Three drug discovery jobs, three different GPU bottlenecks — memory, none at all, and raw parallelism. The interesting part is how little they have in common.
Last November I had time on RTX 6000 Blackwell cards — 96 GB each — at a Swiss data centre, and put together some use cases for a talk at DINAcon. Nine months on, the models have moved and it is worth writing down what the memory ceiling really meant. Three jobs, three different bottlenecks.
1. A protein that does not fit
Most proteins are small. A kinase domain is around 300 residues; a GPCR about the same. Then there is TRRAP: 3859 amino acids, a PIKK-family scaffolding protein, recurrently mutated in cancer.
Cofolding models scale quadratically in sequence length, because the pair representation does. I measured this on Boltz-2 in 100-residue steps and fitted the curve:
My largest successful Boltz-2 run was 2700 residues. The fit puts full-length TRRAP at roughly 215 GB — more than twice what a 96 GB card holds. Bigger hardware alone was not going to close that gap.
What closed it was OpenFold3’s low-memory mode. In default settings OpenFold3 will not take 3859 residues on 96 GB either; in low-mem mode the full chain came back in about sixteen minutes.
That is the useful lesson, and it is not the one I expected going in. The memory wall is real, but you attack it from two sides. Buying a card with more VRAM moves it a little — quadratically little. An implementation that trades compute for memory moves it a lot. Check what the code can do before pricing the hardware.
2. Mutants, in minutes
Osimertinib – and Poziotinib, below – is a covalent inhibitor for EGFR-mutant lung cancer, developed to survive the T790M gatekeeper mutation. Resistance keeps evolving, so the recurring question in any kinase program is what the site looks like once a new variant appears.
My test was whether a model reproduces a pose we already know. Below, Boltz-2 models of the double mutant T790M/L858R, are overlaid on the corresponding crystal structures.
Both drugs reproduce the crystallographic binding mode perfectly – rmsd values were below 1 Angstrom. A mutant–compound pair takes minutes on one GPU. So you no longer pick the two or three variants worth modeling — you run the whole clinical panel against your series.
This case needs no special hardware. A 24 GB consumer card is enough, and a Mac with enough unified memory will get there too if you are patient. Data centers earn their keep on the jobs in sections 1 and 3 — not on this one.
3. Design: a throughput problem
The measured number: 2000 protein designs generated and ranked in just under 4 hours on 4 GPUs. The Boltzgen authors suggest sampling around 60,000 for a serious campaign — call it a long weekend on 8 cards, assuming throughput scales cleanly, which I verified in a different project at that scale.
Memory never came close to be an issue here. A nanobody plus a GPCR is a small system. What this job wants is cards, plural, kept busy.
None of these designs have seen a wet lab. A ranked list from a generative model is the start of a project, not a result — the point is that the list now costs a weekend of GPU time rather than months of immunization and panning.
Since November 2025
OpenFold3 left preview in March 2026: full stack, weights, training and inference code, evaluation scripts and the training data itself, permissively licensed. For consulting that matters more than a benchmark: I can run it inside a client’s firewall, tell them exactly what the model was trained on, and know it will still be there, unchanged, when the project ends.
BoltzGen has meanwhile accumulated a broad body of validation across target classes and binder modalities. Regrettably it did not deliver when I used this for peptide design in a live project.
The practical version
- Large single structures are a memory problem — but check for a low-memory implementation before you check the price of a bigger card.
- Routine cofolding and mutant modelling is not a hardware problem at all. One 24 GB card under a desk keeps up with a medicinal chemistry program.
- Design campaigns are a throughput problem: several cards you can saturate, and a pipeline that survives one of them failing.
A pair of RTX 3090s covers most of what clients ask me for. The value of occasional 96 GB access is not that everyday work gets faster — it is that a category of question I used to decline becomes answerable.
I run CADD Consulting GmbH in Zofingen, Switzerland — structure-based design, free energy calculations and AI structure prediction for biotech, pharma and academic groups. All structures shown are from non-confidential internal work. Thanks to stepping stone AG for GPU access.
Models: Boltz-2 (repo, MIT) · BoltzGen (repo, MIT) · OpenFold3 (repo, Apache 2.0)