FWIW, Newegg is now listing the GIGABYTE 64GB version at $4,999.
At that price, I’m struggling to see the value proposition. For running larger, more capable LLMs, 256GB is a much more practical starting point.
Getting to 256GB with these systems would require four units, at a total cost of $20,000. By comparison, a Mac Studio configured with 256GB of unified memory costs just $10,000.
Granted, the NVIDIA platform has advantages in CUDA support and the broader AI software ecosystem, but is that enough to justify twice the price for the same aggregate memory capacity?
I really think the 64GB models will be a tough sell at this price point. What am I missing?
Gufo experience is first hand, I have Strix for far longer than Sparks and it changed dynamics recently (use to be a toy to run Qwen 3.5122B, not it’s serious) , I run my backup llm on it.
Even more than this I was able to make gufo work in WSL with native GPU passthrough, meaning that I play steam games and when I need run LLM no reboots (performance is same as linux), that’s ROCm though. Unfortunately I can’t play and run llm same time with gufo. With Vulkan backend in the past with llama + qwen 122B I was able to. Playing Diablo 2 Resurected all on Max same time with active inference (2x decode drop) was magical :)
With the 64GB Spark you are paying a large proportion of the total cost for the NIC, you really need to have some very specific use cases for that to make sense.
A RTX Spark (running linux) with 64GB makes more sense to me: not necessarily for inference but rather as a box on which to develop and test CUDA applications which demand more RAM than most GPUs provide (and where the slower VRAM bandwidth is not a blocker). Also as a box on which to run Blender, DaVince Resolve, etc and you do not want to use MacOS.
Imho - only one reason to buy 64GB Spark is try t2image production pipeline and you’re really newbie in all ai stuff so build own ai server or take AMD is pain for you.
It can fit normally flux2, etc. But no i2v pipelines.
And price is crazy. AMD AI Max 395+ cost same for 128 and 95+% of comfy workflows work on it too. And you can try video pipelines too.
I have 2 sparks for 128. My wife need ai computer for image/video processing and Spark 64gb vs x4 5060ti 16gb or 2x 4080 32GB from China vs Ryzen Ai MAX 128gb looks not worthy.
You all understand correctly. I can’t imagine what real thing you could even run there? That’s catastrophically little, even for Qwen 3.8 27B (I’m exaggerating a bit).
Essentially, you said it right — a Mac machine (even a Mac mini with 64GB) would be a better choice than this underpowered Spark.
But if NVIDIA is looking to the future and soon 128GB for AI will simply cost an astronomical price, then perhaps it will find its buyer.
Ouch $8,499 before tax and shipping for128GB and $4,999 for 64GB is a long way from $3,999 for 128GB just a year ago.
With these prices the M5 Ultra 512GB looks like a better deal than stacking four $8,499 Sparks for $33,996 or eight $4,999 for $39,992 unless your workflows are CUDA dependent.
Double Comfy. That’s my second use for my Strix besides backup LLM.
Use to be bad, because Comfy had bad habit on machine with unified memory and carved 96Gb of VRAM load models into remaining 32Gb - system ram and crashing. 2 nights with GLM Flash fixed this nasty behavior for all models I use. So now it’s stable.
Downside though, it’s about 2x slower than Spark from what I can tell, hope entuziasts will fix that for new models eventually :)
With over 4x compute performance of the M5U from the Sparks, and given that CUDA is lightyears ahead of Apple, the Sparks’ prefill performance and high concurrency are a clear winner for my simultaneous multiple agent and subagent workloads.
The Spark price hike is more than a response to RAM price increases we saw that move from $3,999 to $4,650, GB10s got repriced with the last price hike to $6,950. It will be interesting to see how the price continues to move with this weeks market FUD.
The problem is that prices are rising catastrophically, and here’s what we actually have in Europe right now (I suspect prices will reach this level in the US within a couple of weeks)
For example, two weeks ago I bought an MSI 4T for €6,200 — now it costs €9,500 / apparently it’s time to wait for a new solution and alternatives.
Ok that is wild! Micro Center has spoiled me. I had no idea prices were still skyrocketing elsewhere. Sounds like my math will need adjustments before too long unfortunately.
As i read - Strix on linux can offer more than 96gb as “vram”.
Why it x2 times slower? As we count spark as a bit slow 4070, so Strix would be slower than 5060 ti?
