ComputeCosts observatory report 006

Abstract

This study evaluates a current scientific-compute workstation built around a refurbished DellPrecision T7610 with 512 GB ECC system memory and two NVIDIA GPUs with separate roles: anRTX PRO 4500 Blackwell with 32 GB VRAM for scientific compute and a Quadro RTX 8000 with 48GB VRAM for local LLM inference, programming assistance and retrieval-augmented generation. Ittests whether this architecture is a practical alternative for memory-intensive scientific computingwhile new high-capacity memory and professional workstations remain expensive, difficult tosource and still increasing in price. The study measures actual system performance, the cost of theslower host platform and what happens when scientific working sets exceed the physical memoryof a modern 128 GB-class workstation. The economic pressure is not only price. High-capacity DDR5 remains expensive and supply istight, affecting both the availability and cost of high-memory components and complete professionalplatforms. TrendForce reported an undersupplied server-DRAM market in Q3 2026, forecastanother 13–18 percent quarter-over-quarter increase in contract prices and stated that quotationscould continue to rise during the quarter. Its 4 September 2026 analysis still described tighteningsupply and rising contract prices [1, 2]. Current industry forecasts indicate that DRAM supply willremain tight through 2027, with meaningful new capacity expected mainly from the second half of2027 and substantial additional output not expected until 2028 [10, 11]. The observed 512 GBDDR5 ECC options in this report cost approximately EUR 19,700 to EUR 20,700 excluding VAT.The observed 512 GB DDR3 ECC capacity cost approximately EUR 1,669 excluding VAT. A currentprofessional 512 GB compute configuration captured for comparison cost EUR 58,823 excludingVAT [3, 4, 5, 7]. The intended workloads are scientific machine learning, bioinformatics and other researchworkloads in which large datasets or scientific state must remain directly accessible in memory.The workstation is intended as a complete local research system rather than bare hardware, withcurrent Debian, CUDA, PyTorch and open-source scientific software stacks installed, tested andready to use. The two-GPU layout keeps the RTX PRO 4500 available for scientific computation1while the RTX 8000 runs a local language model or RAG service alongside it. To evaluate theplatform under these conditions, this study uses FrankenBench v3.1, an in-house developed,fit-for-purpose benchmarking tool that measures CPU throughput, memory performance, GPUfeeding, large file-backed datasets, full RAM residency and large scattered-memory workloads.The T7610 performed well. It sustained 585.9 GFLOP/s in the CPU matrix test, 18.44 GB/sconcurrent NUMA memory-copy throughput and 12.98 GB/s pinned host-to-GPU transfer. The RTXPRO 4500 reached 100 percent sampled utilisation in GPU-resident neural-network training.Continuous RAM-fed training reached up to 78 percent of the GPU-resident rate, while the tunedfixed-batch sweep reached 94.7 percent at batch 1024. The machine kept a 450 GiB file-backeddataset resident across the second pass and completed the three-iteration scattered-memoryworkload in 7.1 minutes at 300 GiB and 10.3 minutes at 400 GiB. The T7610 is slower than a current high-end desktop on small in-memory workloads. The Ryzen 9reference system measured approximately 4.0 times more CPU matrix throughput and 4.6 timesmore simple RAM-copy bandwidth. Those differences matter for CPU-bound work, but they areseveral-fold rather than order-of-magnitude differences. The T7610 still provides substantial hostcompute performance and, based on the author’s experience, is more than adequate for mostchemical-engineering machine-learning, process-monitoring and control, and sensor-fusionresearch workloads. For GPU-driven scientific computing, the host-side difference matters lessbecause the measured system already feeds the RTX PRO 4500 effectively. Once the workload exceeds the physical memory of the modern reference system, the differencebecomes much larger. At 300 GiB the T7610 completed all three random-access iterations in 7.1minutes; the 128 GB-class modern workstation ran for four hours without completing its firstiteration. At 400 GiB the T7610 completed all three iterations in 10.3 minutes; the modernworkstation again ran for four hours and completed only about 42 percent of its first iteration. Linearextrapolation of the measured record-processing rate gives estimated full-job times ofapproximately 15.9 hours at 300 GiB and 27.8 hours at 400 GiB, versus measured T7610 times of7.1 and 10.3 minutes. These extrapolated values are estimates, not benchmark measurements.The measured result is sufficient: for the tested large scattered-memory workload, enough physicalmemory changes execution time from many hours to minutes.