Throughput Isn't One Problem: A Pareto-Limited View of Reconstruction Bottlenecks

The Problem Isn't Always Where You'd Expect

In equipment throughput optimization work generally, the instinct is often that throughput is primarily a mechanical and motion-control problem — faster stages, tighter tolerances, less time wasted moving between positions. That instinct isn't wrong, but it's frequently incomplete. Motion control and detector acquisition speed are real contributors, but as those improve, a different bottleneck often becomes the limiting factor: data reconstruction.

That's a common pattern in throughput optimization generally, and it's worth naming explicitly: improving one stage of a pipeline doesn't improve overall throughput once a different stage becomes the binding constraint. You don't find out reconstruction is the bottleneck until you've already fixed the things that were bottlenecks before it.

Why Reconstruction Is a Genuinely Parallel Problem

The reason GPUs matter so much for this specific bottleneck is that image and tomographic reconstruction is unusually well-suited to parallel computation. Published work on multi-GPU CT and tomographic reconstruction consistently shows large speedups from adding GPU resources — one study using three GPUs for iterative statistical reconstruction reported roughly a 72x speedup over a CPU baseline, and an analytical FDK-based reconstruction approach reported speedups over 1000x using multiple GPUs versus CPU. Other multi-GPU ptychographic and tomographic reconstruction work has reported superlinear speedups when moving from fewer to more GPUs, specifically because splitting a large dataset across more GPUs lets each one fit its share of the problem into fast on-board memory, avoiding slower memory-swapping penalties that hurt performance on a single, memory-constrained GPU.

In practice, this means adding a second GPU to a reconstruction pipeline can come close to halving reconstruction time — not always exactly linear, since real-world scaling depends on problem size, memory constraints, and communication overhead between GPUs, but close enough that it's a legitimate, high-leverage lever for any throughput optimization effort to pull.

The Lever That Got More Expensive While We Were Reaching for It

Here's where the story stops being purely technical. Adding GPUs is the obvious fix for a reconstruction bottleneck — but 2026 turned out to be one of the worst possible years to be buying GPUs and the DRAM that feeds them.

The underlying cause is a genuine, structural shift in the memory market, not a temporary blip. AI data centers are projected to consume something like 70% of the world's memory output in 2026, up from roughly 20-30% in 2022, as the three dominant DRAM manufacturers — Samsung, SK Hynix, and Micron — redirect wafer capacity toward the high-bandwidth memory that AI accelerators need. That memory carries far richer margins than conventional GPU or system DRAM, so the reallocation isn't a supply hiccup; it's where the money is. TrendForce reported DRAM contract prices rising somewhere in the range of 50-90% quarter-over-quarter at points during 2026, and GPU vendors have passed a meaningful share of that cost increase through to buyers. Industry forecasts generally don't expect meaningful relief before 2027 or 2028, when new fabrication capacity currently under construction actually comes online.

That's the environment any throughput optimization effort is operating in right now: the fix for a real engineering bottleneck (add GPUs) runs directly into a real macroeconomic bottleneck (GPU and memory costs climbing sharply, driven by demand that has nothing to do with the problem at hand).

Why This Is a Pareto Problem, Not a Solvable One

Once cost and throughput are both real constraints, you're no longer looking for "the" answer — you're looking at a Pareto front. In multi-objective optimization, a solution is Pareto-optimal if you can't improve one objective (throughput) without making another objective (cost) worse. The set of all such solutions forms a curve, and the honest job of an engineering team isn't to find a point that maximizes both simultaneously — that point usually doesn't exist — it's to characterize the tradeoff curve clearly enough that leadership can make an informed decision about where on it to sit.

For a team optimizing equipment throughput, that reframes the whole conversation. Instead of "how do we make reconstruction faster," the real question becomes: how much additional throughput is a given GPU budget actually worth, given current, inflated hardware pricing, and is that the best use of that budget relative to the other levers still available (motion control refinement, detector timing, algorithmic improvements to the reconstruction pipeline itself that reduce compute requirements without adding hardware)?

The Practical Takeaway

Throughput problems in complex equipment systems are rarely single-variable, and the "obvious" fix for whichever bottleneck is currently binding may not be the best use of budget once you account for what that fix actually costs right now. Right now, GPU-based reconstruction acceleration is a genuinely powerful lever — the technical case for it is strong and well-documented — but it's an unusually expensive lever to pull in 2026 specifically, because of a memory market shift that has nothing to do with any one team's reconstruction pipeline and everything to do with global AI infrastructure demand. Leading a throughput effort well means being honest about that tradeoff rather than treating more GPUs as a free win.

Rob Rainer is Director of Controls & Electrical Engineering, with experience leading cross-disciplinary throughput optimization efforts spanning systems engineering, electrical, motion control, and data reconstruction in semiconductor capital equipment environments. He previously spent over 15 years in controls and accelerator operations at Brookhaven National Laboratory's NSLS-II.

← Back to Articles