CAD Laboratory Refresh at Valley College
Valley College's CAD laboratory has 60 desktop PCs, each five years old with 8 GB RAM, a 1 TB hard disk and integrated graphics. The new engineering course uses 3D assemblies of about 1.5 GB and renders that take 20 minutes on the current machines. The college must support 60 students, with a peak of 45 working concurrently. Option A replaces each PC with a fat client: 32 GB RAM, 1 TB NVMe SSD, a discrete GPU, at $1,800 per seat. Option B uses thin clients at $350 each plus one shared server costing $45,000, with two data-centre GPUs, 256 GB RAM, 10 TB NVMe storage and a 10 GbE network card. The server room has power and cooling capacity for either option. The network upgrade to 10 GbE is already funded. The college keeps equipment for four years and values predictable lesson starts.
1. Calculate the four-year hardware cost of each option and identify two non-cost factors that could change the decision.
Read the worked answer
Option A, fat clients: 60 × $1,800 = $108,000. Option B, thin clients plus server: 60 × $350 = $21,000; $21,000 + $45,000 = $66,000. The apparent saving is $42,000, about 39% of the fat-client cost.
Non-cost factors:
- Single point of failure and contention. One server serves all 45 concurrent users. If it fails, the whole laboratory stops; if two GPUs are shared by 45 users, each GPU supports about 22.5 interactive sessions. Render jobs may queue, so lesson starts are less predictable than with local GPUs.
- Network dependency and latency. Thin clients depend on the 10 GbE link and switching. A 1.5 GB assembly can be loaded over the network, but interactive rotation and rendering add latency. The fat client keeps the model local.
A fair comparison must also include server redundancy, UPS, support and four-year warranty. The $42,000 gap may shrink if a second server is needed for resilience. The college should not choose on purchase price alone.
2. Explain how RAM, SSD and GPU each affect the 20-minute render, and identify the most likely bottleneck on the current PCs.
Read the worked answer
RAM holds the operating system, CAD application and the 1.5 GB assembly. With only 8 GB, the model plus application can exceed available memory, so the OS pages data to the hard disk. A 1 TB hard disk has high seek latency, so paging can dominate. The SSD reduces load time and makes paging far less painful, but it does not perform the render calculation.
The GPU accelerates viewport display and, for many CAD packages, GPU rendering. Integrated graphics share system RAM and have far less parallel compute than a discrete GPU, so complex shading and ray tracing are slow. The CPU may still be the renderer if the software uses CPU rendering; in that case the GPU helps interaction but not the final image. The 20 minutes could therefore be CPU-bound, GPU-bound or I/O-bound depending on the renderer.
Most likely on the current PCs: memory pressure causing paging to the hard disk during load and rendering, compounded by weak integrated graphics. The correct diagnosis is to measure memory use, disk queue length and GPU utilisation during a real render. Adding RAM and an SSD may cut load time dramatically but may leave a GPU-bound render unchanged.
3. Recommend an option and describe how you would validate it before purchase. Your answer should account for the peak of 45 concurrent users.
Read the worked answer
Recommendation: a hybrid is safer than either pure option. Use fat clients for the 45 peak concurrent CAD seats, or at least for the most render-heavy seats, because local GPUs avoid GPU contention and network latency. Thin clients could serve overflow, light editing and general access. If the college insists on one option, the thin-client server needs a second GPU node or a second server to keep 45 concurrent users predictable; that changes the cost comparison.
Validation before purchase:
- Build a pilot with 10 fat clients and a thin-client server configured for 45 simulated sessions.
- Benchmark the actual 1.5 GB assembly: cold load time, viewport frame rate at a stated resolution, and final render time.
- Run 45 concurrent sessions, not 10, and record GPU queue time, server CPU, RAM and 10 GbE utilisation.
- Test a lesson-start scenario where all 45 students open the same assembly within two minutes.
- Compare measured results with the current 20-minute baseline and with the college’s acceptable lesson-start delay.
Choose the option that meets the measured requirement with redundancy, not the one with the lowest sticker price.