NVIDIA showed a 4-node DGX Spark ring with no switch at SC26. I cabled mine the same way and compared it with my 200G switch.
ASUS GX10 (DGX Spark) 128GB ×4, DeepSeek V4.1 Flash, original weights, same recipe, same bench
Ring vs 200G switch
128K prompt read: 8,823 vs 8,722 tok/s
980K prompt read: 6,173 vs 6,101 tok/s
Concurrency 16 (code): 1,006 vs 999 tok/s
Concurrency 1 (code): 126 vs 123 tok/s
The ring comes out about 1–2% ahead. I don't know why. I only ran the ring 2–3 times, so I wouldn't read this as "the ring is faster".
NVIDIA's own numbers have the switch about 10% ahead at 4 nodes (Qwen3.8-27B, 32.82 vs 36.34 tok/s), and I'd trust their measurement over mine.
Either way the gap is small. For 4 nodes running one model, I don't think you need a switch. 4 cables are enough.
The ring needs a normal LAN on every node, and all 4 run as one group only.
Also tried on the switch
MTU 9000: link 185 → 196 Gb/s, serving speed the same
PFC (lossless): no change
RouterOS 7.20.8 → 7.24.6: packet loss on long prompts is gone
spark-ds41f @christopherowen · switch: MikroTik CRS812
Pushing Qwen Next Flash on my Mac Studio M5 Ultra (256GB) 🚀
🔥 17 concurrent requests (16 actively decoding)
⚡ 453 tokens/sec median aggregate decode
📈 ~438 tokens/sec sustained right now
That's roughly 27–28 tok/s per request, with all 16 decoding in parallel.
And the interesting part? This isn't a synthetic benchmark. It's a real workload, running continuously.
Apple Silicon + MoE models are a pretty fascinating combination. 🍎
Why I returned the Mac Studio M5 Ultra 512GB, since many of you asked:
I never run just one LLM session at a time. There are always several going at once, and that's where it didn't fit.
DeepSeek V4.1 Flash, same bench, vs four GX10s at about the same price:
- One session: close, 53 vs 73 tok/s
- 16 at once: 137–148 vs 330–847 tok/s total
- Long prompts: the GX10s read them 2.3–3.1x faster
The Mac did win on idle power (~5 W vs 215 W) and cold start (36 s vs ~150 s).
If you run one session at a time, it's a great machine.
Returned my Mac Studio M5 Ultra 512GB today.
A week of tests: LLM serving, video generation, power, cold start.
It's a great machine, just not the right fit for what I do.
Even if a Mac server ships, I'm not sure it'd fit either.
Maybe the Mac server is just a fantasy 😅
I’ve been thinking about this for a while, and I’ve seen some community recipes too. Now that NVIDIA is officially supporting it, I should actually try it.
A supercomputing lab on your desk. ⚡
New at #SC26: the NVIDIA DGX Spark HPC Developer Kit, built for researchers, academics and developers.
🟢 Connect up to 4x DGX Spark systems in a ring over NVIDIA ConnectX-7
🟢 No network switch needed
🟢 Up to 512GB of memory
Returned my Mac Studio M5 Ultra 512GB today.
A week of tests: LLM serving, video generation, power, cold start.
It's a great machine, just not the right fit for what I do.
Even if a Mac server ships, I'm not sure it'd fit either.
Maybe the Mac server is just a fantasy 😅
Follow-up: how long to cold start?
Mac Studio (M5 Ultra) 512GB ×1 vs ASUS GX10 (DGX Spark) 128GB ×4
DeepSeek V4.1 Flash, original weights, file cache emptied first
Start to first token (2 runs)
Mac: 36 s
GX10 ×4: 150-153 s
Where the time goes
Mac: almost all of it is reading 294 GB from one 8TB SSD at ~8 GB/s
GX10 ×4: each unit reads 156 GB from its 1TB SSD at ~4.6 GB/s, done in ~35 s. The other ~115 s is container start, cluster init and engine setup.
Four drives together read faster (~18 GB/s), but the Mac is ready 4× sooner.
SSD sizes differ (8TB vs 1TB each), and bigger SSDs are usually faster, so other configs may read at different speeds.
Read speeds are averages during load, not drive limits. No reboot, so OS boot time is not included.
LLM speed, video generation, power draw.
I've run out of things I want to test on the Mac Studio.
I can't find a reason to keep it.
Is there anything I should try before I decide?
Mac Studio 512GB ×1 vs ASUS GX10 128GB ×4
DeepSeek V4.1 Flash, original weights, same bench
Price: $25,799 vs $24,000
Concurrency 1: 53/84 vs 73/122 tok/s (text/code)
Concurrency 16: 137/148 vs 330/847 tok/s (text/code)
128K prefill: 3,325 vs 7,522 tok/s
DGX Spark recipes keep coming and keep getting faster. That's a big plus.
Mac's plus: one person, one session at a time? It's also your dev machine. Code and run the LLM on the same box, while the Sparks still need another computer to work from.
Mac: mlx-serve @ddalcu · GX10: spark-ds41f @christopherowen
Follow-up: how much power?
Mac Studio (M5 Ultra) 512GB ×1 vs ASUS GX10 (DGX Spark) 128GB ×4
DeepSeek V4.1 Flash, same bench, measured at the wall
Power
Idle: ~5 W vs 215 W
Concurrency 1 (text): 187 W vs 482 W
Concurrency 16 (code): 256 W vs 541 W
Energy per 1M tokens
Concurrency 1 (text): 0.96 vs 1.81 kWh → Mac uses half
Concurrency 16 (text): 0.49 vs 0.45 kWh → about equal
Concurrency 16 (code): 0.48 vs 0.16 kWh → GX10 uses a third
GX10 ×4 is faster, but pulls ~500 W even for one request.
Solo use favors the Mac. A busy server favors the GX10s.
GX10 numbers are all 4 units combined. Switch not included.
Thank you all for your comments and experiences on MLX vs CUDA.
For now, the @nvidia GB10/GPU stacking will continue instead!
As much as I wanted this Mac, our training scripts are all built with CUDA, and while we could get training running on MLX, local AI is accelerating too fast for us to pivot to MLX and push the stack from scratch. Would be a fun project, but I am unable to afford stacking both the Mac and GB10s at this time 🤣
Would love to get further into MLX in the future though, and to those of you wanting more MLX quants of the Qwopus models, I hear you, and we can start shipping those more often too! (They can be made on the GB-10 units as well)
I know it’s unlikely, but if you have a GB10 laying around you’re not using, I would be interested in buying. For now, I’ll just pay scalper price and consider it a dollar cost average 🤣
Alright guys the Qwopus team needs advice, I’m torn! Mac or more GB10’s?
I have an M5 Ultra 512 on preorder!
Primarily thinking the bandwidth and large memory pool would really help us with internal dataset generation with much larger models. But also interested in getting into the MLX training workflows. It theoretically will be awesome for RL in particular, but I haven’t messed with ecosystem a lot yet. I’m sure we could figure it out promptly.
Our goal is always to enhance the local AI experience people can have with actually affordable hardware. We have a lot of solid working theories on domain specific distillation that are proving to be very interesting in testing.
BUT we’ve done most of the work to this point on a single Nvidia GB10 unit and the CUDA training environments we have are really excellent.
Despite them being higher cost now, I could pick up two or maybe three more GB10 units and have the larger memory pool, plus parallel compute.
I know a lot of people are saying the Mac is overpriced, but right now for the bandwidth and the memory it’s still one of the best options. We could also use it primarily for dataset work, and continue training in the GB10 we have.
Please let me know your thoughts, this is an important decision, and completely self funded by me lol ultimately just want to make the best long term decision.
Regardless, we’re so excited to keep shipping and expanding our lab. Things are going to get even crazier for local AI in the coming months!
Voices I’d love advice from:
@TheAhmadOsman, @QwenDevs, @0xSero, @MiaAI_lab, @coldniko
MiniMax H3 video generation: Mac Studio 512GB vs GX10 (DGX Spark) vs RTX 5090 vs RTX PRO 6000
Compared with the Mac Studio, on a 5 s clip at 1344×768:
• RTX PRO 6000 ×4: 6.3× faster
• RTX PRO 6000 ×1: 2.1× faster
• GX10 ×4: 1.9× faster
• RTX 5090: 1.8× faster
• GX10 ×1: 1.5× slower
Four GX10s working on one clip together are about as fast as a single RTX 5090.
Same original BF16 weights, prompts and seeds on every machine. No quantization.
Earlier this week, our cloud service saw abnormal traffic that surged roughly 34-fold in just one hour, peaking at roughly 20 million requests per day and was still growing linearly at a steep rate.
At that pace, we were only hours away from the service becoming unavailable, putting everyday printing for millions of users at risk.
To keep the service up, we temporarily blocked the anomalous traffic after an urgent review.
Some users were affected while we adjusted the policy. The service is stable now.
To learn what happened and how the problem was solved, please check our post on the Bambu Lab Forum.
forum.bambulab.com/t/a-postm…