The Two-GPU PhysX Trick Is Back: Pairing an RTX 5060 With a 5090 to Fix Nvidia's Own Mess

When Old Tech Bites Back

Here's a sentence I didn't expect to write in 2026: a $2,000+ flagship can still get knocked around by a 17-year-old physics middleware. But that's where we are. PhysX — the GPU-accelerated physics tech Nvidia inherited from Ageia in 2008 — quietly turned into a landmine for RTX 50-series owners when 32-bit CUDA support got axed at the start of 2025. Anyone who fired up Batman: Arkham Asylum , Mirror's Edge , or Borderlands 2 on a brand-new 5090 watched their framerate fall off a cliff because the workload was suddenly being dumped onto the CPU.

Nvidia walked it back with the 591.44 Game Ready driver in December, which restored 32-bit PhysX hardware acceleration on Blackwell for an initial nine titles, with Arkham Asylum added shortly after in early 2026. That fix gets these games playable again. What it doesn't do is stop PhysX from eating an enormous chunk of your GPU budget when you turn it on. And that's where a trick from 2010 — a dedicated secondary GPU just for physics — gets dragged out of the graveyard. Fresh testing of a 5090 + 5060 pairing in the Arkham series has produced numbers loud enough that I want to dig into what's actually going on, why it works, and whether anyone outside of YouTube benchmarkers should care.

What PhysX Actually Costs You on a Single Card

Even after the driver fix, PhysX is expensive on modern hardware. Not because Blackwell is bad at physics — it's because the original PhysX implementations in those late-2000s/early-2010s games were tuned for CUDA cores that were a tiny fraction of today's GPU resources. The shader scheduling, the memory access patterns, the 32-bit code paths — none of it was written with 21,760-CUDA-core monsters in mind. So when you crank PhysX to High in Arkham City , your 5090 isn't politely sharing cycles between rendering and physics. It's serializing legacy 32-bit physics work on a tiny slice of its execution units and choking on the round trip.

The numbers tell the story. On a 9800X3D platform at 4K maxed out, dropping a dedicated 5060 in for PhysX duty produced a 76% jump in average framerate in Arkham Asylum (from roughly 220 FPS to nearly 400), a 66% jump in Arkham City (up to 244 FPS average), and a more modest 25% bump in Arkham Origins with a slight regression in 1% lows. Even Arkham Knight , which uses 64-bit PhysX and runs natively fine on Blackwell, picked up 19% in averages and 13% in 1% lows .

What jumped out at me reading those results: the 5060 wasn't even sweating. Average utilization for the dedicated PhysX card ranged from 14% in Arkham City to 42% in Arkham Knight . Peak utilization topped out at 61%. In other words, you don't need a beefy second GPU. You need a second GPU. Anything CUDA-capable with enough VRAM to hold the physics scene will do — which means that mythical drawer of retired 1660 Supers, 3050s, and 2060s sitting in every enthusiast's closet just got a job again.

How the Setup Actually Works

The mechanism here is dead simple and hasn't really changed since the late 2000s. In Nvidia Control Panel, under "Set PhysX Configuration," you pick the secondary card from the dropdown. There's a "Dedicate to PhysX" checkbox if you want to lock that card out of being used as a render target. Driver handles the rest — the primary GPU does the rasterization, ray tracing, and shading; the secondary GPU runs the physics scene; data shuffles over PCIe.

This is not SLI. It never was. SLI required matched GPUs, NVLink/SLI bridges, profile support, and cooperated on the same frame. Dedicated PhysX is asymmetric, asynchronous, and works with mismatched cards across multiple architectures. SLI's been dead since Ampere, but this trick survived because it's structurally different — and because the games that benefit from it are old enough to be permanently pinned to legacy code paths.

A few things worth flagging before anyone tries this. Both cards need a PCIe slot — a dedicated PhysX card runs fine even at PCIe x4 or x1 because the data volume is small, but you need a physical slot wired up, and plenty of "premium" boards now skimp on the second long slot. Both cards need power, too: a 5060 is 145W, a 5090 is 575W, so you're looking at a 1000W PSU minimum and 1200W recommended. Case airflow matters more than people think — two GPUs stacked on top of each other is the worst possible thermal layout, with the bottom card choking the top card's intake. And the driver has to play nice: the 32-bit PhysX restoration is per-title and hand-curated, covering Arkham Asylum , City , Origins , Mafia II , Mirror's Edge , Metro 2033 , Metro: Last Light , Borderlands 2 , AC IV: Black Flag , and Alice: Madness Returns as of the latest drivers. Arkham Knight is 64-bit and works regardless.

Why the Performance Delta Is So Big

I've been benchmarking GPUs long enough to be skeptical of any claim that promises a 76% framerate gain from a config change, so let me put my own framing on these numbers. The reason the gap is enormous in Arkham Asylum and City and shrinks in Origins and Knight comes down to how each game uses PhysX, not the raw amount of physics work.

Asylum and City lean on dynamic cloth, particle, and rigid-body effects that hit GPU shader ALUs hard and contend with rendering work for the same execution resources. When physics and rendering are fighting over the same SMs, you don't just lose physics performance — you lose render throughput too, because pipeline bubbles cascade. Offloading physics to a separate card breaks that contention completely. The 5090's render pipeline gets all 21,760 cores back, and you see compounding gains.

Origins uses PhysX more conservatively. Less cloth, less per-frame particle volume, fewer simultaneous rigid bodies. The contention was never as bad to begin with, so removing it produces a smaller win. Arkham Knight is even more telling — its PhysX is simpler (interactive smoke, paper debris) and it runs as 64-bit code, which Blackwell handles natively. The 19% gain there is genuine but it's the gain from offloading any GPU compute task to a second card, not from rescuing a broken legacy code path.

The 1% low regression in Origins is the kind of thing I'd expect from PCIe latency on small physics scenes. When the workload is light, the cost of shipping physics state across the bus per-frame can occasionally exceed the cost of just doing it locally. It's a wash on heavy physics scenes; on light ones, it can hurt. Worth knowing, not worth panicking over.

Test Setup Context

The bench used for these results is a 9800X3D, 64GB DDR5-6200 CL30, X870-class motherboard, PCIe Gen5 NVMe boot drive, RTX 5090 primary + Asus Dual RTX 5060 OC as PhysX, Windows 11 25H2. That's a high-end-but-not-exotic AM5 setup. On my own AM5 bench — flagship X3D-class CPU, fast DDR5, PCIe 5.0 NVMe — I'd expect the same general scaling pattern. The bottleneck in these games at 4K is overwhelmingly the GPU side, and CPU choice within the X3D family barely moves the needle once PhysX is offloaded.

The piece I'd want to see tested next is PCIe slot bandwidth impact — does running the secondary card at x4 vs x16 matter? Almost certainly not for PhysX scene sizes, but data would settle it. The other obvious gap is lower-end secondary cards . Would a 1050 Ti or a 1660 Super deliver 90% of the same uplift at zero marginal cost from the parts bin? Given the 14-42% utilization numbers, I'd bet yes, and that changes the value calculation completely.

The Competition: Other Ways to Make These Games Run Right

Who This Is Actually For

Almost nobody, and that's fine. If you bought a 5090, you bought it for Cyberpunk path tracing, Alan Wake 2 at 4K with MFG, Black Myth: Wukong with everything turned up — not for hitting 400 FPS in Arkham Asylum . The 5090 alone, with the PhysX fix in place, runs every Arkham game at 4K maxed out at framerates that comfortably exceed any reasonable monitor's refresh rate.

This is for you if: you're an enthusiast who already has a spare CUDA-capable GPU (anything from a 1060 up), a case with airflow for it, a PSU with the headroom, and you genuinely care about peak 1% lows in revisits of legacy PhysX titles. The cost of "free" is just slot space and a few extra watts.

Skip it if: you'd need to buy a second card, modify your case, or upgrade your PSU. The math doesn't work. A used 3050 + a PSU upgrade + a bigger case to hit 400 FPS instead of 230 FPS in a 2009 game is the definition of diminishing returns.

What this whole episode actually demonstrates isn't that dual-GPU PhysX is back — it's that Nvidia's decision to drop 32-bit CUDA support without warning was sloppy, the U-turn was correct but incomplete, and the legacy of PhysX as a proprietary, vendor-locked physics library is still causing problems sixteen years after it was acquired. Had PhysX been opened to run on any GPU, we'd have ten times the games using it and none of this drama. Instead, in 2026, we're rigging up two-GPU setups to work around a driver decision that should have been a simple no-op.

The good news: the trick works. The better news: most of you don't need it.

---

Score: 7.5/10 (as a technique , not a product)

Strong gains where they appear, but narrow applicability and real practical friction.

Pros

• Genuinely large framerate gains in Arkham Asylum and City (66-76% averages)
• Works with cheap, low-end secondary GPUs (a 5060-class card barely breaks 50% utilization)
• Setup is trivial — Nvidia Control Panel checkbox, no SLI bridges, no matched cards
• Bypasses the lingering CPU-fallback penalty for less-than-perfectly-optimized titles

Cons

• 1% lows can regress in lighter PhysX workloads ( Origins is the cautionary tale)
• Requires PCIe slot, PSU headroom, and case space many builds don't have
• Solving a problem that mostly didn't need solving once 591.44 shipped
• Per-game support means future titles aren't guaranteed to benefit
• Two-GPU thermals are a real headache in stacked layouts