getting faster

[A J]

Hi All,

This doesn't directly affect the small and medium size Bots but it is a big jump in speed and endurance.

   Right now some B200 GPUs rent for about $5 - $20/hr and have memory bandwidth less  than 5 TB/s. The current generation Blackwell has memory with 4 x 2 TB/s. The next generation  could be online in some with 8 x ? TB/s but rumors put it at 20 TB/s. Later versions or generations could go to 12 x ? TB/s or 40 TB/s with 3.3 TB/s stacks (by 2030).

   Hopefully this would make the B200 GPU popular for free and subscription services used  by Bot designers.

   For electric car batteries they seem to be on track to reduce the weight of large lithium systems by 50%. This might not track for batteries used by Bi-peds and Quad-peds but given the interest in these Bot there would be motivation to innovate. 

  If things get faster will it break some kind of thresh-hold by 2030? At the small Bot level changes in RAM or Batteries would make a difference?

AI Overview

Yes, Samsung plans to have HBM memory faster than

2TB/s by 2030, with their next-generation HBM4 technology targeting bandwidth of up to

3.3TB/s. They have already achieved 1.3

TB/s with their HBM3E technology and are securing deals with major partners like NVIDIA to use their HBM4. 

p.s. This does come at a very high cost in power.

https://www.techradar.com/pro/ai-gpu-accelerators-with-6tb-hbm-memory-could-appear-by-2035-as-ai-gpu-die-sizes-set-to-shrink-but-theres-far-worse-coming-up

AI Overview

Yes,

significant breakthroughs in lithium battery technology, primarily the development of solid-state batteries and lithium-sulfur batteries, are on track for commercialization and mass production around or before 2030. These advancements promise lighter and longer-lasting car batteries. 

Key Breakthrough Technologies

• Solid-State Batteries (SSBs): Regarded as the "ultimate solution" to current EV battery limitations, SSBs replace the flammable liquid electrolyte with a solid one, offering higher energy density, increased safety, faster charging, and a longer lifespan.
  • Commercialization Timeline: Small-scale production and installations in premium EV models are expected between 2027 and 2028, with wider mass-market adoption projected by the early to mid-2030s.
  • Performance: Expected to provide ranges of 600 miles (965 km) or more and reduce battery weight by up to 50%. Energy density targets exceed 400 Wh/kg, with some projections as high as 600-800 Wh/kg by 2030.

[Chris Albertson]

A few minutes after this email, I got a notice from Nvidia. They are offering an “enthusiast’s” level computer for AI. Basic specs: 1 petaflop FP. That means one trillion 32-bit floating operations per second, and 128 GB RAM. It fits in a small-size box. $4K.

What is impressive is that this is FLOPS, not TOPS. The computer is suitable for training new models, not just running pre-trained models.

Also, they just released about 300,000 video/Lidar/radar clips of street scenes taken from cars all over the world and in different weather. The data is free, even for commercial use. More importantly, they have a new kind of model that is trained with this data. Unlike others, this one tells you in English what it sees. (A bike cycle to the left…)

They are trying to solve the problem that Tesla and others have where even if the model gets it right, you can never know why. It appears they have taken a step back to the 1980s and have a reasoning language that it can follow rules that you can input and read. So rather than just not hitting the bike, the car “thinks” "I see a bike, I don’t want to hit it.” and you can ask why the car moved to the right. This seems to all be just a few days old (or I just found out years after everyone else already knew…)

[Marco Walther]

On 12/2/25 15:55, Chris Albertson wrote:
> A few minutes after this email, I got a notice from Nvidia. They are offering an “enthusiast’s” level computer for AI. Basic specs: 1 petaflop FP. That means one trillion 32-bit floating operations per second, and 128 GB RAM. It fits in a small-size box. $4K.

Are you talking about the DGX Spark here? That 1 petaflop is for FP4 as
per specs;-)

https://marketplace.nvidia.com/en-us/enterprise/personal-ai-supercomputers/dgx-spark/

And the 128GB are unified memory which runs a bit slower than the
dedicated memory of graphics/AI cards. But you have more of it;-)

>
> What is impressive is that this is FLOPS, not TOPS. The computer is suitable for training new models, not just running pre-trained models.

Your mileage will vary;-)

-- Marco

[Chris Albertson]

Yes, you are likely correct. I saw someone’s “repost.”

But still, what else has a petaflop? I will likely never buy a $4K computer but rather rent time on a much bigger one if needed.

What is much more interesting to me is this reasoning language. That is something I do want to try out. I’ve thought for a long time that we need to combine Prolog-like reasoning with our current LLM-based AI. I wonder if it is anything like that? No time to read tonight, maybe tomorrow or so.

[Thomas Messerschmidt]

Does it run windows?


Thomas Messerschmidt

-

Need something prototyped, built or coded? I’ve been building prototypes for companies for 15 years. I am now incorporating generative AI into products.

Contact me directly or through LinkedIn:

https://www.linkedin.com/in/ai-robotics/

On Dec 2, 2025, at 5:29 PM, Marco Walther <marc...@gmail.com> wrote:

--
You received this message because you are subscribed to the Google Groups "HomeBrew Robotics Club" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hbrobotics+...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/07957513-7a45-4c42-abb8-be11095721d8%40gmail.com.

[Stephen Williams]

For robotics, we likely want the NVidia Jetson Thor: It has similar specs as the DGX Spark, but it is an embedded board.  A big one, but still a Jetson compute card that has or will have multiple main boards available.  A bit expensive at something like $3500 before discount.  The DGX Spark is interesting, but: It is $4k rather than the $3k they projected, and it has been reported to overheat then get throttled to 100W which has disappointed some users.  Also, some of the quoted capabilities may be counting on small floats and sparse data / models.  Still impressive for what it can do for the size & power footprint.  The Spark seems to be the desktop development node for you to create programs to run on many large cloud GPUs.  Not so great as the only compute resource.

https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-thor/

I am in a discount program with NVidia which is how I managed a good discount on the RTX Pro 6000 server GPU.  That also applies to the Thor now, but the NVidia rep has been slow to get back to me.

If any of you want to get this at a discount, let me know and I will see how many I can get with the same discount.  Whenever they respond to me again.  As I am only partially crazy, I will only be buying one at this time for myself.

Stephen

--

Stephen D. Williams Founder: VolksDroid, Blue Scholar Foundation 650-450-8649 | fax:703-995-0407 | s...@lg.net | https://VolksDroid.org | https://BlueScholar.org | https://sdw.st/in

[Stephen Williams]

The Jetson Thor and Orin are on sale from NVidia at Amazon:

https://www.youtube.com/watch?v=iWFl7wF6yCw

Jetson AGX Thor Developer Kit (20% off):  https://amzn.to/4oezEZn
Jetson AGX Orin Developer Kit (50% off): https://amzn.to/4inNE1A

Scienteer

--
You received this message because you are subscribed to the Google Groups "HomeBrew Robotics Club" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hbrobotics+...@googlegroups.com.

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/324b61c6-9bbd-4d22-9c03-3cc5c1c0c0ba%40lig.net.

[A J]

The 4090 and 5090 have pretty good LLM power but draw a bit of power.

I would definitely be easier to put a DGX Spark into a Bot.

RTX 5090 (Blackwell) AI Performance (FP16/BF16)

• FP16/BF16 (Dense/Tensor): ~1.676 PFLOPS (1676 TFLOPS)
• FP/BF16 (Dense): ~209.5 TFLOPS
• Key Feature: 5th Gen Tensor Cores, FP4 support, Transformer Engine. 

• 
  

RTX 4090 (Ada Lovelace) AI Performance (FP16/BF16)

• FP16/BF16 (Tensor): ~0.66 TFLOPS (661 TFLOPS)
• FP/BF16 (Dense): ~165.2 TFLOPS
• Key Feature: 4th Gen Tensor Cores, strong FP16/INT8 performance. 

[Steve " 'dillo" Okay]

Has anybody here tried one of these hacks ? 

https://www.jeffgeerling.com/blog/2025/nvidia-graphics-cards-work-on-pi-5-and-rockchip

'dillo