NVIDIA RTX Spark: The Benchmarks Are In And The Future Of PCs

Yeah, so really exciting week, right? NVIDIA lifted the embargo on actual independent testing and analysis of RTX Spark.

They've disclosed plenty of information on architecture and the chip itself over the last few months.

But I have a machine here.

Dave had a machine.

He went to the Microsoft event last week.

And we finally had a chance to, on our own, with our own software, with our own tools, install what we wanted, run what we wanted.

And now we can give a accurate real-world picture of where systems based on Spark fall in the market.

So yesterday was the lift for the platform previews, explaining what Spark is and giving an idea of performance, power, things like that.

Today and moving forward, full reviews of retail-ready product are allowed to be published.

We covered the platform yesterday at Hot Hardware and today, the first retail review of the HP Omnibook 16 Ultra, really beautiful premium RTX Spark-based machine.

Before we get into it, I'm going to challenge you just a little bit, Mark, on what you said about speculating.

Because we know the architecture here, we knew Spark was off-the-shelf ARM cores with the CPU complex co-developed with MediaTek, and we know the core accounts on the GPU.

We knew the power limits and the memory bandwidth.

I think it was a little fair to speculate in terms of performance on Spark.

And I will say some of my speculation turned out to be true, right? So there's two RTX Spark N1X chips hitting the market, at least initially.

We know there are many more on the roadmap, not only based on this generation, but Vera and the Feynman generations going out into the future.

Jensen has already spoken about them.

But we have a 20-core...

CPU in the top-end chip with 6,144 CUDA cores, and then an 18-core variant with 5,120 CUDA cores in the GPU.

Those CPU cores are split between efficiency and performance cores.

NVIDIA's goal with this platform, and this really is what I think is the future of lots of mobile platforms and small form factor and special devices, is this unified memory architecture where all of the compute resources can access a large pool of memory.

Now, that won't necessarily help every workload, but it's key for AI workloads because if the larger model you can fit into memory, the better the experience is going to be.

I'm simplifying, but that's the case.

If you have to swap in and out of memory performance tanks.

What we've seen so far in initial data is what we speculated.