It is the same CPU previously used in the overpriced DGX Spark.

It has 10 big Cortex-X925 cores, which are competitive with the Intel P-cores and with the AMD Zen cores, plus 10 small Cortex-A725 cores, which are similar in performance with the older Intel E-cores, from the Meteor Lake, Raptor Lake and Alder Lake generations. The current Intel E-cores are similar to Cortex-X4, i.e. they are much faster.

This Arm based CPU is more powerful that any Arm-based CPU previously used in a non-Apple PC, but in multi-threaded applications it is inferior to AMD Strix Halo CPUs.

The GPU of this is different from that of DGX, which was good only at ML/AI, but poor for graphics.

Here the GPU is likely to be good for graphics, and the top model will have up to 6144 FP32 execution units compared to 2560 of Strix Halo. But I assume that at least the top models will also be much more expensive than Strix Halo.

This NVIDIA CPU+GPU is limited to 128 GB of DRAM, while the successor of Strix Halo, which has been announced recently, offers up to 192 GB of DRAM, so NVIDIA continues its tradition of always providing less memory than its competitors, in order to have better profit margins.

From somewhere in the middle of Nvidia’s endless press waffle:

“The RTX Spark superchip features an NVIDIA Blackwell RTX GPU with 6,144 CUDA cores and fifth-generation Tensor Cores with FP4 precision, connected via the NVIDIA NVLink®-C2C chip-to-chip interconnect to a high-performance, 20-core NVIDIA Grace™ CPU.

MediaTek, a market leader in Arm-based system-on-a-chip designs, collaborated with NVIDIA on the custom CPU design, contributing to its best-in-class power efficiency, performance and connectivity.“

https://nvidianews.nvidia.com/news/nvidia-microsoft-windows-...

Nvidia Grace is an ARM core.

It's an ARM64 CPU with a nVidia GPU rather than garbage Adreno, not bad

As best as I can tell its something like the Apple M series SoC, but for Windows: CPU + GPU with unified memory.

It has 6,144 CUDA cores is similar to a RTX 4070 (5,888) but a lot less than a 4090 (16,384), but what it does have is support for FP4.

When they claim "1 Petaflop AI compute", thats what they mean. For comparison, a RTX 4090 has ~1.3 Petaflops of FP8 processing.

The second big deal is the NVLink-C2C interconnect, which provides up to 900 GB/s of bidirectional bandwidth between GPU and CPU. For comparison, the Apple M4 has 120 GB/s and the M3 Ultra has 819 GB/s. Notably, the Apple M series does not have FP4 support, so this could mean a significant performance improvement over Apple's offerings.