Quote from: yee on May 30, 2026, 08:07:27Quote from: opckieran on April 26, 2026, 00:21:56There's NO WAY this performs close to a desktop 5060 with essentially half the power budget.QuoteThe chip will be support up to 128 GB of LPDDR5x-8533 RAM and be manufactured on an unspecified 3 nm node from TSMC.Gaming performance is based on memory bandwidth, let's calculate it:
N1X: 273 GB/s = 256-bit * 8533 MT/s / 8 / 1000.
Strix Halo: 256 GB/s = 256-bit * 8000 MT/s / 8 / 1000.
Basically the N1X falls into the 256-bit APU (e.g. Strix Halo (=4060 Mobile/Laptop performance)) category.Quotethe same as the DGX SparkAnother confirmation. en.wikipedia.org/wiki/Blackwell_(microarchitecture): same GB10 chip.
3dmark.com/search - Steel Nomad:
Strix Halo' Radeon 8060S iGPU: Average score: 2031
RTX 4060 (notebook): Average score: 2262
Memory bandwidth of a 4060 Laptop is 224 - 256 GB/s (aligns perfectly with Strix Halo' gaming performance).
Memory bandwidth of a 5060 (Ti) desktop is 448 GB/s (64% more than N1X).
Memory bandwidth of a 5050/5060/5070 Laptop is 384 GB/s.
(en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units)
Based on the memory bandwidth alone, I'd agree with you. But maybe due to it being based on TSMC' 3nm node variant, a bit more of performance can be squeezed out.
But between "between an RTX 5070 and RTX 5070 Ti laptop", would be very surprising, if N1X' memory bandwidth is at 273 GB/s.
Quote from: Why dream on tho? on Yesterday at 20:46:28Just to make it clear why memory bandwidth is so important:
While the AI / LLM token generation (tg) (the generation of text) depends on the memory bandwidth, the prompt processing (pp) (the processing of input text) depends on the compute of the GPU (basically, the GPU's 3DMark score / FPS performance), but this, again, depends on the memory bandwidth (as you can see in the quote above).
That doesn't mean that a workstation GPU, like a RTX PRO 6000 (96 GB VRAM) (same GB202 chip as a 5090 (32 GB VRAM)), can't have 3 times the VRAM memory, which it has, but it also has 1.8 TB/s of memory bandwidth (thanks to its 512-bit chip and using the, much faster per bit, GDDR7 VRAM chips).
If this N1X or Strix Halo had a bigger variant, using the same 512-bit memory bus width, its memory bandwidth would only be:
546 GB/s = 256-bit * 8533 MT/s / 8 / 1000.
(again, a RTX 5090, using the same 512-bit wide memory bus, has 1.8 TB/s)
This APU, on the other hand, uses LPDDR5X memory, so while it can have up to 128 GB RAM (not VRAM, that uses GDDR6 or GDDR7), its system memory bandwidth is still rather (very) slow at 273 GB/s.
As such, depending on how big your input text size is, 273 GB/s, or a 4060 Laptop of performance that this memory bandwidth allows for, may take a while to process the amount of text.
On reddit.com/r/localllama many users don't seem very happy with their Strix Halo' (=256 GB/s memory bandwidth) prompt processing speed in e.g. agentic workloads.
What you instead may want to build for AI / LLMs is a simple desktop gaming PC with as fast of a GPU and as of VRAM that you want to afford. If AI / LLMs are a priority, look at the workstation cards (RTX PRO 4000, PRO 4500, PRO 5000 and the mentioned PRO 6000, tho they are expensive), because they have double or more VRAM (see given example of a 5090 (32GB VRAM) vs PRO 6000 (96GB VRAM), but it's the same GB202 chip (see en.wikipedia.org/wiki/Blackwell_(microarchitecture))), than the desktop mainstream gaming GPUs using the same GPU chips.
Quote from: yee on May 30, 2026, 08:07:27Gaming performance is based on memory bandwidth
Quote from: yee on May 30, 2026, 08:07:27Basically the N1X falls into the 256-bit APU (e.g. Strix Halo (=4060 Mobile/Laptop performance)) category.
Quote from: opckieran on April 26, 2026, 00:21:56There's NO WAY this performs close to a desktop 5060 with essentially half the power budget.
QuoteThe chip will be support up to 128 GB of LPDDR5x-8533 RAM and be manufactured on an unspecified 3 nm node from TSMC.Gaming performance is based on memory bandwidth, let's calculate it:
Quotethe same as the DGX SparkAnother confirmation. en.wikipedia.org/wiki/Blackwell_(microarchitecture): same GB10 chip.
Quote from: opckieran on April 26, 2026, 00:21:56There's NO WAY this performs close to a desktop 5060 with essentially half the power budget.
Deutsch
English
Español
Français
Italiano
Nederlands
Polski
Português
Русский
Türkçe
Svenska
Chinese
Magyar