Somewhere between the third H100 announcement of the year and a cloud GPU bill with five zeros, a reasonable question gets lost: do structure-based virtual screening campaigns actually require a GPU at all? If your team runs docking workflows through HPSee, the answer is a clean no. And in a market where GPU infrastructure costs have become a serious line item in drug discovery budgets, that is worth talking about.
The GPU Price Reality in 2026
The past two years have normalized numbers that would have seemed extraordinary not long ago. A single NVIDIA H100 80 GB card lists between $25,000 and $40,000 depending on form factor and vendor. A production-ready 8-GPU server, once you factor in InfiniBand networking, power infrastructure, and cooling, lands between $200,000 and $400,000. Cloud rental is cheaper per unit of time, but at typical on-demand rates of $3 to $4 per GPU per hour, a continuously running 8-GPU cluster costs upward of $200,000 annually before any reserved-instance discounts. Maintenance contracts for enterprise GPU systems add another $20,000 to $50,000 per year.
None of this is inherently unreasonable if the workload demands it. Deep learning model training, large-scale generative chemistry, and multi-billion-parameter protein structure prediction genuinely benefit from GPU acceleration. But structure-based virtual screening with physics-based docking is a different computational problem, and conflating the two is an expensive mistake.
HPSee 2.3 “Electra” is the current release of BioSolveIT’s scalable virtual screening workflow environment. It orchestrates large-scale docking campaigns using FlexX for pose generation and HYDE for scoring, and scales across CPU cores with near-linear efficiency.
No GPU required. No specialized driver stack. No six-figure hardware procurement. Just cores, memory, and a clean admin dashboard.
Why Docking Is a CPU Problem
FlexX, the docking engine that powers HPSee’s virtual screening campaigns, is a fragment-based incremental construction algorithm. Pose generation proceeds by placing a base fragment in the binding site and extending the ligand through rotatable bonds, evaluated against a physics-informed scoring function. This is a combinatorial, tree-search process. It is embarrassingly parallel: each ligand docking job is entirely independent of every other, meaning throughput scales directly with the number of available CPU cores.
GPUs are optimized for dense matrix operations where thousands of threads share the same instruction stream on closely coupled data. Docking tasks look nothing like that. Distributing a library of 500,000 compounds across 128 CPU cores is computationally natural; forcing the same workload through a GPU shader pipeline is not. The result is that a well-provisioned CPU cluster running HPSee delivers throughput that is genuinely competitive for this class of problem, without the engineering complexity that GPU infrastructure demands.
- Independent tasks scale linearly: Each ligand docking run is self-contained. Adding CPU cores directly translates to higher throughput, with no coordination overhead between jobs.
- No CUDA dependency: HPSee runs natively on Linux (recommended) and Windows. There is no GPU driver stack to maintain, no CUDA version compatibility to manage, and no vendor lock-in to a single hardware ecosystem.
- Deploy on what you have: HPSee can be installed on a dedicated workstation, an in-house cluster, a cloud VM, or even a laptop. The minimum specification is 32 logical CPU cores, 2 GB of RAM per core, and 500 GB of disk space.
Hardware Cost Comparison
The numbers below represent realistic market ranges as of 2025/2026, combining published list prices, cloud provider rate cards, and publicly documented infrastructure cost estimates. They are intended to illustrate order-of-magnitude differences, not to substitute for a formal procurement analysis.
| Setup | Hardware (Purchase) | Cloud / Rental (Annual) | Annual Upkeep | Suitable for HPSee? |
|---|---|---|---|---|
| 1x NVIDIA H100 (PCIe) | $25,000 – $40,000 | $13,000 – $35,000 | Power, cooling, drivers | Not required |
| 8x H100 SXM Server (full cluster) | $200,000 – $400,000 | $200,000 – $280,000+ | $20,000 – $50,000 support contract | Not required |
| Dual AMD EPYC 9654 (192 cores, 512 GB RAM) HPSee team server, on-premise |
~$25,000 – $40,000 | N/A (owned) | Standard server maintenance | Yes — ideal |
| Single-socket AMD EPYC 9354P (32 cores, 64 GB RAM) HPSee minimum viable, workstation or small server |
~$5,000 – $8,000 | N/A (owned) | Standard server maintenance | Yes — entry point |
| CPU cloud instance (e.g. 64-core, 256 GB RAM VM) HPSee on cloud CPU |
No purchase needed | ~$3,000 – $8,000 | Pay-as-you-go | Yes — flexible |
GPU pricing sourced from published market data (2025/2026): H100 single-card MSRP $25,000–$40,000; 8-GPU server total cost of ownership $200,000–$400,000; cloud rental rates $3–$4/GPU/hr on-demand. CPU server pricing based on current retail configurations. Annual rental figures assume sustained 24/7 usage.
What HPSee 2.3 “Electra” Brings to Your Workflow
The latest release of HPSee focuses on making large-scale campaigns faster to set up, easier to monitor, and more resilient in production. For teams running regular docking campaigns or Chemical Space Docking runs, several improvements translate directly to less time spent managing infrastructure and more time spent interpreting results.
- Prefiltering for spaces and libraries: HPSee 2.3 can reduce data volume before computationally demanding steps such as Chemical Space Docking, cutting down both processing time and resource load. This is particularly relevant for large Chemical Spaces where filtering on generic properties upstream meaningfully narrows the search.
- Transparent database handling: Database size is now displayed directly in the Dashboard, and upload progress is communicated through persistent messages and a navigation lock that prevents accidental cancellation mid-upload. No more guessing whether a large space upload completed successfully.
- First-in, first-out workflow execution: Submitted tasks now process in the order they were received, making it straightforward for teams with multiple users to predict when their jobs will run. Sortable admin tables and editable metadata for spaces and workflows round out the administrative improvements.
- Cloud workflow stability: For teams running HPSee on cloud infrastructure, the 2.3 update adds retry logic after crashes and automatic handling of spot instance shutdown events, reducing the risk of losing progress on long-running campaigns.
- Live license renewal: Licenses can now be renewed directly from the admin dashboard without restarting HPSee, removing a friction point for teams in the middle of active screening campaigns.
HPSee and Chemical Space Docking: Scaling Without GPU Costs
Perhaps the most computationally demanding workflow HPSee supports is Chemical Space Docking (C-S-D), BioSolveIT’s iterative fragment-based method for screening ultra-large Chemical Spaces containing billions to trillions of virtual compounds. Even at that scale, the underlying computation remains CPU-native.
C-S-D works by first anchoring synthon fragments in the binding site, then iteratively extending the growing molecule using FlexX’s incremental construction. Because each extension step evaluates a manageable set of fragment combinations rather than enumerating all possible full molecules explicitly, the approach is computationally tractable on CPU hardware. HPSee manages the orchestration, queuing, and result preparation behind the scenes, while SeeSAR handles the visual front end on the researcher’s workstation. The result is access to chemical matter at a scale that was unthinkable a decade ago, running on hardware that fits within a normal IT procurement budget.
HPSee 2.3 “Electra” is available now. It works with SeeSAR 15 and supports both regular docking campaigns and Chemical Space Docking on CPU-based hardware, from a single workstation to a multi-node cluster.