Alright, let's make it clear:
QuoteNvidia has not shown off any performance metrics of the RTX Spark
Very sus, isn't it? But no problem, the gaming / AI performance talk for this N1X / NV Spark is here: notebookchat.com/index.php?topic=310661.0.
In short,
the memory bandwidth determines everything: The prompt processing (pp) (because that's 3D-performance / FPS-based and that, again, is bandwidth-based) and token generation (tg) speeds:
- AMD Strix Halo: 256 GB/s = 256-bit * 8000 MT/s / 8 / 1000.
- NVIDIA Spark (GB10 chip): 273 GB/s = 256-bit * 8533 MT/s / 8 / 1000.
(en.wikipedia.org/wiki/Blackwell_(microarchitecture)#Blackwell_dies)
3dmark.com/search - Steel Nomad:
- AMD Strix Halo' Radeon 8060S iGPU: 256 GB/s: "Average score: 2031".
- NVIDIA RTX 4060 (notebook): 224 - 256 GB/s: "Average score: 2262".
- NVIDIA RTX Spark: 6.6% faster than Strix Halo, so.., but let's wait for official, or better, independent benchmarks. The 3 nm TSMC node + 6,144 CUDA cores might show, to a certain degree, that this Spark is faster than what 273 GB/s would suggest. Let's say it's 15-30% faster than that, that's 5060 Laptop ("Average score: 2632") to 5070 Laptop ("Average score: 3005") performance, not bad and it aligns with performance rumors.
If it's 5070 Laptop performance, then it's weird that NV wouldn't talk about it, because that's pretty good. Of course, price determines Spark's success. For proper AI performance, building a simple desktop PC may still give the most AI performance bang for buck.
Quote from: Nv on Today at 12:16:31Nvidia just killed desktop Snapdragon CPUs. Arm architecture with proper GPU performance and support.
IDK, several posts in a span of almost 24h apart in reddit/r/hardware have 0 upvotes, as also the comments suggest.
The Snapdragon X Elite / X2 Elite are 128-bit wide memory bus APUs, while the Spark is a 256-bit wide one, totally different animals and price-points.
Deutsch
English
Español
Français
Italiano
Nederlands
Polski
Português
Русский
Türkçe
Svenska
Chinese
Magyar