Digital pathology GPU server: whole-slide AI on the RTX PRO 6000 Blackwell Server Edition

What a GPU server does in digital pathology

A GPU server for digital pathology runs the AI that reads scanned glass slides: it cuts each whole-slide image into tiles, turns every tile into an embedding with a foundation model, and lets small task models classify, grade or search the slide from those embeddings. For a pathology institute that scans a few hundred slides a day, a server with two NVIDIA RTX PRO 6000 Blackwell Server Edition cards carries the embedding pipeline and the research work on it; a laboratory with a few thousand slides a day plans four cards. The slide counts per card below are our estimates from published compute figures, and a pilot on your own slides sets the final number.

We wrote this for the head of a pathology institute or a laboratory network who has to judge what such a server does and what it changes, and for the IT team that connects it to the scanners and the slide archive. Radiology works differently: CT and MRI are 3D volumes with their own models and the PACS, which our guide to the medical imaging AI server covers, and language models for clinical documentation are the subject of our guide to a private LLM server for hospitals.

One slide in numbers

A whole-slide image is too large for any model to take whole. A 2026 benchmark paper on arXiv (2607.18218, 20 July 2026) uses a slide from the EBRAINS dataset that splits into 31,469 non-overlapping tiles of 256 × 256 pixels at 20× magnification. That is about 2.06 billion pixels, or 6.2 GB as uncompressed RGB, by our arithmetic. A scan at twice the linear resolution holds four times the pixels, about 24.7 GB uncompressed. The OpenSlide project states that whole-slide images "often" occupy tens of GB uncompressed and routinely exceed RAM sizes, so software reads them tile by tile from a pyramid of resolutions.

Tile count from arXiv 2607.18218 (GigaPath-Flash paper, read 24 October 2026); output size of 1,536 from the H-optimus-0 model card on Hugging Face (read 24 October 2026); the rest is our arithmetic. Scanners store slides compressed, so the files are much smaller than these figures; use the average file size from your own archive.

Tile size changes the count. H-optimus-0 expects tiles of 224 × 224 pixels at 0.5 µm per pixel, and cutting the same tissue into 224-pixel tiles gives about 1.3 times as many tiles, by our arithmetic. The real count depends on the tissue area of each slide.

Open pathology foundation models and their licences

A pathology foundation model is a vision transformer trained on millions of tiles without labels. It turns each tile into a vector, and a small model on top, often an attention-based aggregator, learns the diagnostic task from labelled slides. The licence of the encoder decides whether a company or a hospital may use the model at all.

Model cards on Hugging Face (bioptimus/H-optimus-0, bioptimus/H-optimus-1, MahmoodLab/UNI, prov-gigapath/prov-gigapath), read 24 October 2026. All four are gated: access requires accepting the conditions, and H-optimus-1 and UNI ask for an institutional e-mail address.

The H-optimus-0 card describes a "1.1B parameter vision transformer" trained on "more than 500,000 H&E stained whole slide histology images", and its file list shows a 4.54 GB checkpoint, which matches 1.1B parameters stored at four bytes each. In FP16 the weights take about 2.2 GB, by our arithmetic, a small part of a 96 GB card; the model card recommends mixed precision for faster inference. The Prov-GigaPath card carries an Apache 2.0 tag and still places deployed uses, commercial ones included, outside its intended scope. Read each licence and card with your legal department before a model enters a diagnostic workflow.

Slides per day per card

Embedding a slide is pure compute: the same tile encoder runs tens of thousands of times. The benchmark paper computes the work per slide for its 31,469-tile example, and we divide it by the card's tensor rate.

NVIDIA's page for the RTX PRO 6000 Server Edition lists 1 PFLOP of FP16 and BF16 without saying whether that figure assumes sparsity, and NVIDIA publishes no dense rate for this edition. Its whitepaper gives 503.8 dense BF16 TFLOPS for the Workstation Edition; the Server Edition has the same core count and a 4 per cent lower FP32 rate, so we take about 480 dense TFLOPS, by our estimate.

TFLOP per slide from arXiv 2607.18218, Table 2 (31,469 tiles; the GigaPath figure includes its slide encoder); floor and ceiling by our estimate at about 480 dense TFLOPS per RTX PRO 6000 Server Edition, full rated rate, running all day. Real runs reach a fraction of the rated rate.

These are upper bounds. No published measurement of these encoders on a Blackwell card was found as of October 2026, and a real pipeline also reads, decodes and filters tiles. An encoder that takes 224-pixel tiles, such as H-optimus-0, sees about 1.3 times as many tiles per slide, so its work per slide is not the GigaPath figure either. If a card reaches half of its rated rate, which a pilot has to confirm, a 1B-class encoder takes about 60 seconds per slide and one card embeds about 1,450 slides in a day of continuous running. A department with 500 slides a day then needs about 8.3 hours of one card; a laboratory with 2,000 slides a day needs about 17 hours on two cards or about 8 hours on four.

The task models on top are light. An aggregator trained on precomputed embeddings reads about 97 MB per slide instead of the image, so training a new classifier takes a small part of the compute of the embedding run. Fine-tuning the encoder itself is heavier and follows the rules of our guide to GPU servers for training vision models.

We supply the RTX PRO 6000 Server Edition as cards and in servers built to order; send us your slides per day, your scanners' file formats and the models you plan, and our team replies with a configuration.

Reading slides fast enough

The GPU waits if the server cannot read tiles quickly. OpenSlide, a C library under the LGPL 2.1 (version 4.0.1 of 7 June 2026), reads Aperio, DICOM, Hamamatsu, Leica, MIRAX, Philips, Ventana, Zeiss and other formats, with the caveat on its site that this support is "not endorsed by their respective vendors and may be incomplete". NVIDIA's cuCIM, under Apache 2.0, offers an API that matches OpenSlide's, reads Aperio SVS, Philips TIFF and generic tiled TIFF, and its cuslide2 module uses nvImageCodec for GPU-accelerated TIFF decoding. Packages exist for CUDA 12 and CUDA 13.

Plan the processor side for this. NVIDIA's guide for certified systems asks for at least six cores per GPU and system memory of twice the GPU memory, and tile decoding on the CPU adds to that. Test the reader with your scanners' own files before you settle the core count.

Configuration options for a pathology institute

The options follow the reference configurations on our AI servers page and NVIDIA's guide for certified systems. Slide counts assume a 1B-class encoder at half the card's rated rate, which a pilot confirms or corrects.

Reference configurations on our AI servers page; slides per day by our estimate (about 60 s per slide per RTX PRO 6000 Server Edition, running all day); power by our estimate as in our server guides (cards, processors and 15 per cent for the rest). NVIDIA publishes no dense tensor rate for the RTX PRO 4500 Server Edition, so we give no slide count for it.

Two cards. The entry server has two RTX PRO 6000 Server Edition, two AMD EPYC processors, 384 GB of DDR5 ECC and two NVMe drives. One card embeds the daily slides; the other splits by MIG into up to four isolated 24 GB instances, as our guide to RTX PRO 6000 MIG profiles shows, for task models, a test instance and the researchers. It fits one 230 V 16 A feed, which carries 2.94 kW at 80 per cent.

Four cards. The inference node with 768 GB and four NVMe drives carries a laboratory network or a reference centre. Three cards embed the daily slides and the fourth trains new task models, or all four embed during the night. It draws about 3.4 kW with the cards at 600 W and needs 32 A or three-phase; capped at 450 W it draws about 2.7 kW.

RTX PRO 4500 Server Edition. Each passive single-slot card has 32 GB and 165 W and splits into two 16 GB MIG instances. A 1.1B encoder in FP16 fits easily, but a card with less compute embeds fewer slides; this server suits task models and inference at a smaller site, while the embedding of large volumes stays on the RTX PRO 6000.

The RTX PRO 6000 Max-Q is the workstation card, a 300 W blower card; in a rack it belongs only in a chassis whose maker lists it, and for continuous operation we recommend the Server Edition. A pathologist's viewing station is a separate question from the server.

IVDR, GDPR and models built in-house

For pathology software, the regulation to check first is the one on in vitro diagnostics. Article 2(2) of Regulation (EU) 2017/746 (IVDR) defines an in vitro diagnostic medical device as, among other things, "software" intended by the manufacturer "for the examination of specimens, including blood and tissue donations, derived from the human body", to provide information "concerning a physiological or pathological process or state". A tissue section on a scanned slide is such a specimen.

Article 5(5) matters for institutes that build their own models. Except for the general safety and performance requirements of Annex I, the IVDR's requirements do not apply to devices "manufactured and used only within health institutions established in the Union" when all of its conditions are met: among them, the device is not transferred to another legal entity, manufacture and use run under an appropriate quality management system, the laboratory complies with EN ISO 15189 or national provisions, and the institution justifies that the patients' needs cannot be met by an equivalent device on the market. The institution also publishes a declaration and reviews clinical experience with the device, and the paragraph does not apply to devices manufactured on an industrial scale.

Slides and their reports are data concerning health, a special category under Article 9(1) of the GDPR, and so are the embeddings and logs derived from them while they can be linked to a patient. Under Article 6(1) of the AI Act, an AI system that is a product, or the safety component of a product, covered by the legislation in Annex I (the IVDR is listed there) is high-risk when that product needs a third-party conformity assessment; Article 113, as amended by Regulation (EU) 2026/1744, applies these rules from 2 August 2028; our medical imaging guide sets out the pending proposal to move the MDR and the IVDR within Annex I. Whether a model is a device, whether Article 5(5) applies and on which legal basis slides are processed are assessments for your regulatory affairs function and your data-protection officer. General information on EU law as of October 2026, from the texts on EUR-Lex; national rules can add conditions.

Storage and network for the slide archive

The slide archive grows faster than anything on the GPU server. Size it from your scanners' average file size times slides per day times the retention period your national rules set for slides; the embeddings add about 97 MB per slide in FP16 by our arithmetic, about 29 GB a day at 300 slides. The GPU server reads from the archive and keeps a working set on local NVMe, as our guide to NVMe storage for an LLM server explains for model files. A slide of several GB that moves to the server for each run makes the network link part of the throughput, so place the server close to the archive.

Notes for the IT team

1. Pipeline. Tile reading (OpenSlide or cuCIM), tissue detection, the encoder in FP16 or BF16 with a large batch, embeddings written to disk, task models reading the embeddings. Each step runs as its own container, so a new model replaces one step.
2. Separation. Keep research jobs apart from any model with a declared diagnostic purpose: a separate card or MIG instance, separate containers and documented versions, as the quality system asks.
3. Network. Scanners, archive and GPU server sit in the clinical network; the BMC belongs on the management network, and the hardening steps in our guide to securing a GPU server apply.
4. Software on delivery. On request the server arrives with the operating system, the NVIDIA driver, CUDA and a container runtime; cuCIM publishes packages for CUDA 12 and 13.
5. Acceptance. Run a pilot on 50 to 100 of your own slides and measure seconds per slide before you fix the card count; our guide to GPU server acceptance testing covers the test of the delivered system.

We set up the encoder pipeline, the task models and the logs on these servers through our Private AI/ML service; describe your scanners and your first use case, and we propose a pilot and the server for it.

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. We build GPU servers to order with two or four RTX PRO 6000 Server Edition cards, or with RTX PRO 4500 Server Edition cards, assembled and tested before delivery, delivered anywhere in the EU on one EU contract and invoice, with manufacturer warranty on every component handled through us; quantities and lead time are confirmed in writing with the quote. Our Private AI/ML service, with engineering by our partner Vixen.UNO, deploys the models and the pipeline. Regulatory assessment and validation stay with your institute. Our guide to buying a server with two RTX PRO 6000 Server Edition lists the entry configuration in detail.

FAQ