Replacing a DGX A100 or HGX A100: 8× H200 NVL, 8× RTX PRO 6000 Server Edition or DGX B300

Upgrading a DGX A100 or HGX A100: what you replace and the three options

An eight-GPU A100 SXM system, a DGX A100 or an HGX A100 server from another maker, is best replaced by an eight-card H200 NVL server when the work needs HBM bandwidth, FP64 or NVLink within four GPUs, by eight RTX PRO 6000 Blackwell Server Edition cards when each model copy fits 96 GB, and by a DGX B300 when jobs need all eight GPUs in one NVLink domain and the rack can deliver about 15 kW. This is a whole-system replacement: the A100 SXM modules sit on a baseboard with NVSwitch and cannot be swapped for PCIe cards. Our guide to replacing A100 PCIe cards covers the card-for-card case.

NVIDIA DGX A100 user guide, "Introduction to the NVIDIA DGX A100 System", last updated 8 September 2026, read on 24 October 2026. Supplies rated 3,000 W at 200 to 240 V AC, 16 A; the system runs on three working supplies.

NVIDIA's A100 product page gives the 80 GB SXM module 2,039 GB/s of HBM2e and 400 W, and lists Tensor Core rates for TF32, BF16, FP16, INT8 and FP64, with no FP8. We found no end-of-life date for the DGX A100 published by NVIDIA, and third-party end-of-service lists are not NVIDIA statements, so the support contract you hold sets the deadline. On the software side, NVIDIA's vGPU lifecycle page, updated 2 October 2026, still lists the NVIDIA A100 among GPUs that have not reached the end of full support, and CUDA 13.0 dropped Maxwell, Pascal and Volta, not Ampere. The reasons to replace are memory, number formats and work per kilowatt, not a cut-off.

The replacements side by side

DGX A100 user guide and NVIDIA A100 page; H200 NVL figures from NVIDIA's H200 page as used in our eight-card H200 NVL server guide; RTX PRO 6000 Server Edition product page; DGX B300 user guide and NVIDIA HGX page as used in our DGX B300 specs and FP64 server guide. FP64 totals are our multiplication of per-GPU ratings (A100 9.7, H200 NVL 30, RTX PRO 6000 Server Edition about 1.9 TFLOPS); 10 TFLOPS is NVIDIA's HGX B300 figure for eight GPUs. Power of the PCIe servers is our estimate with two AMD EPYC 9555 and 15 per cent for the rest. MIG counts from NVIDIA's MIG user guide (13 October 2026) and our B300 article.

The H200 NVL server is the closest to the DGX A100 in character: HBM, FP64, NVLink and seven MIG instances per GPU, with 1.76 times the memory and 2.4 times the bandwidth per GPU. Its NVLink differs, though. Dell's XE7740 guide states that eight H200 NVL cards need two 4-way bridges, so the server has two NVLink domains of four, 564 GB each, and traffic between the domains crosses PCIe. The RTX PRO 6000 Server Edition server has less bandwidth per GPU than the A100 and no NVLink at all, and gains FP4, 96 GB per card and graphics. The DGX B300 is the only option that keeps an all-to-all NVLink fabric across eight GPUs, three times faster per GPU than the DGX A100's.

Power and rack: reusing the DGX A100's feed

A DGX A100 draws up to 6.5 kW. Both eight-card PCIe servers draw about 6.3 kW each by our estimate, which puts them in the same envelope: a rack position wired for a DGX A100 takes one of them, and a rack that held two DGX A100 systems at 13 kW takes two. Dell gives about 8 kW as an indicative maximum for an eight-card node, so the feed is planned on the maker's figure for the chosen server, and a measurement on the delivered system confirms our estimate or replaces it.

The DGX B300 is a different class of load. NVIDIA's user guide gives 14.5 kW, the datasheet of February 2026 lists 15.1 kW for the AC version with twelve 3.2 kW supplies, and the system is 10U and up to 168 kg. NVIDIA's data-centre guide for the DGX B300 plans two systems per rack on AC distribution at 39.4 kW peak. One DGX B300 needs more than twice the feed of the DGX A100 it replaces, and the rack and cooling plan comes first, as our DGX B300 specs and power requirements explain.

NVIDIA DGX A100 and DGX B300 documents as above; PCIe servers by our estimate. A 400 V 16 A three-phase feed is 8.87 kW at our 80 per cent planning margin, as in our guide to 16 A, 32 A and three-phase feeds.

Air matters as much as power. The DGX A100 accepts 5 to 30 °C at the inlet, the DGX B300 up to 30 °C, and Dell sets 30 °C for eight H200 NVL in the XE7740. A room that runs a DGX A100 at its limit therefore runs any of the three at theirs. We check rack, feed and airflow before we quote; send us the feed type, rack layout and inlet temperature of your DGX A100 position, and we say which replacement fits it without facility work.

Which jobs need all-to-all NVLink

The DGX A100's six NVSwitches let every GPU talk to every other at 600 GB/s. Whether the replacement needs the same depends on how your jobs use the eight GPUs.

• Tensor parallelism over eight GPUs: every layer runs two all-reduces across all GPUs. On a DGX B300 that stays inside NVLink at 1.8 TB/s. On eight H200 NVL, a model that fits one domain of four runs as TP=4, and a larger one as TP=4 within a domain and pipeline parallelism between them, as our eight-card H200 NVL guide shows.
• Fully sharded training and expert parallelism: FSDP or ZeRO-3 gathers parameters every layer, and mixture-of-experts routing sends tokens all-to-all. Across eight GPUs this traffic is what the NVSwitch fabric was built for; on the PCIe servers it crosses PCIe Gen5 at 128 GB/s per card.
• Data-parallel training of a model that fits one GPU or one domain: one gradient all-reduce per step, which the H200 NVL server handles inside its domains and the RTX PRO 6000 Server Edition server over PCIe.
• Serving copies of a model: one copy per card or per domain has no traffic between copies, and the interconnect does not matter.
• MIG tenants and notebooks: no traffic between instances at all; NVIDIA's MIG guide states that NCCL is not supported with MIG.

If your scheduler logs show that most jobs request one, two or four GPUs, the eight-card PCIe servers replace the DGX A100 in the same power. If they show regular eight-GPU jobs with FSDP or TP=8, the DGX B300 is the like-for-like successor, and a pair of four-domain H200 NVL nodes is the compromise, as our comparison of one 8-GPU server or two 4-GPU servers weighs.

Number formats and measured throughput

The A100 has no FP8 Tensor Cores. vLLM runs FP8 checkpoints on it as weight-only W8A16, so the memory halves and the arithmetic stays 16-bit. The H200 NVL computes in FP8; the RTX PRO 6000 Server Edition and the B300 add FP4. For an institute whose models moved to FP8 or NVFP4 checkpoints, this alone changes the throughput per GPU.

In the Kempner Institute's single-GPU training test of 15 July 2026, a GPT-style model of about 0.94B parameters in BF16 trained at 30,191 tokens per second on an A100 40 GB, 77,163 on an H200 and 41,370 on an RTX PRO 6000, 2.6 and 1.4 times the A100 by our division. The A100 there is the 40 GB version, and the handbook does not name form factors. MLCommons publishes no A100 result we cite here; for the replacements, MLPerf® Inference gives these Llama 2 70B figures per node.

MLPerf® Inference: Datacenter v6.1 (entries 6.1-0071, 6.1-0070, 6.1-0050; benchmark Llama 2 70B, llama2-70b-99.9 for 6.1-0070 and 6.1-0050) and v6.0 (entry 6.0-0021, llama2-70b-99), closed division, offline scenario. v6.1 retrieved from mlcommons.org on 24 October 2026, v6.0 on 24 September 2026; results verified by MLCommons Association. Different rounds and accuracy targets, so the H200 NVL row is not a like-for-like comparison. The MLPerf name and logo are registered and unregistered trademarks of MLCommons Association in the United States and other countries. All rights reserved. Unauthorized use strictly prohibited. See www.mlcommons.org for more information.

By our division of the per-node results, not an MLPerf metric, the DGX B300 delivered about 3.6 to 3.7 times the eight-card RTX PRO 6000 systems and about 3.4 times the eight-card H200 NVL server on this benchmark, at a little more than twice their power.

FP64: where the DGX B300 is the wrong successor

Many DGX A100 systems in research run double-precision codes, and here the newest system goes backwards. Eight A100 GPUs deliver about 78 TFLOPS of FP64 and 156 on their Tensor Cores, by our multiplication of NVIDIA's 9.7 and 19.5 TFLOPS per GPU. NVIDIA's HGX page lists 10 TFLOPS of FP64 for all eight B300 GPUs together. Eight H200 NVL deliver 240 TFLOPS of FP64 and 480 on the Tensor Cores, three times the DGX A100. Eight RTX PRO 6000 Server Edition cards come to about 15 TFLOPS, enough for code to run and too little for a solver. For CFD, molecular dynamics, VASP or Quantum ESPRESSO on a DGX A100, the eight-card H200 NVL server is the replacement; our FP64 GPU server guide covers the codes.

MIG: slices before and after

A DGX A100 with 80 GB GPUs offers up to 56 instances of 1g.10gb. Eight H200 NVL offer the same count of 1g.18gb instances, which NVIDIA's H200 page gives at 16.5 GB each on the NVL, so every notebook or small service gets 65 per cent more memory in the same number of slices. Eight RTX PRO 6000 Server Edition cards offer 32 instances of 24 GB, fewer and larger, with graphics profiles. A DGX B300 offers 56 instances of 1g.34gb. The profile names, GPU Operator layouts and Slurm types change with the card, as our guide to A100 MIG layouts on H200 NVL and RTX PRO 6000 maps profile by profile.

Recommendation by workload

• FP64 simulation, mixed HPC and AI, many MIG tenants: eight H200 NVL in one server, in the DGX A100's power envelope.
• Inference of models that fit 96 GB, FP4 serving, graphics or Isaac Sim next to AI: eight RTX PRO 6000 Server Edition, also within 6.5 kW by our estimate.
• Regular eight-GPU training with FSDP, TP=8 or expert parallelism, models above 1 TB: a DGX B300, with a facility plan for about 15 kW per system.
• A DGX A100 that serves several departments: two four-card servers, one with H200 NVL and one with RTX PRO 6000 Server Edition, keep a failure from taking out every service, at about the same total power.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. Our training node on AI servers has eight H200 NVL with NVLink bridges, two AMD EPYC processors, 2.3 TB of DDR5 ECC, eight NVMe drives and 400G networking, and we build eight-card RTX PRO 6000 Server Edition servers to order, assembled and tested before delivery, on one EU contract and invoice with manufacturer warranty on every component. The DGX B300 we supply per project; before we can quote it, the distribution asks for the company, the project and the intended use, and we check price and availability for your project. We do not supply the DGX A100 or the A100. Send us the job mix from your DGX A100, with GPU counts per job and the share of FP64 work, and we propose the replacement.

FAQ