Workstation Strategy for AutoCAM and CAD/ML Development¶

Home / Infrastructure / Workstation Strategy for AutoCAM and CAD/ML Development

Final configuration ordered¶

Based on the analysis below, the first workstation was ordered from Puget Systems as a 2U rackmount configuration.

Component	Selection
Platform	Puget Rackstation Ryzen X870E R110-2U
Motherboard	ASUS ProArt X870E-Creator WiFi
CPU	AMD Ryzen 9 9950X — 4.3 GHz base, 16 cores, 170 W
RAM	64 GB DDR5 (2×32 GB)
GPU	NVIDIA GeForce RTX 5090 32 GB
Sound	Onboard
Networking	Integrated Ethernet, WiFi, and Bluetooth
Storage	2 TB NVMe PCIe Gen5 M.2 SSD (primary drive)
Case	InWin IW-RL200 Liquid Cooled 1200 W 2U Rackmount
PSU	Included with system
CPU Cooling	Included with system (liquid)
OS	Ubuntu 24.04 LTS Server Edition (64-bit)
Warranty	Lifetime labor and tech support, 3-year parts

Subtotal: $9,192.30

1. Goal¶

We need a Linux workstation strategy that works well now and scales cleanly over the next 12 months. Near term, the machines should feel fast for interactive AutoCAM development, support GPU-based simulation and CAD/ML research, and stay responsive when background jobs are running. Longer term, we should be able to add more machines incrementally and later introduce a larger shared workhorse without throwing away the initial investment. (NVIDIA)

2. Requirements and desired experience¶

Primary workloads¶

The main workloads are:

AutoCAM development on ModuleWorks
GPU-accelerated simulation
AI/ML research on B-Rep or topological CAD data
Light shared Linux server / CI use

ModuleWorks publicly positions GPU-accelerated simulation as a major speedup path for simulation workloads, so GPU capability matters for both ML and verification. (camelcamelcamel.com)

Desired user experience¶

The machine should feel responsive in normal engineering use: IDE, shell, local builds, kernels, tests, and simulation should not feel sluggish. It should stay stable on Ubuntu 24.04, support CUDA cleanly, and be easy to upgrade later without redesigning the whole machine. AMD lists Ubuntu x86_64 support for the Ryzen 9 9950X, and NVIDIA provides current Ubuntu/CUDA installation guidance for Linux. (Pangoly)

3. Hardware priorities¶

CPU¶

The best default CPU for this use case is the AMD Ryzen 9 9950X. It gives 16 cores / 32 threads and up to 5.7 GHz boost, which is a good fit for work that needs both strong single-core responsiveness and enough headroom for background jobs or a second user. AMD also lists 24 usable PCIe lanes on this class of desktop system. (Pangoly)

The main Intel alternative is the Core Ultra 9 285K. It also boosts to 5.7 GHz, but it has 24 total threads rather than 32, and Intel lists 125 W base power and 250 W maximum turbo power. It is worth knowing about, but it is not the default recommendation here. (Amazon)

GPU¶

For this workload, 16 GB VRAM is the floor and 32 GB is the preferred target for the ML-heavy machine. NVIDIA’s RTX 5080 is listed at 16 GB GDDR7 and “starting at $999,” while the RTX 5090 is listed at 32 GB GDDR7 and “starting at $1,999.” That makes the 5080 viable for today and the 5090 much more comfortable for larger CAD/ML experiments and future growth. (NVIDIA)

Practical framing:

RTX 5080: acceptable starting point.
RTX 5090: better if ML and simulation are likely to become a major daily bottleneck soon. Street pricing in early 2026 has often been above MSRP, especially for the 5090. (Best Value GPU)

RAM¶

A strong starting point is 96 GB, and on AM5 the preferred way to get there is 2×48 GB. AMD lists the Ryzen 9 9950X as a 2-channel DDR5 platform and shows a meaningful difference in official supported memory speed between 2 DIMMs and 4 DIMMs: DDR5-5600 with 2 DIMMs versus DDR5-3600 with 4 DIMMs. That is the clearest reason to prefer 2×48 over 4×32 where possible. (Pangoly)

Practical implication:

2×48 GB: preferred.
4×32 GB: usable, but less attractive on this platform.
3×32 GB: avoid. (Pangoly)

Storage¶

For this workload, storage should be kept simple. A single larger NVMe drive is a sensible default, with the option to add a second drive later if real workflow needs emerge. Current SSD pricing guides still treat high-end Gen4 as sufficient for many desktops, while Gen5 remains a premium option. For example, Samsung lists its 9100 Pro Gen5 at $419.99 for 2 TB, while current price tracking shows some 2 TB Gen5 drives falling into the high-$100s to low-$400s depending on model and stock. (Samsung be)

The better reason to use two drives is separation of OS/tools from datasets, logs, checkpoints, and scratch. For current workloads like part CAD, tool models, machine models, and toolpaths, a second dedicated “data firehose” drive is probably not required on day one. (PCWorld)

Power, cooling, and headroom¶

PSU margin and chassis shape matter because a future GPU swap is easier when the first machine already has enough electrical and physical headroom. NVIDIA’s RTX 5090 page lists a 1000 W PSU requirement. Puget’s Ryzen workstation line supports larger ATX PSUs, while System76’s Thelio Mira uses 750 W by default and 1000 W SFX with RTX 5080 / 5090 GPU. (NVIDIA)

4. Recommended strategy¶

Phase 1: one or two strong workstations now¶

The right near-term move is to buy one or two high-end AM5 Linux workstations optimized for current engineering work. The target profile is: Ryzen 9 9950X, 96 GB RAM in 2×48 if available, and either a 5080 or 5090 depending on budget and how quickly ML/simulation is expected to become the bottleneck. (Pangoly)

Phase 2: scale by adding machines, not overbuying¶

As the team grows from 2–3 engineers to 5–6, add machines incrementally instead of trying to make the first box do everything. This keeps the current purchase focused on real near-term productivity rather than speculative future capacity.

Phase 3: later add a larger shared workhorse¶

When benchmarking, batch generation, larger CI runs, or heavier shared experimentation become important, add a Threadripper-class machine. That platform is designed for more cores, more memory channels, and much more PCIe connectivity than mainstream Ryzen. (Pangoly)

5. Specific options considered¶

Puget Ryzen workstation¶

Link: Puget AMD Ryzen workstations

Why it is attractive:

Strong support posture
Conventional workstation build style
Easier future GPU and power expansion because of standard tower layouts and larger PSU options

Main caution:

Confirm whether they will support 96 GB as 2×48 GB rather than forcing 64 GB or 128 GB via 4×32. That detail matters. (Pangoly)

System76 Thelio Mira¶

Link: System76 Thelio Mira custom

Why it is attractive:

Linux-first vendor
Good default fit for Ubuntu / Pop!_OS workflows
Clear support for 96 GB via 2×48 in current configurations
Good choice if the priority is a polished Linux workstation now

Main caution:

Less PSU headroom than a larger ATX tower if the goal is a later 5090-class upgrade with lots of margin. (TechRadar)

Lower-cost mainstream prebuilt option¶

Link: Lenovo Legion Tower 7i Gen 10

Worth knowing about as a lower-cost out-of-box option, but it is more consumer/gaming-oriented and less aligned with the Linux-first, workstation-first priorities above.

6. Practical buying guidance¶

Prefer Ryzen 9 9950X. It is the cleanest fit for the current mix of interactive engineering work and background compute. (Pangoly)
Prefer 96 GB as 2×48 GB. This is the best balance of capacity and platform behavior on AM5. (Pangoly)
Treat 16 GB VRAM as the floor, not the goal. A 5080 is fine to start with; a 5090 is the better long-term ML/simulation choice if budget allows. (NVIDIA)
Start with one good NVMe drive unless there is a clear reason to split storage now. Add a second drive later if real usage justifies it. (Tom's Hardware)
Do not ignore power and cooling. If future upgrades are a priority, extra PSU margin and a conventional tower layout are worth paying for. (NVIDIA)
Ask vendors direct questions. Specifically: RAM population, expected memory speed, future GPU clearance, PSU margin, M.2 slot layout, and how easy it is to add storage later.

7. Bottom-line recommendation¶

The best current strategy is to buy a high-end Ryzen 9 9950X workstation now, configured for strong day-to-day engineering productivity rather than maximum theoretical future capacity. The sweet spot is 9950X + 96 GB (2×48) + one solid NVMe + RTX 5080 or 5090 depending on budget. Add more machines as the team grows, and add a Threadripper shared workhorse later when higher concurrency, RAM ceiling, or PCIe expansion becomes a real need. (Pangoly)

Component index with approximate prices¶

These are rough March 2026 reference prices for quick comparison. Real street pricing varies, especially for GPUs and memory.

CPUs¶

AMD Ryzen 9 9950X — about $520–$540 street price recently. (camelcamelcamel.com)
Intel Core Ultra 9 285K — about $560 current Amazon price; roughly $550–$585 typical recent range. (Amazon)
Threadripper platform — much more expensive; usually a later-stage purchase rather than first-box territory.

GPUs¶

RTX 5070 Ti — roughly $750–$900+ depending on board partner and availability; exact pricing has been volatile. (PC Gamer)
RTX 5080 — $999 MSRP, but often around $1,250–$1,400 in recent street pricing. (NVIDIA)
RTX 5090 — $1,999 MSRP, but real market pricing has often been far higher, commonly $3,000+ and sometimes much more. (NVIDIA)

Memory¶

96 GB (2×48) DDR5-5600 — very roughly $350–$850, with wide variation depending on brand and retailer; recent trackers show about $350 average for some kits, but retail listings can be much higher. (Pangoly)
64 GB (2×32) DDR5 — generally cheaper and easier to find; often the more common default in workstation configurators.
128 GB (4×32) DDR5 — higher total cost and less attractive on AM5 from a memory-speed standpoint. (Pangoly)

Storage¶

2 TB PCIe Gen4 NVMe SSD — roughly $130–$250 depending on model and quality tier. (PCWorld)
2 TB PCIe Gen5 NVMe SSD — roughly $180–$420; Samsung’s 9100 Pro is listed at $419.99, while some other Gen5 drives have recently tracked much lower. (Samsung be)
4 TB PCIe Gen5 NVMe SSD — roughly $345–$840 depending on model and retailer. (Tom's Hardware)

Candidate systems¶

Puget AMD Ryzen workstation — premium workstation pricing; better support and expansion posture. (TechRadar)
System76 Thelio Mira Custom — premium Linux-first workstation pricing; strong out-of-box Linux fit. (TechRadar)
Lenovo Legion Tower 7i Gen 10 — lower-cost mainstream prebuilt option, but less workstation- and Linux-oriented.

Appendix: One shared workhorse vs. two separate workstations¶

There are two realistic ways to spend the next chunk of infrastructure budget.

The first is to buy one larger shared workhorse: a Threadripper Pro system with server-style memory, more PCIe capacity, and a professional GPU. The second is to buy two separate high-end workstations: for example, two Ryzen 9 9950X systems with strong GPUs. Both approaches can work. They optimize for different things.

Option A: one shared workhorse¶

A shared workhorse is the better fit when the goal is to create a central internal machine for batch work, benchmarking, CI, long-running experiments, and later expansion. Threadripper Pro platforms are built for that kind of role. Compared with mainstream Ryzen, they offer many more PCIe lanes, more memory channels, ECC RDIMM support, and much larger practical RAM ceilings. For example, the Threadripper Pro 9965WX platform supports 8 memory channels and 128 PCIe 5.0 lanes, while the Ryzen 9 9950X sits on a smaller desktop platform with 2 memory channels and 24 usable PCIe lanes.

The downside is that a shared machine becomes a shared dependency. It is easier for it to turn into a queue, a bottleneck, or a “don’t touch that box right now” resource. It also creates more single-machine risk: if that system is down, a large chunk of capacity disappears at once.

A shared workhorse makes the most sense when:

we expect lots of CPU-heavy shared jobs
we need very large RAM soon
we want a machine that lives in a rack or server room
we are intentionally moving toward more centralized internal infrastructure

Option B: two separate workstations¶

Two separate workstations are usually the better fit for a small team that still values flexibility over centralization. A Ryzen 9 9950X workstation gives very strong single-core speed, enough multicore headroom for heavy local work, and a simpler ownership model. The 9950X boosts up to 5.7 GHz, while Threadripper Pro parts in the same generation are generally tuned a bit lower per core and are better understood as platform-scale upgrades rather than “faster silicon” on each core.

The main practical advantage of two machines is that work can happen in parallel without much coordination. One person can run a long training or simulation job on one box while another person still has a full machine available. It also reduces operational risk. If one workstation has an issue, the team still has another strong system available.

Two separate workstations make the most sense when:

the team is still small
heavy GPU use will usually come from one person at a time
we want to avoid creating a central bottleneck too early
we want simpler upgrades and lower operational risk

How to think about the choice¶

The shared-workhorse approach is the better long-term infrastructure shape. The two-workstation approach is usually the better near-term startup shape.

In plain language: if we were building a more mature internal compute environment today, the Threadripper Pro route would be easier to justify. But for a small team that needs immediate engineering throughput, two strong workstations are often the better business decision. They are easier to use, easier to schedule around, and more fault-tolerant in practice.

Current recommendation¶

For the current stage of the company, the better default is two separate high-end workstations first, then revisit a shared Threadripper-class machine later when the need for centralized compute, larger RAM, or heavier shared workloads becomes real.

That keeps today’s purchase focused on developer productivity while leaving a clear path to a true workhorse box later.