As the "legitimate successor" to the RTX 5000 Ada, how much new performance has NVIDIA actually squeezed out this time? Is it a genuine upgrade or just a "rebranded card"?

Today, we'll put these two generations of graphics cards to the test and see if the Blackwell architecture is really pulling its weight!

 1.1  FurMark: The "Veteran Doctor" of Stress Testing

There's no need to introduce FurMark, is there? Known as the "graphics card roaster", it beats all comers.

Test method:

· 4K resolution OpenGL rendering

· Long-term stress test at room temperature (to see if it will "throttle down and surrender")

Here are the results:

· OpenGL 4K performance of RTX PRO 5000 Blackwell: 1.61 times that of RTX 5000 Ada

· Full-load temperature: Steadily kept below 86 degrees Celsius (the target temperature is exactly 86 degrees, indicating a very stable cooling strategy)

86 degrees Celsius, this temperature control is as precise as NVIDIA's "masterful calibration".

With a higher power limit, the performance release is aggressive, yet the temperature remains under control. I'd give this operation 82 points out of 100, and the remaining 18 points will be issued in the form of 666.

1.2 3DMark: The "Gaokao" of Graphics Performance

3DMark is the "standard exam" in the graphics card world, and we ran three test items:

· Steel Nomad (DX environment): Testing DirectX performance

· Port Royal: Tests ray tracing performance

· Steel Nomad Volcano:

Score Report:

· Port Royal (Ray Tracing): 1.29x ↑

· Steel Nomad (Vulcan): 1.35 熊 ↑

Whether it's DX, Vulkan or ray tracing, the entire Blackwell series outperforms Ada by a wide margin.

Although the ray tracing performance improvement is relatively modest (1.29x), considering this is a professional card rather than a gaming card, this result is already quite impressive.

1.3 Omniverse: NVIDIA's "beloved" software

Omniverse is NVIDIA's in-house collaboration platform that supports multiple users to create 3D scenes online simultaneously and can also use DLSS for a performance boost.

We tested two scenarios: enabling DLSS and disabling DLSS.

Turn on DLSS, and the performance skyrockets! What does 2.6x mean? It means a scene that used to take 1 hour to render now takes just over 20 minutes.

Even without turning on DLSS, there's still a 25% performance boost, which is quite a thoughtful "guaranteed performance" feature.

2.1 V-Ray Benchmark: The "Benchmark" in the Rendering Industry

V-Ray has two engines:

· RTX engine (using ray tracing)

· CUDA engine (using CUDA cores)

Scores:

· RTX Rendering: 1.7x ↑

· CUDA Rendering: 1.4x ↑

2.2 Blender Benchmark: The Champion of the Open Source World

Performance Comparison:

· Monster: 1.39x

· Junkshop: 1.81x (Highest)

· Classroom: 1.47x

2.3 Octanebench: The "Dignity Battle" of a Veteran Renderer

As one of the earliest renderers to support ray tracing, Octane also delivers impressive performance results:

Score: 1.6x ↑

For offline rendering, Blackwell delivers an average improvement of 40% to 80%, which is the highlight of this upgrade.

Score: 1.3x ↑

A 30% improvement doesn't seem earth-shattering, but considering that text-to-image generation is already fast (producing images in a few seconds), the actual experience may not be very noticeable.

But if you need to generate images in batches, this 30% improvement can save you a lot of time.

3.2 Text-to-Video: ComfyUI + Hunyuan Large Model

Video generation is far more resource-intensive than image generation, and it's also a great way to test GPU memory and computing power.

We used the default text-to-video workflow of Hunyuan Large Model in ComfyUI:

Score: 1.38x ↑

Video generation has improved by 38%, which is quite good.

But I'd like to remind you here: the biggest problem with video generation isn't slowness, but the video crashing directly due to insufficient video memory. So 48GB of video memory is the real "killer feature" of this card.

3.3 Large Model Inference: MLPerf Client v1.5

This test mainly looks at two metrics:

· TTFT (Time to First Token): How long it takes for the AI to "think

· TPS (Tokens Per Second): how fast the AI "speaks

Scores:

· Time to First Token (TTFT): 20%~50% faster ⚡

· Generation Speed (TPS): Increased by 70%~90%

TPS nearly doubles? ! This means that when you run a 70B large model, Blackwell can deliver nearly twice the generation speed for you.

For players who deploy large models locally, this is simply a "physical cheat"!

After all that, let's give you a "total account":

Offline Rendering (Comprehensive) ⬆ 40%~70%

Text-to-Image (ComfyUI) ⬆ 30%

Text-to-Video (ComfyUI) ⬆ 38%

Large model inference (TTFT) ⬆ 20%~50%

Large Model Inference (TPS) ⬆ 70%~90%

So is the RTX PRO 5000 Blackwell worth buying?

If you are such a user:

✅ Gamers running large models (48GB/72GB VRAM is a real bargain)

✅ Professional rendering professionals (users of V-Ray/Blender/Octane)

✅ Enterprise users with deep pockets

If you are such a user:

❌ Only dabble in design

❌ Limited budget (you know the price of professional GPUs)

❌ No need for 48GB of VRAM

The performance of the Blackwell architecture on professional graphics cards this time is indeed much more genuine than the "minor upgrade" approach on gaming cards.

Especially the improvement in AI performance is clearly aimed at the current AI boom. The 48GB/72GB VRAM plus the significantly increased AI inference speed are obvious hints that you should buy it to run large models!

But then again, the price of professional graphics cards... Well, let's just say: NVIDIA's pricing strategy has never failed to hit the mark.

Do you think this generation of Blackwell professional graphics cards is worth buying?

Test data source: Leadtek