Nvidia RTX Spark Analysis - Finally an answer to Apple's M5 Pro & M5 Max?

[Redaktion]

RTX Spark is Nvidia's first SoC and combines a custom ARM-CPU with a Blackwell GPU and up to 128 GB RAM. The new chips are designed for AI workloads, content creators and you can play games as well. We testes the new RTX Spark SoC and compare it with Apple, AMD, Intel & Qualcomm.

https://www.notebookcheck.net/Nvidia-RTX-Spark-Analysis-Finally-an-answer-to-Apple-s-M5-Pro-M5-Max.1422820.0.html

[rutto]

the Photoshop benchmark is the only one showing the N1X, otherwise it doesn't appear in any other table. lol, I guess

[dada_dave]

the Photoshop benchmark is the only one showing the N1X, otherwise it doesn't appear in any other table. lol, I guess

Yeah which is not formatted the same as their normal charts ... but as a picture. As for the others, they must've grabbed the wrong data tables for the article or something?

[Andreas Osthoff]

We are dealing with a small technical issue and we are working on it now, sorry for the delay!

Edit: Problem is fixed!

[dada_dave]

We are dealing with a small technical issue and we are working on it now, sorry for the delay!

Edit: Problem is fixed!

Great! Now I get to look forward to the full review. :)

BTW I like the power analyses you do at NotebookCheck so much, I actually re-visualize them as interactive x-y bubble charts. I think my account is too new to do a link, I was previously just posting as a guest, but I host the graphs online at powervsperformance dot tiiny dot site.

[Linux?]

Can Linux be installed? I read the bootloader is locked?

[ye idk]

Here, the GPU performance is comparable to powerful versions of the Nvidia GeForce RTX 5070

Clicking on "Steel Nomad Score", Spark only 12% faster than the 5060 Laptop (Legion 7 16 AGP11) and 4-9% faster than the Radeon 8060S in "Cyberpunk 2077 / Ultra Preset (FSR off)", which checks out considering the ~300 GB/s (could be lower in this particular device, depending on the LPDDR5X MT/s used) (vs Strix Halo's 256 GB/s) memory speed.

Spark N1X: 301 GB/s = 256-bit * 9400 MT/s (assumed) / 1000 / 8.

Local AI powerhouse. [..] The new chips are designed for AI workloads, [..]

PS: For AI, AMD's Gorgon Halo, a Strix Halo rename/refresh with software-patched support from up to 128 to up to 192 GB unified memory/RAM, at a similar bandwidth, is an alternative (NVIDIA's CUDA is not required for end-user inferencing tasks (maybe there are special cases, I don't know)); if you want to fine-tune or even train AI models (which would may be slow on these ~300 GB/s )), CUDA is the standard, but it doesn't mean that it can't be done on AMD hardware) for those who require more than 128 GB unified memory/RAM.
Gorgon Halo: 273 GB/s = 256-bit * 8533 MT/s / 1000 / 8.

I wish these (256-bit) APUs had double the unified memory (to 256 GB) and double to quadrupled the memory speed, to be really interesting. Of course, price is another issue too, but first, the specs need to double (and no, I'm not buying 2 of these).

AI bang for buck performance alternative
For more AI bang for buck performance and upgradable parts, get a consumer desktop PC 96 to 128 GB RAM (a consumer desktop PC uses 128-bit memory interface (aka dual-channel) and the RAM's MT/s is typically between 5600-6200 MT/s, which equals to ~96 GB/s (vs Spark's ~300 GB/s), but the GPU(s) will make up for it (see calculation below) (32 GB VRAM at ~1800 GB/s, when running 2 5070 Ti (each at ~900 GB/s) in tensor parallel) + 1 GPU (at lest like a 5070 Ti (16 GB VRAM at 900 GB/s, it will have much faster prompt processing and you will have 128+16 GB of total memory) (a 5080 24 GB VRAM is rumored to be released in 2027 Q1)) or may get a motherboard that supports bifurcation (geizhals.eu/?cat=mainboards&xf=2833_AMD+AM5~2873_ab+2&pagesize=30&promode=true) and get 2 GPUs of the same GPU architecture and run them in tensor parallel mode, which will double the VRAM memory in size and speed (reddit.com/r/LocalLLaMa/search/?q=tensor parallel).

Let's calculate our arbitrary AI score (memory size times memory speed):
N1X Spark: 38400 = 128 GB unified memory * 300 GB/s.
Desktop PC build using two 5070 Ti and running them in tensor parallel: 69888 = 128 GB * 96 GB/s + 32 GB VRAM * 1800 GB/s.

[ye idk]

The AI models that are best suited for these ~300 GB/s APUs (Strix/Gorgon Halo, Spark N1X, APPLE's M Pro) are of MoE (as opposed to dense) architecture, see e.g.:

artificialanalysis.ai/?models=qwen3-5-122b-a10b-non-reasoning%2Cglm-5-3-flash%2Cqwen3-5-122b-a10b%2Cmimo-v2-5-0424%2Cnemotron-3-5-lightning%2Cqwen3-6-27b-non-reasoning%2Cminimax-m2-7%2Cmuse-glimmer%2Cnvidia-nemotron-3-super-120b-a12b%2Cqwen3-8-27b-medium%2Cqwen3-6-35b-a3b-non-reasoning%2Cstep-3-7-flash%2Cinkling-small%2Csolar-open2-250b%2Cqwen3-6-35b-a3b%2Cqwen3-6-27b%2Cqwen3-8-flash-next%2Cqwen3-8-27b%2Cmotif-3%2Chy3%2Cg9v3-39a5b%2Cgemma-4-31b%2Cgemma-4-31b-non-reasoning%2Cmistral-medium-3-5%2Cling-3-0-flash%2Cgpt-oss-120b%2Cdeepseek-v4-flash%2Cnemotron-3-5-bolt-74b-a9b%2Cnemotron-3-5-super-vl&intelligence=artificial-analysis-intelligence-index

Qwen3.8-Flash-Next is basically the winner overall (highest performance per size) and has only 6B active parameters, perfect for ~300 GB/s/Spark N1X devices.

NVIDIA also offers specific NVFP4 builds of Qwen3.8-Flash-Next for Spark specifically, since it also supports NVFP4:
huggingface.co/nvidia/Qwen3.8-Flash-Next-NVFP4 (133 GB but the n-gram may be streamed from the SSD (-> 133GB - 30% (streamed from a SSD) = 93 GB -> fits in a 128 GB unified memory Spark config)
huggingface.co/nvidia/Qwen3.8-Flash-Next-FP4-Local (this one is new and should definitely fit in a 128 GB unified memory config)

[James Parker]

Can Linux be installed? I read the bootloader is locked?

I second the question.

[Pierce Wiederecht]

How does a 20 core ARM based chip come behind 16 core Strix Halo in multicore performance per watt when Zen5 is a full two years old now and on it's way out? Not to mention the N1X is on 3nm while Strix Halo in on 4nm... That just doesn't make any sense. Zen6 is going to be a huge jump over the N1X considering ots already beating Nvidia's custom Vera chips and the N1X is using an off the shelf core that's significantly weaker than Vera.

[dada_dave]

How does a 20 core ARM based chip come behind 16 core Strix Halo in multicore performance per watt when Zen5 is a full two years old now and on it's way out? Not to mention the N1X is on 3nm while Strix Halo in on 4nm... That just doesn't make any sense. Zen6 is going to be a huge jump over the N1X considering ots already beating Nvidia's custom Vera chips and the N1X is using an off the shelf core that's significantly weaker than Vera.

Part of the answer is that these ARM cores are pretty old themselves! Nvidia released this same chip in Linux workstation form last year and other chips with these cores started appearing a year before that. They may be fabbed on a newer node than the AMD chips (though we aren't sure which 3nm node it is for the Nvidia chips) but the core design is from roughly the same era even though these devices are new.

The other major part is that 20 vs 16 cores is slightly misleading. The Nvidia device has 10 P-cores and 10 M-cores while the AMD chip has 16 P-cores + SMT2. SMT2 should give the AMD chip a roughly 25% additional boost in MT workloads (highly variable).

These are the two main reasons why this "new" CPU is so lackluster compared to its x86 PC counterparts, merely competitive with them when a priori it should be beating them.

[Seems possible]

Can Linux be installed? I read the bootloader is locked?

I second the question.

According to Wendell of Level1tech, he has hinted that he will definitely be delving into Linux on these things and to watch for future  upcoming videos on his Linux channel (Level1linux).