On an invoice, a VFX workstation and an editing workstation look like the same machine. Same brands, same parts list, same case. Then you cache your first flip sim and find out they’re different machines: the GPU that plays back 4K timelines all day refuses to even launch the render, the 32GB of RAM that Premiere never filled is gone by frame 40, and the “fast” SSD chokes on a cache folder that grew to 400GB before lunch.
This guide is the VFX-specific answer. We already published a full GPU guide for editors and colorists, and if your day is Resolve and Premiere, go read that one instead. This article is for the Houdini, Nuke, Redshift, and Octane side of the pipeline, where the requirements are different in ways the vendors put in writing: SideFX now requires 12GB of VRAM just to meet Houdini 21’s GPU spec, Karma XPU only runs on CPU and NVIDIA OptiX devices, and Foundry’s own hardware partners describe Nuke users routinely filling 128GB of RAM. Everything below is built from those requirement pages and Puget Systems’ published test data, with prices pulled from live listings. I don’t run a benchmark lab; the numbers here are cited so you can check them.
One thing to accept before you spend anything: this is the worst component market in a decade. The AI buildout ate the memory supply, so everything containing GDDR7, DDR5, or NAND is inflated. GPUs are running 34 to 125 percent over MSRP depending on the tier, and 64GB DDR5 kits that cost $150 to $200 in 2024 now run $600 to $900 and up. Gartner doesn’t forecast meaningful relief before late 2027. So part of this guide’s job is telling you what to skip, and there’s a $13,250 graphics card further down that I’m going to spend a whole paragraph talking you out of.
Last updated: July 2026. Some links below are affiliate links. Buy through them and it helps keep the site running, at no extra cost to you. I only point at gear I’d be comfortable handing to a friend. Every price here was verified on July 11, 2026, and in this market they will drift, so read them as “around this much at the time of writing.”
The quick verdict
If you just want the answer: for most working VFX artists, the core of the machine is an RTX 5070 Ti (16GB), a Ryzen 9 9950X, 64GB of DDR5-6000 (the Lexar ARES Gen2 kit below), and a Samsung 990 PRO 4TB as the scratch drive. That’s around $3,200 in core parts at today’s verified prices, and every component clears the documented requirements of Houdini 21, Redshift, Nuke, and Octane with headroom. If your scenes outgrow 16GB of VRAM, the RTX 5090 (32GB) is the only sane step up. If your bottleneck is RAM rather than render speed, that’s what the Threadripper 9970X platform is for. And the Mac answer is the Mac Studio M3 Ultra, with two caveats you should read before ordering one.
Budget GPU: MSI GeForce RTX 5060 Ti 16GB
The cheapest card that actually clears the bar. Houdini 21 quietly raised its GPU requirement to 12GB of VRAM, with 16GB listed as ideal for larger simulations, and that single line disqualifies the entire 8GB card class for VFX work. This 5060 Ti has 16GB of GDDR7 at the lowest price of any current NVIDIA card, which matters because VRAM, not raw speed, is what kills a VFX card first. It clears Redshift’s 8GB minimum with double the headroom, and full CUDA and OptiX support means every major GPU renderer runs on it: Redshift, Octane, Karma XPU, and Cycles. No AMD card at this price can run all four.
The downsides: at $589.99 it’s selling about 38 percent over its $429 MSRP, courtesy of the GDDR7 shortage. The 128-bit memory bus puts its rendering throughput at roughly half of a RTX 5070 Ti. And while lookdev and lighter scenes are comfortable, production-density scenes will spill out-of-core and slow down. This is the card for learning GPU rendering and doing paid freelance work, not for a farm node.
Best GPU for most VFX artists: NVIDIA GeForce RTX 5070 Ti (16GB)
This is the value knee of the 50 series for rendering: 8,960 CUDA cores and 16GB of GDDR7 on a 256-bit bus. Maxon’s own hardware guidance says Redshift fits roughly 20 to 30 million unique triangles per gigabyte of VRAM, so 16GB works out to somewhere around 300 to 480 million triangles held in-core. That comfortably covers Houdini 21’s 16GB-ideal guidance and most commercial scenes. By the numbers, it’s the best rendering hardware per dollar you can buy right now.
Now the part that saves you $257. The RTX 5080 ($1,256.99 today) carries the exact same 16GB of VRAM. The extra money buys roughly 10 to 15 percent more speed and zero additional scene capacity, and in VFX, capacity is the spec that ends renders. This is the honest stop-buying point for most artists: past here, the next rung that adds actual VRAM is a $4,300 RTX 5090. The 5070 Ti’s own downsides: it’s about 34 percent over its $749 MSRP, and 16GB is a real ceiling for heavy production scenes. Redshift’s out-of-core rendering keeps you working past that ceiling, but it costs speed.
The money-no-object GPU: NVIDIA GeForce RTX 5090 (32GB)
The only consumer card with VRAM headroom to spare: 32GB of GDDR7 and 21,760 CUDA cores. Puget Systems measured it at roughly 35 percent faster than the RTX 4090 in offline GPU rendering (early Redshift-specific gains were smaller, about 7 to 30 percent depending on the test and driver maturity). But the speed isn’t really the pitch. For heavy Redshift, Octane, and Karma XPU scenes, 32GB is the difference between staying in-core and watching out-of-core spill eat the speed advantage you paid for.
The workstation card is the one to skip. The RTX PRO 6000 Blackwell has 96GB of ECC GDDR7 and 24,064 CUDA cores, and it launched at $8,565 in March 2025. NVIDIA’s own listing now has it at $13,250, a 55 percent increase blamed on the same memory shortage. That’s three RTX 5090s for one card, and unless you’re a solo artist whose single scenes exceed 32GB in-core (giant volumetrics, fur at scale, full scene assembly) or you’re running 70B-class LLMs locally, it buys you nothing that out-of-core rendering plus a 5090, or plain render-farm credits, doesn’t do for a fraction of the money. If what you need is more VRAM per dollar in a quiet single-slot card, the RTX PRO 4000 Blackwell 24GB is the one to cross-shop instead. The 5090’s own reality check: at $4,329.99 for the ASUS ROG Astral sold by Amazon, it’s more than double its $1,999 MSRP, the worst gouging in the whole stack, and it pulls 575W of board power, so budget a 1000W-plus PSU and serious case airflow.
The CPU for most VFX workstations: AMD Ryzen 9 9950X
The VFX CPU problem has two halves that pull in opposite directions. Puget’s Nuke guidance is blunt: most Nuke tasks are single-threaded, so high clock speed matters more than core count. Houdini is the reverse, with fluid and particle sims scaling across every core you feed them. Sixteen Zen 5 cores with top-tier single-core clocks is the best answer to both questions in one socket, and at $509.99, CPU prices are not inflated. Logic silicon dodged the memory crisis entirely. This is the one component in the whole build that’s a normal deal in 2026.
A note on its sibling: the Ryzen 9 9950X3D costs $160 more for 3D V-Cache that helps games, not renders or sims, so the plain X is the better buy unless the machine moonlights as your gaming rig. The limits are the platform’s: AM5 is dual-channel and caps out around 192GB of RAM, which is a genuine wall for deep Nuke comps and big sims, and its 24 usable PCIe 5.0 lanes constrain a multi-GPU, multi-NVMe build. If your work is CPU rendering at production scale or you’re RAM-bound, that’s the Threadripper 9970X conversation below.
The CPU for heavy sims and deep comps: AMD Ryzen Threadripper 9970X
The real reason to buy Threadripper in 2026 is not render speed. GPU renderers barely care what CPU feeds them. You buy it for the platform: quad-channel memory up to 512GB and more than 80 PCIe 5.0 lanes. Puget’s data shows Nuke users regularly filling 128GB of RAM and complex Houdini simulations wanting 256GB and beyond, numbers AM5’s ceiling simply can’t serve. The 32 Zen 5 cores also make it a legitimate CPU-render node for Arnold and Karma CPU work. At $1,999.99 on Amazon, down from its $2,499 MSRP and the lowest price it’s ever hit, this is the cheapest entry to that tier that makes sense.
Read the fine print first. TRX50 motherboards start around $700 and the platform wants RDIMMs, so the real premium over an AM5 build lands at $1,500-plus before you buy a single stick of RAM. Single-core clocks are lower than the 9950X, so Nuke interactivity is a touch worse on the more expensive machine. If you’re not RAM-bound or CPU-render-bound, stop at the 9950X and put the difference into VRAM or the render farm.
The RAM baseline: Lexar ARES Gen2 RGB 64GB (2x32GB) DDR5-6000 CL30
64GB is the documented floor for VFX work, not a nice-to-have. SideFX “strongly recommends” 64GB for fluid simulations in the Houdini 21 requirements, and Maxon’s guidance wants system RAM at about twice your total VRAM plus multitasking headroom, which lands a 16GB-GPU machine with Houdini and Nuke open at the same place. DDR5-6000 CL30 with EXPO is the sweet spot for Ryzen 9000, and at $899.99 this was the cheapest name-brand 2×32 kit at that spec I could verify, with comparable Corsair kits running $920 to $980 the same day.
Yes, that price is awful. This kit cost roughly a quarter of that in 2024, and Gartner sees no meaningful relief before late 2027, so waiting for normal has a two-year horizon. I still won’t tell you to cut here: skimping to 32GB on a sim and comp machine is the one cut that bites every single day. The RGB is a bit much for a workstation, but it came in $20 under everything comparable, so I can live with it.
The high-capacity kit: Corsair Vengeance 96GB (2x48GB) DDR5-6000 CL30
If your dailies involve pyro caches or hundred-node comp scripts, this is the sensible step up: 96GB across two DIMMs at the same DDR5-6000 CL30 spec, for $1,799.99. Two sticks matters, because it keeps a two-DIMM board viable and leaves expansion room on a four-slot one, and 96GB covers the vast majority of Houdini sim and Nuke comp workloads without crossing into Threadripper territory. In a normal market I’d call this configuration boring. In this one, it’s the least-bad way to buy real headroom.
The shortage tax is unavoidable: this capacity also sits around four times its 2024 price. And if you do need 128GB, the one to buy is the Kingston FURY 2x64GB DDR5-5600 kit at $2,497.76, but check your actual peak usage in a heavy session before spending SSD-array money on RAM. Most artists who think they need 128GB are watching a cache that belongs on the scratch drive.
The scratch drive: Samsung 990 PRO 4TB NVMe
VFX scratch is a sustained-write problem. A Houdini sim cache or a Nuke disk cache hammers a drive with hundreds of gigabytes per session, and a single sim cache can eat 500GB on its own, so capacity matters as much as speed: a nearly-full NVMe throttles its write speed exactly when you need it most. The 990 PRO’s 7,450MB/s reads on a proven TLC-plus-DRAM design have made it the reliability benchmark for this job, and 4TB is the size that keeps you away from that throttle as the cache folder grows.
At $799 it carries the NAND shortage markup like everything else, roughly two and a half times its 2024 street price. It’s PCIe 4.0, and if your board has a Gen5 M.2 slot and you move giant caches all day, the Samsung 9100 PRO 4TB (14,800MB/s) is the step-up for $140 more, though most VFX I/O patterns won’t feel the difference. One rule that isn’t optional: like all consumer NVMe drives, it has no hardware power-loss protection, so scratch data only. This drive should never hold the only copy of anything.
The Mac answer: Apple Mac Studio M3 Ultra (96GB)
Four years ago I wrote there were zero reasons to consider a Mac for VFX. In 2026 that take is out of date. Octane X is Metal-native on Apple Silicon, Redshift officially supports M-series chips, Cycles has a Metal backend, and Nuke and Houdini both ship native ARM builds. And for RAM-hungry comps, the M3 Ultra’s 96GB of unified memory is something the PC side can’t match right now: 96GB of DDR5 alone costs about $1,800 in this market, and on the Mac the GPU shares all of it. For a Nuke, Flame, or Resolve-centric finishing suite that does some 3D, this is now a defensible machine, and the full Apple decision tree lives in our Mac buying guide. If the workstation has to travel with you, the laptop guide covers the 24GB-VRAM machines that come closest to this tower.
But the caveats are the headline. Karma XPU will not run on it, period; SideFX supports CPU and NVIDIA OptiX devices only, and V-Ray GPU is CUDA-only too, so a Houdini-centric GPU rendering pipeline is a hard no. In most cross-platform GPU render benchmarks the M3 Ultra trails a $999 RTX 5070 Ti. And the timing is bad: Apple raised the base price of this config by $1,300 in June, to $5,299, Apple discontinued the 256GB and 512GB memory options so 96GB is now the ceiling, and MacRumors currently rates the Mac Studio “don’t buy now” with an M5 Ultra refresh expected around October. If you can wait until fall, wait. One buying note: Amazon’s Mac Studio listings are third-party and marked up around $600 over Apple’s own price, so the link below goes to Apple direct.
How the picks compare
| Part | Role | The spec that matters for VFX | Verified price (July 11, 2026) |
|---|---|---|---|
| MSI RTX 5060 Ti 16GB | Budget GPU | 16GB GDDR7, full CUDA/OptiX, 128-bit bus | $589.99 ($429 MSRP) |
| NVIDIA RTX 5070 Ti (ZOTAC) | GPU for most artists | 16GB GDDR7, 8,960 CUDA cores, 256-bit bus | $999.99 ($749 MSRP) |
| NVIDIA RTX 5090 (ASUS ROG Astral) | Money-no-object GPU | 32GB GDDR7, 21,760 CUDA cores, 575W | $4,329.99 ($1,999 MSRP) |
| AMD Ryzen 9 9950X | CPU for most builds | 16 cores / 32 threads, top single-core clocks | $509.99 |
| AMD Threadripper 9970X | CPU for sims / deep comps | 32 cores / 64 threads, quad-channel, 512GB RAM ceiling | $1,999.99 ($2,499 MSRP) |
| Lexar ARES Gen2 64GB | RAM baseline | 2x32GB DDR5-6000 CL30, EXPO | $899.99 |
| Corsair Vengeance 96GB | High-capacity RAM | 2x48GB DDR5-6000 CL30 on two DIMMs | $1,799.99 |
| Samsung 990 PRO 4TB | NVMe scratch drive | PCIe 4.0, 7,450MB/s reads, TLC + DRAM | $799 |
| Mac Studio M3 Ultra | The Mac answer | 28-core CPU / 60-core GPU, 96GB unified memory | $5,299 (Apple direct) |
What actually matters in a VFX workstation
VRAM is the first wall you hit. The vendors have stopped being subtle about it: Houdini 21 requires 12GB and calls 16GB-plus ideal, and Redshift’s own math (roughly 20 to 30 million unique triangles per GB) means every gigabyte is literal scene capacity. When a scene outgrows VRAM, renderers with out-of-core support keep going by paging to system RAM, at a real speed cost; renderers without it just fail. That’s why every card in this guide has 16GB or more, and why the 8GB cards a tier down are a dead end no matter how the price looks. If your machine also pulls editing duty, the RTX 5060 Ti 16GB is also the cheapest card with hardware 10-bit 4:2:2 decode, which we covered in the GPU guide.
Check your renderer’s hardware page before you buy anything. This is the CUDA lock-in problem. Karma XPU currently supports CPU and NVIDIA OptiX devices only. OctaneRender on Windows supports NVIDIA CUDA only. V-Ray GPU: CUDA only. “But doesn’t Redshift support AMD now?” It does: RDNA 2 and later, officially, on Windows. And Blender Cycles is the most backend-agnostic renderer going (CUDA, OptiX, HIP, oneAPI, and Metal). But a workstation is a five-year bet on your pipeline, and betting that you’ll never touch Karma, Octane, or V-Ray GPU is how you end up rebuying a graphics card. NVIDIA is the default here for boring, documented reasons.
Budget more RAM than feels reasonable. Houdini 21’s official ladder is 16GB minimum, 32GB recommended, 64GB strongly recommended for fluid sims, and Puget’s real-world numbers run higher: Nuke users consistently filling 128GB, complex Houdini sims wanting 128 to 256GB and beyond. The platform question follows from that appetite. Consumer AM5 tops out around 192GB; Threadripper’s TRX50 platform goes to 512GB on quad-channel. Buy the platform for your peak RAM day, not your average one.
Clocks for comps, cores for sims. Nuke is mostly single-threaded, so a higher-clocked 16-core beats a lower-clocked 32-core for interactive comp work. Houdini’s fluid and particle solvers are the exception; they’ll use every core you have. Know which one is your bottleneck before you pay the Threadripper premium, because it’s possible to spend $1,500 extra making Nuke slightly slower.
Spend around the shortage, not into it. The 2025-26 memory crisis didn’t inflate everything equally, and the pattern is consistent: the further up the GPU stack, the worse the gouging. Verified against MSRP today: the 5060 Ti is 38 percent over, the 5070 Ti 34 percent, the 5080 26 percent, and the 5090 sits at 110 to 125 percent over. Meanwhile CPUs are selling at or below MSRP. So the shortage-era strategy writes itself: buy your CPU and platform happily, buy the cheapest GPU tier that fits your scenes, buy RAM once instead of twice, and skip anything you can’t justify with a specific scene that’s failing today. If your current card still renders your work, keep it another year and let someone else pay the markup.
FAQ
How much VRAM do I need for VFX work in 2026?
The documented floor is 12GB, because that’s what Houdini 21 now requires to meet spec. In practice, 16GB is the working minimum (SideFX calls it ideal for larger sims, and Redshift’s triangles-per-GB math gives it real headroom), and 32GB is for production-density scenes that have to stay in-core. Nothing in a VFX build should have 8GB.
Can I use an AMD GPU for VFX?
Partially, and that’s the problem. Redshift now officially supports AMD RDNA 2 and later on Windows, and Blender Cycles runs on AMD via HIP. But Karma XPU supports only CPU and NVIDIA OptiX devices, OctaneRender on Windows is CUDA-only, and V-Ray GPU is CUDA-only. Unless you can guarantee your pipeline never touches those three, NVIDIA remains the safe bet.
Is a Mac good enough for VFX now?
Better than it’s ever been, with a hard boundary. The Mac Studio M3 Ultra runs native ARM builds of Nuke and Houdini, Metal-native Octane X, and M-series Redshift, and its 96GB of unified memory is a real advantage for comp work. But Karma XPU and V-Ray GPU won’t run on any Mac, and its GPU rendering trails a $999 PC card. It’s a finishing and comp machine that can do 3D, not a render box. Also: with a June price hike to $5,299 and an M5 Ultra expected around October, MacRumors rates it “don’t buy now,” and I agree.
Should I wait out the memory crisis before building?
If your current machine renders your current work, yes, wait; this is the worst pricing in a decade and analysts don’t see relief before late 2027. If a specific job is failing on VRAM or RAM today, build now and buy exactly what the failing scene needs, because two years is a long time to bill proxy-speed hours waiting for a discount that isn’t scheduled.
Where this fits in your pipeline
This machine is one node in a bigger workflow. If your work leans editorial and color rather than sims and renders, the editing GPU guide and the Resolve 21 breakdown are the better maps. And when a single box stops being enough, the answer usually isn’t a $13,250 graphics card, it’s the cloud: I built an entire VFX studio on AWS years ago, and renting render capacity by the hour is still how I’d absorb a deadline that outgrows the workstation.
And that’s it! The usual caveat applies double here: these picks are built from the vendors’ documented requirements, Puget’s published data, and live prices, not from a benchmark lab in my garage, and this market moves fast enough that I’ll be re-verifying every price on a regular basis. If your workflow hits a wall these specs don’t explain, drop it in the comments. The rest of our buyer’s guides live on the Gear Guides page.
