Microsoft reveals new Mac Studio rival with Nvidia RTX Spark and 128 GB RAM

[Redaktion]

Microsoft has revealed a new Surface-branded device featuring the Nvidia RTX Spark chipset. Arriving a day after the Surface Laptop Ultra, the Surface RTX Spark Dev Box is a mini-PC with 128 GB of RAM and more ports than its laptop counterpart.

https://www.notebookcheck.net/Microsoft-reveals-new-Mac-Studio-rival-with-Nvidia-RTX-Spark-and-128-GB-RAM.1313672.0.html

[wrong, not yet a rival]

Asus reveals pro-grade Mac Studio rival mini PC: Nvidia RTX Spark

Not a rival: A NVIDIA Spark is based on a NVIDIA's GB10 chip (273 GB/s) (en.wikipedia.org/wiki/Blackwell_(microarchitecture)#Blackwell_dies), while a Mac Studio is/was based on Apple's M4 Max (up to 546 GB/s) or M3 Ultra (819.3 GB/s) chips.

Unfortunately Apple doesn't sell higher RAM configs anymore:
en.wikipedia.org/wiki/Mac_Studio: "Configurable up to 512 GB until March 4, 2026;[30] 256 GB until May 5, 2026[31]".

If the Spark had 192 GB or 256 GB RAM configs, then, even with the lower GB/s, one might call it a rival, because at least then it would have more RAM than a Mac Studio.

[It is a rival]

That is outdated info.

The maximum you can configure the M4 max is 64 GB and for M3 ultra it's 96 GB now.

So RTX spark is fine.

You finished trolling?

[Me myself and I]

Yeah but it doesn't have the neural processing units that you have on an M5 Max I have 128 GB at 1 and I will be purchasing at 256 Ultra as soon as it comes out if they make the 512 I'll be getting that one.

[Michael Davidson]

I own a Macbook Pro M5 Max with 128GB of ram and a DGX Spark (GB10) with 128GB of ram. The Macbook Pro is leagues faster, almost 3x the memory bandwidth. The GB10's saving grace is of course CUDA and the ability to run NVFP4 models. Beyond that, this chipset is already a year old, this WILL NOT outperform a similarly spec'ed Macbook Pro and most definitely not a Mac Studio.

[wrong, not yet a rival]

The GB10's saving grace is of course CUDA and the ability to run NVFP4 models.

Problem is, even after a year nobody NVFP4 models are basically unavailable.

Maybe the bits per weight (BPW) are simply not enough after all, so nobody even bothers training them in native NVFP4?
Q4_K_M is the most popular quant when it comes to size vs performance and it is 5.24 BPW, after that, the performance drops off sharply. Gpt-Oss-120B is such a natively trained (QAT) (not on NVFP4 tho) 4-bit model and it constantly makes so many mistakes that I use it less and less, the BPW may be simply not enough: Forget 4-bit for now and move to 4.5 or 5.24-bits in the meantime. NVIDIA needs to introduce NVFP5 for the Blackwell successor XD, haha.

it's 96 GB now.

Correct and 819.3 GB/s (M3 Ultra) at 96 GB RAM is still better and Spark not a rival.

Total memory size x speed value:
M3 Ultra: 78653 = 819.3 GB/s * 96 GB
GB10/Spark: 34944 = 273 GB/s * 128 GB

But I agree, offering only 96 GB is crazy low, and a waste of the huge 1024 bits APU potential (exactly like giving a RTX 5090 6 GB of VRAM (5.33 times lower: 512 GB -> 96 GB = 32 GB -> 6 GB), so low in fact, that it's going to push some users to get AMD Strix Halo or NVIDIA Spark, depending on their use case, especially considering the price.

Problem is tho, 96 GB @ 819.3 GB/s can fit and run dense models that outperform MoE models on the slower 128 GB RAM ;) and compensate hugely for the fact that it's not 128 GB RAM. If Qwen releases a dense e.g. 50B to 90B model, then the 819.3 GB/s start to shine even more. But we don't have to wait: Qwen3.6-27B dense is a good example.