Count Stacula

Local AI Enthusiast

6x Asus Ascent GB 10s Clustered with a Mikrotik 812 Switch
1x Nvidia DGX Spark
1x Mac M3 Ultra 512 GB
2x 3090s
And growing...

[Count Stacula @CountStaculaAI]

@mr_r0b0t Fixed, try now! Sorry about that, thanks!

[Count Stacula @CountStaculaAI]

I haven't ran it yet but plan to do so within the next few days but from what I've seen so far, DeepSeek V4 Flash 0731 is right up there with GLM 5.2. I've got that running on my 6x cluster of GB10s but with the 2x RTX 6000s, I think you can pump out some serious tokens with minimal degradation in quality from what I'm seeing. Looking into getting some RTX 6000s soon for my own testing, just going off what I've seen/read so far.

[Count Stacula @CountStaculaAI]

"A real 502,639-token request served end-to-end: 18.3 min prefill, 66 tok/s decode at 500k depth."

Is this not really slow prefill for the 2x RTX 6000 + DeepSeek V4 Flash 0731, even at 500k token context window that seems long for this model and hardware? Thinking about buying a 2x setup, so I'm curious on prefill speeds.

[Count Stacula @CountStaculaAI]

@mr_r0b0t Look at the left side, its actually 6x of them on a 812 Mikrotik switch, I messed up and bought the shorter cables which completely messed up my placement/setup, but too lazy to deal with buying new cables and redoing it all for now.

[Count Stacula @CountStaculaAI]

@MiaAI_lab And whatever you do, do not skimp on the length of the connection cables. Learned this one the hard way with no room for proper placement.

[Count Stacula @CountStaculaAI]

@loktar00 @TeksEdge Wow, ya, that's an incredible deal with those discounts. I'm jealous, I'd buy multiple at that price. Thank you for your service!

[Count Stacula @CountStaculaAI]

Here's some things I'm sure you've not considered:
1) The DGX Spark will likely appreciate or at the very least retain its value over the next 2 years, maybe more.
2) Cost of electricity at full throttle all day everyday is < $20 per month for me, per spark.
3) Full privacy: Big labs can't steal your data, use cases, etc
4) Currently the labs you speak of are subsidizing their API/Subscription pricing with investment. If Investment dries up or when it comes time for them to focus on profitability, the cost of API/Subscription rises and possibly drastically.

If you can afford it, in my opinion it's a no brainer, but to each their own.

[Count Stacula @CountStaculaAI]

I reached out to customer support, the first person said there was nothing they could do but offer a return + refund. After going back and forth with them for 10 minutes, I asked to be transferred to their manager who basically gave me the same answer and no alternative solution, however, they said they would send me an email where I could upload proof and that customer service would then reach out to me via email. The email I received to provide the proof is attached below, which is incorrect for my issue, so rather than reach out again and waste my time, I just gave up and figured I'd take it to X to warn others. Their customer support has definitely gone down hill recently, which is disappointing.

[Count Stacula @CountStaculaAI]

PSA to all DGX Spark Buyers:
If you buy "New" from Amazon, there's a 2 in 7 chance (from my experience) that you will receive an open box returned machine.

I purchased 7 New:
1x Nvidia DGX Spark
6x Asus Ascent GB10's

What I got was:
1 New Nvidia DGX Spark
4 New GB10s
2 Open Box, Used GB10s

When I reached out to Amazon customer support, they did allow me to return them, but I need these for a new cluster, so I don't have time to waste dealing with that and then risk getting more open boxes or even my own returns sent back to me, so I guess I'm keeping them.

If receiving an open box computer is of concern to you, beware purchasing from Amazon, as they clearly consider an open box, returned computer as "New".
Here's 1 of the shipments showing me opening the Amazon box and receiving 1 open box Asus Ascent GB10.

@amazon @AmazonHelp @ASUS @nvidia

[Count Stacula @CountStaculaAI]

@kkiran "Ships from and sold by Amazon."

Here's the specific link where I made the purchases:
https://t.co/eyRVl9pkO3

Thanks!

[Count Stacula @CountStaculaAI]

After leaving Twitter many years ago, I have returned to X, with a new focus: Local AI. I have wiped all of my old posts and follows and will be starting fresh.