Why is FAH still assigning the same work to CPUs that an RTX 5090 finishes 40x faster?

[Ncard00]

I’ve been looking at the PPD (Points Per Day) stats for the 2026 hardware lineup, and the gap has become absurd. An RTX 5090 is pulling 100 million PPD, while a high-end CPU struggles to hit 2 or 3 million. Even though the 5090 uses double the watts, it's still 20 to 40 times more efficient per unit of science.

Why are we still allowing this overlap in the scheduler? It feels like we’re subsidizing massive energy waste just to keep "inclusive" support for hardware that should have been retired years ago. Every work unit that sits on a CPU for days is a unit that a modern GPU could have cleared in minutes.

How can we justify this "carbon tax" on the project in 2026? Are we afraid to set an efficiency floor because it might offend people with older rigs, or is there a genuine scientific reason to keep wasting cycles like this? I’m curious how the rest of you balance the "volunteer spirit" with the reality that a single modern GPU does more for the science than a room full of legacy CPUs ever will.

[muziqaz]

Not you again.
CPU projects are there for a reason.
GPUs cannot do them. Plain and simple.
Any workloads which can be run on the GPUs are being run on GPUs.
While GPUs are super fast, they cannot do certain simulations.

[Ncard00]

Not you again.
CPU projects are there for a reason.
GPUs cannot do them. Plain and simple.
Any workloads which can be run on the GPUs are being run on GPUs.
While GPUs are super fast, they cannot do certain simulations.

I'm well aware that certain simulations historically needed the lower latency of a CPU, but it’s 2026. The gap between what a modern GPU (or even a high-end NPU with unified memory) can handle and what a legacy CPU can do is closing fast.

If these "certain simulations" are so specific that they can only run on an architecture that is 40x less efficient, then we need to ask why the software isn't being prioritized for modernization. It feels like a convenient excuse to keep the "inclusive" legacy support going while ignoring the massive carbon footprint of these long-running CPU tasks.

What is the long-term plan here? Are we just going to keep running these specific simulations on "space heaters" forever, or is there an actual push to move these workloads onto more energy-proportional silicon? At some point, the "special simulation" argument stops being a scientific necessity and starts looking like a refusal to adapt to modern hardware reality.

[muziqaz]

It is not a latency. It is the simulation type. GPUs are unable to do that. End of conversation. Your ideas and your assumptions are just that, and have no scientific backing.

[toTOW]

An RTX 5090 is pulling 100 million PPD

No, it's in the 50-60 millions range on Linux with favourable projects.

[appepi]

we need to ask why the software isn't being prioritized for modernization.

....

Does this thread really needed this type of comment?

[muziqaz]

we need to ask why the software isn't being prioritized for modernization.

Yes indeed, so I suggested to Google that: "we need to ask why the folding software isn't being prioritized for modernization". And google asked its AI of course, which (inter alia) found this thread as its first reference, and concluded: "Based on common industry hurdles in 2026, the modernization of specialized, legacy systems like folding software is often de-prioritized due to a combination of high-stakes risk, perceived stability, and competing strategic investments." Reading further down, it seemed that the AI might have been unclear that it was protein folding we were asking about, so I tried: "we need to ask why the protein folding software isn't being prioritized for modernization". This time it said:

While AI tools like AlphaFold have revolutionized static structure prediction, the modernization of broader, dynamic, and physics-based protein folding software faces significant hurdles, including high computational costs, challenges in modeling folding pathways, and funding uncertainties in 2026.

Then I went to https://alphafoldserver.com/about and https://deepmind.google/blog/stopping-m ... ts-tracks/ and https://www.nature.com/articles/s41586-019-1923-7 but my head started to hurt at the thought of 200 GPUs versus my antique workstations and decided to keep doing something that isn't owned by an oligarch.

Does this thread really needed this type of comment?