SparkStack

A stackable, modular, 3D-printable rack for the NVIDIA DGX Spark. One to four machines in a 205 mm footprint.

Printed in ABS or ASA. Every part fits a 220 × 220 mm bed.

Status: built and in use. Two machines sit in a prototype on a desk, and the assembly is fine to live with. The three structural parts are published below; the decorative panels are not yet.

Not affiliated with, endorsed by, or sponsored by NVIDIA. "DGX" and "Spark" are trademarks of NVIDIA Corporation. This is an independent accessory.

What's here

stl/	Ready-to-slice STLs for v3.0
sparkstack.py	Structure generator — base, tiers, cap
sparkstack_panels.py	Decorative panel generator
render_sparkstack.py	Renders: white rack, real Sparks, gold pins
render_hive.py	Renders: dark organic finish, panels on
render.sh	Blender wrapper (see Renders)
validate_stl.py	Checks exports are closed, watertight, single-shell
profiles/	OrcaSlicer profiles for ASA on a FlashForge AD5M Pro

Print set

File	Qty
sparkstack_v3_0_base.stl	1
sparkstack_v3_0_tier.stl	1 per machine
sparkstack_v3_0_cap.stl	1
sparkstack_v3_0_tier_pinned.stl	optional — the tier with the pin bore opened through the rail. See The joint

Three parts. A four-machine stack is one base, four tiers, one cap. The tier_pinned file is the same tier with one bore opened up; print it instead of tier if you want to use pins.

The decorative panels are not in this release. The generator is here and the renders below show them fitted, but the panel STLs are not published yet — see Panels for how they work and why they need a little more time.

Model

1 unit = 1 mm. Everything is generated, not hand-modelled:

blender --background --python sparkstack.py    # writes stl/ and a spec sheet

Key parameters live in the P dict at the top of sparkstack.py:

PITCH	76 mm	tier spacing; sets the air gap above each machine
DEV_W / DEV_D / DEV_H	150 / 150 / 51.2	the machine
SIDE_CLEAR	1.5 mm	post clearance
JOINT_CLR	0.15 mm	bore clearance; the ribs interfere by 0.10 mm/side
RAIL_PROUD	2.0 mm	panel rail standoff
CRADLE_CLR	1.0 mm	corner cradle clearance, sized for ASA shrinkage
BASE_R	14 mm	base corner radius

The constraint that shapes everything

The machine draws air in through its front face (upper and lower) and its underside, both dust-filtered, and exhausts out the rear heatsink fins and the QSFP fin stack. (ChargerLAB teardown)

So there is no shelf under it. The tier ring opens to 168 × 168 mm, wider than the machine's own 150 mm silhouette, leaving a 9 mm annulus around the chassis. The machine rests on four Ø14 mm pads at (±68, ±68) mm, shadowing 3.5% of the underside, all at the corners.

That is also why the panels are open webbing rather than solid skins — a closed front panel would strangle the intake.

PITCH sets the air gap above each machine: at 76 mm it is 11.8 mm. For scale, buoyancy across a full four-tier stack is ~0.26 Pa against tens of Pa from the machine's own blowers, and the intake annulus costs ~0.13–0.25 Pa at a plausible 10–20 CFM. The rack's contribution to the intake path is small, and the gap above the machine is not in that path at all: the top of the chassis is neither intake nor exhaust.

The mechanical design is built and in use. Two machines sit in a prototype on a desk — they locate in the cradles, the tiers stack and hold, and the assembly has been fine to live with.

The numbers on this page are not the reason for that. Every figure above is analysis from published specs and first principles — buoyancy, annulus pressure drop, the rising gap. No temperature logging has been done under sustained load, and the machine dimension used throughout is the published spec, not a measurement of the unit on the desk. Take the numbers as the reasoning behind the shape, not as verification of it.

Panels

Not shipped yet. The renders at the top of this README show them fitted; the STLs are not in this release. Everything below describes the design as built, so the generator in sparkstack_panels.py and the geometry it produces match what you see — but treat the panel files as unreleased until they appear in stl/.

Each tier's front comes from its own seed, so a stack doesn't read as four copies of one part. The sides are identical across tiers.

The panels are flat 4 mm plates with a hook along the bottom of the inner face that drops over a rail on the tier. Printed flat, the panel's inner face is up, so the hook is a step in the top surface — there is no overhang anywhere, which matters because ABS and ASA with near-zero part cooling will string and stay uncured on any unsupported ceiling.

An earlier attempt put the locating groove in the tier's vertical wall instead. That groove's ceiling was a flat bridge and it printed badly. Hence the rail.

Corner cradles

Each pad is backed by a square L-shaped wall, 2.5 mm thick, rising 6 mm above the seat. It stops the machine sliding in either axis. The plan is square on purpose — it matches the footprint it fits — with a 1 mm round-over on the top edge only, so no 90° edge shows in the front view. No snap lip: it was a 0.6 mm ledge hanging over open air, exactly the overhang that strings in ASA.

CRADLE_CLR is 1.0 mm rather than a tight 0.25 mm because ASA shrinks ~0.6%: over a 150 mm opening that is 0.45 mm per side, so a 0.25 mm gap would print as an interference fit.

The joint

Tiers join on four corner tenons into four sockets. The fit is meant to hold a stack up on its own — the pins are insurance, not structure.

Bore 11.30 mm, ribbed tenon 11.50 mm: 0.10 mm of interference per side across four crush ribs over a 7 mm band. Both halves are the same material, so shrinkage moves them together and the fit is unchanged after printing.

Check the fit before you commit to a full tier. sparkstack.py can generate a one-corner coupon — a half-height column carrying both halves of the joint — which is far cheaper to print than a tier. Print two and stack them.

If it comes out wrong, JOINT_CLR and RIB are the two numbers to move.

Gold pins through each joint are optional.

Which tier to print

There are two, and the difference is one bore.

The pin runs along X through a post, and at the joint it has to pass through three things: the tier's socket wall, the ring, and the panel rail. The rail is unioned on after the socket is bored, and it is solid across |x| 98–100 at z 3–9 — exactly where the pin exits. So on tier.stl the rail caps the hole: the tenon above has a clean bore, the socket below looks drilled, and a pin will not go in.

tier_pinned.stl cuts that bore once, after the rail. Verified: a clean 3.4 mm opening at z 4.0–7.0, identical to tier.stl everywhere else, same 200 × 200 footprint, same rail.

The base and the cap have no rail, so their bores were never capped and they are fine as they are.

Pins must be flush. With the rail fitted there is no room for a head — it would sit inside the rail. And a flush pin has nothing to grab: its inner end is buried in the post, so you cannot push it out from the far side either. Use an M3 screw, which the Ø3.4 bore already clears, or leave a small recess at the outer end to pick at.

Renders

Four high	Two stacks of two
Two stacks and a tower	Dark organic finish — panels fitted, not shipped
Honeycomb webbing, drips, wordmark, gold pins

render_sparkstack.py and render_hive.py drive Cycles from a script — no HDRI, no image textures, no manual scene setup. Use the wrapper:

./render.sh --background --python render_hive.py

The wrapper exists because Cycles compiles its Metal kernels on the first GPU render of a session and caches them under $HOME/Library/Caches. Under a sandboxed or read-only home that path is unwritable, the compile repeats on every launch, and you pay 150–200 s per run instead of ~10 s. render.sh points HOME at a local directory so the cache persists. Harmless outside a sandbox.

On an M2 Max at 1150 × 1500, 32 samples:

CPU	31.7 s per frame
GPU, first of a session	152–208 s (kernel compile)
GPU, warm	3.2 s per frame

cycles.preferences.devices is lazily enumerated — it stays empty until you call get_devices(), so setting cycles.device = 'GPU' on its own enables nothing and Cycles falls back to CPU without raising. Both scripts do it properly; if you copy the setup, don't drop that call.

The Spark model is not included. It is a third-party asset and not ours to redistribute. The render scripts will tell you where to download it (free, no signup); the rack itself needs none of it. Those renders are visualisations against an unofficial model, not NVIDIA CAD — don't take connector clearances off them.

Printing

ABS or ASA. Enforced, not a preference: the design leans on those materials' stiffness and temperature resistance.

Slicer profiles for both, for the FlashForge AD5M Pro with a 0.4 nozzle, in profiles/ — a process profile and a filament profile for each. Import them together; the process profile is material-independent apart from a little extra warping margin.

	ASA	ABS
Nozzle	255 °C	250 °C
Nozzle, first layer	255 °C	255 °C
Bed	105 °C	105 °C
Fan	10–15%	10–15%
Brim	6 mm	8 mm
First layer speed	35 mm/s	30 mm/s
Outer wall	120 mm/s	100 mm/s

Everything else is shared, including the near-zero cooling. ABS is the more warp-prone of the two, hence the wider brim and slower first layer — the base at 205 mm is the part that will lift if anything does.

Close the enclosure and keep the room still. ABS puts out styrene while it prints, so ventilate the room, not the printer. Textured PEI at 105 °C, and let the bed cool before removing parts or you will pull the bottom face off with them.

• Bed contact ranges 4,163–13,820 mm² across the parts
• Overhang ≤ 2.1% on every part; there are no intentional unsupported spans
• Print the panels flat — thickness is the build direction, and local z = 0 is the outer face (bed side)
• Every exported STL is verified closed, watertight, and single-shell; python3 validate_stl.py stl/*.stl re-checks it

License

CC BY-NC-SA 4.0 — see LICENSE.

You may download, modify, and print these files for personal use, and share your modifications under the same license. You may not sell prints, sell the files, or ship them inside a commercial product.

Commercial licensing is available. If you want to sell printed units or a kit, get in touch — that's a conversation, not a no.