Puget Systems

08/31/2026

We recently worked with a client who builds their own systems regularly, but needed something outside their usual range: a rock-solid testbed that could image multiple USB drives simultaneously as part of their deployment pipeline. Rather than chase down workstation-grade parts on an unfamiliar platform during a supply crunch, they came to us.

We built them a Threadripper platform with 14 USB ports, letting them image the drives in under 30 minutes; a job that previously took 2-3 separate test benches to accomplish.

This is the value of a supply chain team that maintains real OEM and distributor relationships. Even when the broader market is tight, we can still build the right system configurations for the workflows that need them.

Read the full story: https://hubs.ly/Q04v-N0m0

08/25/2026

Liquid-cooled GPU systems are now closer to home.

The partnership between Puget Systems and Comino Liquid Technology is evolving to bring Comino’s unique liquid-cooled GPU platforms to North American customers with final assembly, integration, testing, and support handled by Puget Systems.

What this means for you:
• Final integration, system validation, testing, and burn-in in the US
• A more direct path from configuration to deployment
• 3-year coverage backed by Puget Systems Support

The lineup is designed for demanding deep learning & AI inference, HPC & scientific research, and other GPU-intensive workflows where system density, deployment timelines, and support ownership matter.

Explore Puget Systems, Powered by Comino: https://hubs.ly/Q04v8sZM0

08/18/2026

The most useful thing we found testing two AMD Radeon AI PRO R9700s wasn't a benchmark number. It was what didn't work.

Stock vLLM multi-GPU does not run on RDNA 4. Not tensor parallelism, not pipeline parallelism. Both abort during RCCL collective initialization before a model ever loads.

If you're speccing a dual-R9700 workstation on the assumption that the standard vLLM recipe applies, you'd find out after the PO cleared.

The working path is llama.cpp's ROCm backend, which distributes over direct HIP transfers with no collectives. It runs Qwen3.6-27B at Q4 across both cards at 23.4 tok/s single-user, 56 tok/s aggregate at eight concurrent users. That's genuinely usable for the 27B tier.

Full testing and setup guide:

Can two Radeon AI PRO R9700 cards deliver production-quality local LLM inference and image generation? How do they compare against Intel's Arc Pro B70 GPUs?

08/13/2026

SIGGRAPH 2026 was a blast and we had more going on than we could fit into one booth!

We teamed up with Form Studios and AMD to bring the Robo Dojo to life a real-time motion capture experience where attendees fought AI-powered robots and walked away with a branded clip generated on the spot.

Down the hall, District Cinema's "Roll for Reality" turned an actual play D&D session into a full-on virtual production experience.

And our president, Jon Bach, led a Birds of a Feather discussion on where workstations are headed as AI workloads and memory constraints reshape infrastructure decisions.

Watch the recap video below!

SIGGRAPH 2026 is a wrap, and what a week it was!From AI-powered m...

08/10/2026

From a viral YouTube series to a real-time AI-powered video game!

SoKrispyMedia's Chalk Warfare is now in beta and episode one of the making-of series is live on YouTube!

This is what it looks like when a creative team with a wild idea gets the right tools and the right partners behind them.

Watch the making-of → https://hubs.ly/Q04sqYlF0
Sign up for the beta → https://hubs.ly/Q04sq-jY0

Upgrade your online protection with an all-in-one security app! ➵ ...

07/28/2026

Gaussian splatting is changing photogrammetry. We tested seven GPUs to find out which one actually keeps up.

Pix4Dmatic added Gaussian splatting support this year; a GPU-intensive processing method that builds more detailed, more accurate 3D scenes than traditional structure-from-motion photogrammetry. It's a meaningful upgrade for teams doing aerial mapping, cultural heritage capture, or as-built documentation.

We tested every NVIDIA GeForce generation back to the RTX 20 series across two real-world datasets: a small pumphouse capture and a much larger cathedral scan.

What we found:
🔹 On smaller projects, GPU choice barely matters. The gap between the fastest and slowest card was just 20%
🔹 On larger, higher-resolution datasets, the gap becomes significant. The RTX 5090 finished splat processing 47% faster than an RTX 3060
🔹 Calibration and mesh generation steps are largely GPU-agnostic; splatting is where the real hardware differences show up
🔹 Systems older than 2018 (pre-RTX 20 series) aren't capable of this workflow at all

Our recommendation: the RTX 5080 is the sweet spot for most Gaussian splatting work, with the 5090 justified only for the largest capture projects.

The full breakdown of both datasets and every sub-step can be found at the link below. What's your team's biggest photogrammetry bottleneck right now; capture time or processing time?

07/27/2026

District Cinema builds centimeter-accurate digital twins of real-world locations using Gaussian Splatting, working on productions for Apple TV and HBO. Their workflows can't touch the cloud due to strict content security requirements, which means every hour of processing has to happen on-prem, on hardware that can't afford to fail.

Technical Director Nolan Kearney put it plainly: "The Puget Systems hardware had been running nearly 24/7 for six months without a single issue". That includes processing massive mesh extraction for HBO's Rooster, work that requires exactly the kind of raw horsepower needed to say yes to client requests that "have just never been done before".

Read the full case study for how District Cinema is using Threadripper PRO and RTX PRO 6000 Blackwell workstations to compress timelines that used to depend on the cloud.

07/24/2026

Just out here looking cooler than we planned.

07/24/2026

Every sales team has a secret weapon 😂