GPTware

🇨🇦 Obsessed with all things Ai
Agents, Loops, LLMs, Fine-Tuning

- Entrepreneur. Builder.

[GPTware @GPTWare]

DeepSeek-V4.1-Flash-1017 on 1× RTX PRO 6000 Blackwell + 2× GB10, using QMesh AFD-TP3 + DSpark:

Full NVFP4 Quant
⚡ 119.78 tok/s peak single-request (code)
⚡ 200.24 tok/s peak at 8 concurrent requests

2M context window.

[GPTware @GPTWare]

@mov_axbx Working on some variants of this.
https://t.co/LaJZg77Otw

[GPTware @GPTWare]

DeepSeek-V4.1-Flash-1017 on 1× RTX PRO 6000 Blackwell + 2× GB10, using QMesh AFD-TP3 + DSpark:

Full NVFP4 Quant
⚡ 119.78 tok/s peak single-request (code)
⚡ 200.24 tok/s peak at 8 concurrent requests

2M context window.

[GPTware @GPTWare]

Repo: https://t.co/vmdZxfFjNl

[GPTware @GPTWare]

DeepSeek-V4.1-Flash-1017 on 1× RTX PRO 6000 Blackwell + 2× GB10, using QMesh AFD-TP3 + DSpark:

Full NVFP4 Quant
⚡ 119.78 tok/s peak single-request (code)
⚡ 200.24 tok/s peak at 8 concurrent requests

2M context window.

[GPTware @GPTWare]

@CTOAdvisor 2 @ 256GB >> 512GB
Just ask Apple.
https://t.co/rOrS0Bfms0

[GPTware @GPTWare]

I said i'd work for 1/2 the price and that i'd bring my own kneepads.

[Roy @usr_bin_roygbiv]

There is someone on applystead right now with 3% response rate and 70 invites

[GPTware @GPTWare]

If I owned this, i'd probably still think I didn't have enough compute. 🤪

[Wall St Engine @wallstengine]

A 300-rack AI supercomputer powered by 9,600 Nvidia A100 GPUs is listed on eBay for $59.7M.

That’s 1,200 GPU compute servers with 384TB of combined GPU memory and roughly 1.4 petabytes of system RAM.

Each compute server connects eight A100 GPUs through $NVDA's NVSwitch technology.

More than six years after A100 launched, The seller is offering these systems refurbished for new AI training and inference deployments.

[GPTware @GPTWare]

@ivanfioravanti Or just buy 4 @ M5 Ultra then sell 2 in 9 months when the price doubles again to pay for the 2 you want to keep.

[GPTware @GPTWare]

@jayleaton 2 @ 256GB is way better than 3 @ 96GB - 224GB more RAM for $1200 more.

[GPTware @GPTWare]

When your RTX Pro 6000 is the shitty end of your compute. 😂

[James Meadlock @JamesMeadlock]

Deepseek-V4.1-Flash for 1 x DGX Station GB300 + RTX Pro 6000

https://t.co/H4OfMzR5Z2

[GPTware @GPTWare]

@sakurayukiai @jayleaton https://t.co/rOrS0Bfms0

[GPTware @GPTWare]

@XadenRyan They can play nicely together.
https://t.co/z5EjPizuQK
= Check out TensorFold.

Moving forward there will be more of this. I'm working on RTX + Sparks...
https://t.co/r6AnSz5sOl

[GPTware @GPTWare]

DeepSeek-V4.1-Flash-1017 on 1× RTX PRO 6000 Blackwell + 2× GB10, using QMesh AFD-TP3 + DSpark:

Full NVFP4 Quant
⚡ 119.78 tok/s peak single-request (code)
⚡ 200.24 tok/s peak at 8 concurrent requests

2M context window.

[GPTware @GPTWare]

@__tinygrad__ Wow! So no room for my PCIe coffee warmer???

[GPTware @GPTWare]

@cornokun Yep... you should have bought 4! 🤪
If you can afford it buy 2 more and sell those in September 2027 to pay for most of or all of the first 2.

[GPTware @GPTWare]

@JoesInvestments RTX, Sparks, M5U they all have their places and increasingly will work well together.

[GPTware @GPTWare]

@softpoo I never said it scaled linearly

[GPTware @GPTWare]

@someRandomDev5 @Larssteenhoff Can be minimized to some extent but good point.

[GPTware @GPTWare]

At this cost to upgrade, why would anyone buy a 512GB M5 Ultra instead of 2 @ 256GB ??? You get 2 full processors and get an approx. 1.8 speedup with TP=2 double everything. That upgrade price is insane.

[GPTware @GPTWare]

@graykevinb @SJRonanMD 1228 vs 1792 for M5U vs RTX6000. Prefill blows though. Mac needs a lot of work.

[GPTware @GPTWare]

For those pointing out interconnect etc. etc. Yes good points, but see:
https://t.co/ICgQnAH4tr

and from Apple themselves:
https://t.co/rOrS0Bfms0
(thank you to @someRandomDev5 for this link)

[GPTware @GPTWare]

At this cost to upgrade, why would anyone buy a 512GB M5 Ultra instead of 2 @ 256GB ??? You get 2 full processors and get an approx. 1.8 speedup with TP=2 double everything. That upgrade price is insane.

[GPTware @GPTWare]

@CloudyBRM From Apple themselves: https://t.co/rOrS0Bfms0

Also see: https://t.co/ICgQnAH4tr

[GPTware @GPTWare]

@josusanmartin @chargebite_ I hear what you're saying, but i'm betting they are just increasing the density of their pricing 😂. There was no mention of that from the M3U.