Published: Jun 22, 2025

Building my Dream Workstation

Sapphire Rapids

In early 2023 Intel launched their 4th gen Xeon Scalable processor, it was originally planned to launch in 2019 so the delay was pretty bad. The generation was called Sapphire Rapids, and was meant to be a huge leap forward in a whole bunch of features: DDR5, PCIE5.0, built-in hardware accellerators (DSA, QAT, DLB, and IAA), CXL1.1, and a whole bunch of new vector and matrix instruction sets. I love it when hardware companies are ambitious, but in hindsight this was way too many new features and it seems inevitable now that it was extremely delayed. But it did launch, eventually.

Whenever Intel launches a new Xeon they tend to drop an absurd number of variant SKUs, and this generation was no different. One thing I didn’t expect though was that they launched a special set of SKUs that were so different that they weren’t even socket-compatible with other Xeons from the same generation; what they eventually called Xeon Max. These Max processors very weird, and most importantly they had 64GB of HBM2e memory physically colocated on the package next to the CPU dies. It was an interesting experiment, and it’s kind of amazing that they try these very niche variants at all. Unfortunately the product line was discontinued after this generation so these are the only HBM x86 CPUs that exist.

I watched this Xeon launch with a lot of interest, since the generation was supposed to include a line of workstation SKUs (that ended up coming even later), and at the time I really lusted after the Max CPUs. But they were extremeley expensive (about $13k USD) and they required special motherboards and such, so these were quickly unobtainable. They ended up in some niche HPC clusters but overall didn’t see as much interest as I thought they deserved.

Brutal depreciation

Sometime in late 2024 one of the HPC clusters that was built on Xeon Max was being decomissioned and the parts were appearing online. This is a very short cycle time even by server hardware standards, and I wasn’t expecting to see these parts available second-hand for a few more years. Since these parts are so niche, they were being sold for a huge markdown (again, even by server hardware standards); about 90% off of their MSRP from only 18-months earlier.

I really wanted one of these things, but it was a big risk to buy one. I first verified that the silicon that was appearing on the market wasn’t pre-production engineering samples. Then I quickly did as much research as I could to find a 4th gen Xeon SP motherboard that supported Max variants (very few of them do, since the HBM requires substantially more power delivery to the socket and they also require the E1C package carrier because they are shaped physically different from other Xeons of that generation) and a cooling solution (similar issue, needs to dissipate more power and a passive cooler won’t work for a workstation since it’ll never get enough static pressure). I was pretty sure I had a build that would work and I quickly placed an order for everything and crossed my fingers.

What a weird machine

I was able to get a Dynatron L45–4677 AIO cooler that somewhat fits into a desktop case, a Gigabyte MS03-CEO motherboard that specifically supports the Max processors in an eATX form factor, and the 9480 itself. Plus the regular boring stuff like a PSU, storage, a case, etc. All-in, the hardware was probably about $3k, for an absolutely monster workstation setup. I wasn’t sure it would work at all so I didn’t bother building it into the case at first, I just built out the components on my kitchen table and shorted the power pins to boot it using a screwdriver. I used a 7-inch VGA display to setup the machine, and had some absurd struggles with updating the CPU microcode, but eventually it POSTed successfully! Kind of lucky to be honest, if anything had failed there are very few resources out there to debug the situation.

lstopo view of the Intel Xeon Max 9480
lstopo view of the Intel Xeon Max 9480

I decided not to bother (for now) with registered DDR5, since it’s still very expensive; so the machine currently boots with no memory populated on the motherboard at all. Very neat! Since the architecture is x86 all of the software installed quite easily and worked as expected.

What’s next?

There’s only two things I plan on adding to the system and neither are urgent. I’d like to add a few hundred gigs of registered DDR5, and I’d like to fabricate a bracket so that the 3U AIO can be mounted better to the case’s 140mm fan holes (it’s currently mounted kind of crooked with only 2 screws, not ideal but also not causing any problems). Overall though I’m extremely happy (and also relieved) with how it all turned out.

I’ve already started playing with some local llama models running fully in the CPU, and also experimenting with some numa configs to bind models to a specific die and HBM stack to try to measure the overhead of inter-die communication and traversing to a non-local HBM stack. It’s been a lot of fun to play with, and I’m excited to start experimenting with some personal projects. I wish more semi companies did weird niche hardware like this, I often think of the hardware lottery paper and what we might be missing with a general lack of ambition in hardware design.

This thing is so cool.