GPU Prices Too High? Try HPSee!

Other

GPU Prices Too High? Try HPSee!

August 21, 2026 10:20 CEST

Structure-based virtual screening of a billion-compound library used to be a question of how much hardware you could buy. Chemical Space Docking® changed the arithmetic: instead of enumerating and docking every product, it docks the reagents that span the space. Cost then scales with the size of the synthon set, not the size of the catalogue. That is why HPSee runs ultra-large campaigns on CPU hardware, and why no GPU is required.

Peer-Reviewed Result
almost 1 billion
commercially available
compounds screened
27 of 69
purchased compounds with
Ki below 10 µM, a 39% rate
2
X-ray structures of leads
confirming their docked poses
ROCK1 kinase campaign with Genentech, published in Nature Communications 3.

Why Chemical Space Docking® Changes the Cost Curve

Conventional virtual screening enumerates a library, then docks each molecule. Compute cost therefore tracks the number of products. At a billion compounds and beyond, that relationship is what forces teams into large infrastructure purchases.

Chemical Space Docking® breaks that relationship. The method anchors synthon fragments in the binding site first, then iteratively extends the surviving anchors along the reactions that define the Space, using FlexX 1 incremental construction and HYDE 2 scoring. Because the expensive work happens at the level of the reagents that span the Space rather than the products it contains, the approach scales roughly with the number of reagents. Beroza and co-workers report that this makes the approach multiple orders of magnitude faster than traditional docking 3.

An algorithmic saving of that size is worth more than any hardware upgrade. It is also the reason a screening campaign against billions of virtual compounds fits comfortably on a CPU server.

Why the Workload Suits CPUs

FlexX pose generation is a fragment-based incremental construction algorithm. It places a base fragment in the binding site and extends the ligand through rotatable bonds, pruning branches as it goes. The result is a branch-heavy tree search with irregular control flow and data-dependent memory access. Workloads of that shape map poorly onto the SIMT execution model that GPU accelerators are built around, which rewards thousands of threads executing the same instruction on regularly structured data.

What the workload does suit is horizontal scaling. Individual docking jobs are independent of one another, so throughput within each stage of a campaign rises directly with the number of available cores. HPSee distributes those jobs across every core on the machine and queues them without any manual chunking of libraries. Because there is no coordination overhead between jobs, adding cores shortens campaigns close to proportionally, and HPSee can be deployed as a swarm across several machines when a single server is not enough.

  • Independent jobs, horizontal scaling: Docking runs within a stage are self-contained. Adding cores translates directly into throughput, with no coordination overhead between jobs.
  • No CUDA dependency: There is no GPU driver stack to maintain, no CUDA version compatibility to track, and no lock-in to a single accelerator vendor.
  • Deploy on what you have: HPSee is delivered as a containerized REST API running under Docker, so the same deployment works on a dedicated workstation, an on-premise server, a VM, or a cloud instance. Linux is supported and recommended; Windows 10 and later is supported via WSL2 and Docker Desktop. macOS is not currently supported for the HPSee server.
HPSee Server: Where to Start

An entry configuration that runs Chemical Space Docking® end to end. Larger configurations shorten the turnaround accordingly.

32
logical CPU cores
the practical entry point
2 GB
RAM per logical core
64 GB at 32 cores, scaling with the core count
500 GB
disk space
at least 10 GB per C-S-D run

The SeeSAR client needs 16 GB RAM to work with C-S-D results and runs on Linux, macOS, and Windows.

These figures describe what is needed to run a campaign from start to finish. Because the workload scales horizontally, the core count is the one variable that decides how long that campaign takes. The result of the screen does not change with the hardware. What changes is how quickly it arrives, and that in turn decides how a project team can work.

Scalable performance

The same Chemical Space Docking® run, on three different core counts.

Results the following week
5 to 6 days
on 32 cores
Set the campaign running, return to it after other work.
Results the next morning
1 to 2 days
on 128 cores
Adjust the binding site or the filters and screen again the same week.
Results the same day
1 to 2 hours
on 1024 cores
Screen, review the poses, and refine the setup in a single session.

Bar length shows the relative time to completion. Figures depend on the Space, the target, and the filters applied.

The step from a weekly rhythm to a same-day one is what changes the method from a set-piece exercise into something a project team uses routinely. For scale, the ROCK1 screen was executed on nine computing units of 32 cores and 64 GB RAM each, 288 cores in total 3. Core count is also not the only lever. The prefiltering introduced in HPSee 2.3 reduces the volume of data entering a workflow before the expensive steps begin, which shortens the same run on the same hardware.

What That Looks Like in Practice

Four configurations cover most teams, each matching one of the rhythms above. Figures below are indicative market ranges for complete systems in 2026 and are meant to show orders of magnitude, not to replace a procurement quote.

Configuration Indicative Cost Best For
Single-socket 32-core server
e.g. AMD EPYC 9355P, 64 GB RAM
~$5,000 – $8,000
purchase
A single group getting started with remote docking and C-S-D. Roughly 5 to 6 days per run, so campaigns are planned by the week.
Single-socket 128 to 192-core server
e.g. AMD EPYC 9965, 256 to 512 GB RAM
~$25,000 – $45,000
purchase
Shared team resource running campaigns continuously. Overnight turnaround makes iteration on binding site setup and filters practical.
Cloud CPU instance
64 vCPU, 128 GB RAM
~$3.30/hr
on demand, roughly a quarter of that on spot
Campaign-driven or seasonal use. HPSee 2.3 handles spot instance shutdown events, which makes spot pricing practical.
Multi-node swarm or cloud burst
around 1024 cores across several machines
~$50 – $55/hr
on demand, for the hours the run takes
Time-critical or high-throughput work. A run that finishes in 1 to 2 hours puts C-S-D inside a working day, for roughly $100 of compute.

The last configuration is the one that changes how a campaign can be planned. Screening a Chemical Space of billions of compounds within a working day, on rented CPU capacity, costs roughly the same as a set of assay plates.

For comparison, a single current-generation accelerator card still lists in the tens of thousands of dollars, and a full eight-GPU node runs into the low hundreds of thousands before power, cooling, and support contracts. Those budgets are justified for model training and generative chemistry. They are not a prerequisite for physics-based screening.

Note: CPU server figures based on current retail configurations for complete systems. Cloud figure based on published on-demand rates for 64-vCPU compute-optimized instances (August 2026); the burst figure scales that rate to roughly 1024 vCPU. Turnaround times are indicative and depend on the Space, the target, and the filters applied. Actual cloud spend depends on utilization; a campaign that runs for two weeks costs a fraction of a continuously provisioned instance.

Free access to SeeSAR and Chemical Space Docking, celebrating 25 years of BioSolveIT
25 Years of BioSolveIT
Free Access to SeeSAR and Chemical Space Docking®

From September 1 to November 15, 2026, SeeSAR and HPSee for Chemical Space Docking® are available free of charge. The license covers all Modes and Modules, including SeeSAR’s third-party applications, and is open to users from academia and industry alike.

There is no commitment and no cost. The one thing worth arranging in advance is the compute described above, which is exactly why we announced the promotion early. Teams that secure more cores before September will get more campaigns done inside the 2.5 months.

Get a Reminder

How SeeSAR, HPSee, and Chemical Space Docking® Fit Together

Desktop Client
SeeSAR
Set up campaigns, define binding sites, monitor jobs, and visualize poses and HYDE scores on your workstation.
Server Engine
HPSee
Containerized REST API that hosts the Spaces, queues the jobs, and executes them across all available CPU cores.
Method
Chemical Space Docking®
Synthon-anchored screening of billions of virtual compounds, tractable on CPU.

SeeSAR is the desktop client that connects to HPSee’s REST API. Researchers submit campaigns, define binding site parameters, monitor job progress, and assess poses and HYDE scores locally. The compute stays on the server, the interaction stays on the desktop. The two products are licensed separately.

The anchoring phase can be guided by co-crystallized ligands or predicted binding poses used as templates, which speeds up the calculation and produces poses aligned with the template molecule. HPSee also supports standard remote docking and HYDE scoring independently of Chemical Space Docking®: teams can dock conventional enumerated libraries and run batch scoring jobs through the same API.

What HPSee 2.3 “Electra” Adds

Released 31 March 2026, HPSee 2.3 focuses on 3D Chemical Space processing: faster workflows, more transparent data handling, and a smoother cloud experience.

  • Prefiltering of Spaces and libraries: Chemical Spaces and libraries can now be prefiltered before they enter a workflow, reducing both the data processed and the data produced. Computationally demanding steps such as those in Chemical Space Docking® benefit most. Database size is now visible directly in the Dashboard.
  • More convenient generic filters: Separate filters for molecules and fragments can be combined within a single request, filters can be connected through OR logic, and annotated properties of libraries and Spaces are reachable more quickly.
  • More transparent Space uploads: Dashboard messages no longer fade automatically, upload windows explain the preprocessing applied during upload, and a navigation lock warns before an open tab is closed mid-upload. Chemical Space deletion now runs asynchronously, so users get immediate feedback while deletion continues in the background.
  • Improved cloud workflows: Workflows that crash can be retried, and HPSee reacts automatically to shutdown events from spot instances. This is what makes spot pricing a realistic option for long campaigns.

Ready to Scale Your Screening Without Scaling Your Hardware?

HPSee 2.3 “Electra” works with SeeSAR 15 “Apollo” and supports both standard remote docking and Chemical Space Docking® on x86 CPU hardware, from a single workstation to a multi-node cluster.

Download HPSee 2.3
  
Download SeeSAR 15
  
Learn More

References
[1] Rarey M, Kramer B, Lengauer T, Klebe G. A fast flexible docking method using an incremental construction algorithm. J. Mol. Biol. 1996, 261, 470–489. doi:10.1006/jmbi.1996.0477
[2] Schneider N, Lange G, Hindle S, Klein R, Rarey M. A consistent description of HYdrogen bond and DEhydration energies in protein-ligand complexes: methods behind the HYDE scoring function. J. Comput. Aided Mol. Des. 2013, 27, 15–29. doi:10.1007/s10822-012-9626-2
[3] Beroza P, Crawford JJ, Ganichkin O, Gendelev L, Harris SF, Klein R, Miu A, Steinbacher S, Klingler FM, Lemmen C. Chemical space docking enables large-scale structure-based virtual screening to discover ROCK1 kinase inhibitors. Nat. Commun. 2022, 13, 6447. doi:10.1038/s41467-022-33981-8