Move over GeForce RTX 5090, there's a new king in town. That's the Nvidia RTX Pro 6000 Blackwell Workstation edition, which is more powerful (and much more expensive!) than the RTX 5090. While the 96GB of ECC-capable VRAM is designed for AI, CAD, 3D modeling, and other professional workloads, the first question I had was "can it game?"
After all, Nvidia can't have forgotten about gamers, and the Pro cards are still compatible with G-Sync, frame generation, and DLSS. We got our hands on one, and I've been feverishly benchmarking it vs an AIB RTX 5090, and it spanks the flagship consumer card. Not by much on some games, but others have a much wider delta. Again, you'll pay for the extra fps, but with GPU prices being what they are now, the gap is much smaller than it was at launch for the $10,000 Pro graphics card.
I went back to native resolution for a week, but Nvidia's DLSS still made it feel like the wrong choice
Technically superior no longer means visually better
Nvidia's RTX Pro 6000 Blackwell is beastly
96GB of GDDR7 VRAM to stuff with gaming textures
The RTX Pro 6000 is based on the GB202 GPU die used by the RTX 5090, so you might expect similar performance. But what you get here is a near-perfect die with more cores enabled and a whopping 96GB of GDDR7 memory for some insane bandwidth.
One quick note, it doesn't support the Game Ready drivers that the RTX 5090 does, but that just means per-game tweaks might not be available for newer games. I tried to mitigate this by using older games with mature ray tracing and DLSS implementations, and the workstation drivers are still capable of handling Nvidia's technology stack.
Nvidia RTX Pro 6000
The RTX 5090 is no slouch, either
The Nvidia GeForce RTX 5090 isn't that far behind the workstation bigger brother, with roughly three thousand shader units fewer. That's not a huge amount, considering it's hard to push the 5090 to full utilization. The additional 18 ray-tracing cores on the RTX Pro 6000 might be more beneficial for gaming, but that depends on the other components being able to feed the GPU data.
Nvidia GeForce RTX 5090
- Shader Units
- 21,760
- Ray Accelerators/Cores
- 170
- Base Clock Speed
- 2,010MHz
- Boost Clock Speed
- 2,410MHz
- Memory Capacity
- 32GB GDDR7
- Memory Bus
- 512-bit
- Memory Bandwidth
- 1,792GB/s
- Power Draw
- 575W
- Architecture
- Blackwell
- Price
- 2000
More specifically, I'm using the Gigabyte GeForce RTX 5090 Gaming OC, partly because it's the only 5090 I have on hand, but also because it's one of the better ones. If I wanted, I could flash the unlocked VBIOS onto it and set it up to pull 800W or more from the 12VHPWR cable, but I'm keeping things stock because I don't want anything to melt.
Gigabyte GeForce RTX 5090 GAMING OC
Let's talk test platform
Because a GPU is useless on its own
All testing was performed on the same system, with only the GPU swapped out after a full set of benchmark runs. As befitting the two behemoth GPUs used, the rest of the test system was equally impressive.
|
GPU |
Nvidia GeForce RTX 5090, RTX Pro 6000 Blackwell Workstation Edition |
|---|---|
|
CPU |
|
|
Motherboard |
MSI MPG X870E Carbon Wi-Fi |
|
RAM |
96GB G.Skill Trident Z DDR5-6000 CL26 |
|
SSD |
Samsung 9100 Pro 8TB |
|
Cooler |
Tryx Panorama 360 ARGB |
|
Case |
Havn 360 BF Flow |
|
PSU |
MSI MEG Ai1600T PCIE5 |
All testing was also done at 3440x1440 on an Alienware QD-OLED ultrawide. Neither of these GPUs should have any issues with the resolution, and we're trying to put them to work, so I'm not even going to entertain running anything at 1080p. This is a shootout from the top tiers of professional and consumer GPUs from Team Green, and should be treated with respect.
I've been running some of the biggest open-weight LLMs for free on Nvidia's cloud
You can use big models for free, though there aren't any promises on speed.
All systems go!
Time to put the Blackwell Beast through its paces
Okay, using a professional workstation graphics card to play games seems a little bit overkill, but we like a bit of that sometimes. This is the best GPU you can buy from retail while still using consumer platforms to run it, and with all that VRAM, it should chew up and spit out any game that it chooses. And it does, with every game with at least a modest improvement to fps scores.
|
Gigabyte GeForce RTX 5090 Gaming OC 32G |
Nvidia RTX Pro 6000 Blackwell |
Percent change |
Notes |
|
|
Doom: The Dark Ages (Ultra Nightmare) |
122 fps |
128 fps |
4.9% |
No DLSS |
|
Doom: The Dark Ages (Ultra Nightmare) |
259 fps |
280 fps |
8.1% |
DLSS Balanced + Frame Gen |
|
Doom: The Dark Ages (Ultra Nightmare Path Tracing) |
285 fps |
302 fps |
6% |
DLSS Balanced + 4X |
|
Cyberpunk 2077 (Ultra) |
167.5 fps |
193.1 fps |
15.3% |
No DLSS |
|
Cyberpunk 2077 (Ultra) |
322.2 fps |
343.3 fps |
6.5% |
DLSS Balanced + Frame Gen |
|
Cyberpunk 2077 (RT Overdrive) |
381.2 fps |
397.2 fps |
4.2% |
DLSS Performance + 4X |
|
Black Myth: Wukong (Cinematic) |
133 fps |
192 fps |
44.3% |
DLSS Quality + Frame Gen |
|
Black Myth: Wukong (Cinematic RT) |
271 fps |
277 fps |
2.2% |
DLSS Balanced + 4X |
|
Star Wars Outlaws (Ultra) |
120 fps |
135 fps |
12.5% |
No DLSS |
|
Star Wars Outlaws (Ultra RT) |
302 fps |
430 fps |
42.4% |
DLSS Balanced + 4X |
Some games outperformed what I'd expected, with Wukong and Star Wars Outlaws both zooming ahead when ray tracing was cranked up. Maybe that's not such a surprise, as the additional cores and memory bandwidth really come in clutch here, but most games were closer to 10% on average.
Professional GPUs can still push out the frames
The line between Nvidia's consumer and prosumer GPUs has always been a little blurry, first with the Titan range sitting above the xx80 cards, and then with the introduction of the xx90 tier to replace that. But the RTX Pro 6000 is another bump above that, built for demanding server and workstation tasks that require the right tools, regardless of cost. Even at the reduced price since launch, it's still over twice the cost of a RTX 5090, but then again if you absolutely need the VRAM, it has no equal.
