, cooling, and networking. Tell us the build you need and we will return pricing and availability.
Research & comparisons
Keep me updated
We'll email you about changes across GPUs — availability, specs, or new units.