FSD v13 and Hardware 4 Partition Limit

[CyberTexas]

This thread is packed with more maybes, possiblys, and what-ifs than a conspiracy theory subreddit. I just need a straight-up yes, you can upgrade to 5, or no, sit your ass down. The speculation in here is heavier than my Cybertruck fully loaded for overlanding hahaha.

I hope that i am wrong. But the fact that v13 is reaching limit of storage of HW4 is worrisome. Maybe Tesla can do some optimizing magic and shrink v13 down 3x??

Sponsored

 

[REM]

This thread is packed with more maybes, possiblys, and what-ifs than a conspiracy theory subreddit. I just need a straight-up yes, you can upgrade to 5, or no, sit your ass down. The speculation in here is heavier than my Cybertruck fully loaded for overlanding hahaha.

 

[DJAlan2000]

yeah, I think we should all just wait until HW8 comes out, I'm sure it will have 5 TB of space and a processor that can run at 10 times the speed of a Ryzen Threadripper 7995WX...

But doesn't need liquid cooling and can fit in the space of a cigarette pack!!

Yeah, let's wait...

Well, I will likely be buying something with HW5 (IF it is due out in Dec) or, if I need to wait another year for HW6 I could do that too...

Maybe for the CyberSUV...

 

[IntoTheRain916]

If I had to trust anyone to make it happen, it’s the Tesla engineers—those wizards make rockets land themselves and cars fart on command.

 

[CyberTexas]

If I had to trust anyone to make it happen, it’s the Tesla engineers—those wizards make rockets land themselves and cars fart on command.

The same way those engineers got FSD v13 to work on HW3??

 

[Outdoors]

yeah, I think we should all just wait until HW8 comes out, I'm sure it will have 5 TB of space and a processor that can run at 10 times the speed of a Ryzen Threadripper 7995WX...

But doesn't need liquid cooling and can fit in the space of a cigarette pack!!

Yeah, let's wait...

Well, I will likely be buying something with HW5 (IF it is due out in Dec) or, if I need to wait another year for HW6 I could do that too...

Maybe for the CyberSUV...

Yeah I had a hardware friend. First it was I am waiting for 3.00 then it was. I am not sure. Maybe they could do something with an overhead camera. The upgrade for 2.5ers confused him. Had past tense. Never bought anything. Always waiting for perfect. Imagine same person solar panel shopping. Yikes.

 

[REM]

The same way those engineers got FSD v13 to work on HW3??

Who says they can't?

 

[REM]

Imagine same person solar panel shopping. Yikes.

How dare you trigger me with this on a wednesday lmao! I have been waiting so long because the prices keep going down ?

 

[scottf200]

Software can always outgrow hardware, Tesla has no incentive to not use all the available space. Why optimize/ shrink the NN if it fits?

Tesla/Musk was bragging about their new "FSD chip" and how it had redundancy. That was 100% talked about on the Tesla’s Autonomy Investor Day. Green has consistently pointed out when they went over their limits and could no longer have redundancy. Some people (not necessarily you) have really short memories on Tesla and what they say vs what they do.

Each of Tesla’s new Full Self Driving Computer boards actually come with two of the new chips for redundancy, and it’s just one of the many redundant features you’d hope and expect to find in situations where you’ll be trusting your life to a computer — you’ll also find redundant power, and even redundant calculations where the system compares results from both processors before it steers the car. “Any part of this could fail, and the car will keep driving,”
says Musk.

https://www.theverge.com/2019/4/22/...g-chip-is-here-and-this-is-your-best-look-yet

 

[CyberGus]

Tesla/Musk was bragging about their new "FSD chip" and how it had redundancy. That was 100% talked about on the Tesla’s Autonomy Investor Day. Green has consistently pointed out when they went over their limits and could no longer have redundancy.

The early FSD computer was almost like a mainframe, with duplicated CPU and RAM doing the same work. When they hit the wall on compute, they instead spread the work across them, basically doubling the power "for free" at the expense of redundancy.

IDK if there's even the pretense of such redundancy in HW4, but I doubt it. It's a rare event to have the two parallel CPUs disagree, and even if they did you'd want the car to stop anyway. The juice is not worth the squeeze.

They probably do have that level of redundancy at SpaceX, but they really can't afford even a one-in-a-million failure until there's a Radio Shack in orbit.

 

[scottf200]

IDK if there's even the pretense of such redundancy in HW4, but I doubt it. It's a rare event to have the two parallel CPUs disagree, and even if they did you'd want the car to stop anyway.

That is unclear. I'm finding references to it in articles and the Munro video.

There is also a question if redundancy is 'required' for unsupervised ("level 4/5") FSD. ala RoboTaxi. [Level 4 Requirements: Level 4 autonomy means the vehicle can operate without human intervention within a specific Operational Design Domain (ODD)]

a)

One significant aspect of Hardware 4 is the reintroduction of redundancy. Each board in the new hardware suite has two nodes, which can compute the same data and compare outputs, allowing the vehicle to identify any inconsistencies. This redundancy feature, previously removed in Hardware 3 to maximize processing power, has been reinstated in Hardware 4, ensuring that the vehicle can continue driving safely even if one node fails.

Tesla Hardware 4 (AI4) - Full Details and Latest News - AutoPilot Review

b)
>>>>>>>>
The presenter mentions that the placement of the main harness connectors on each end of the board is related to the symmetry of the board and helps maintain hardware redundancy for safety [04:35]. They also emphasize Tesla's priority on safety [04:42].
<<<<<<<<

 

[CyberGus]

As for incentive to optimized…one only has to look at the success of DeepSeek. Optimization of software for such a lean company like Tesla (and Musk) should be top priority regardless of available space!

Generally, optimization is done after you get it working. It's pointless to optimize code that you're not even sure you're gonna keep.

The DeepSeek revelation was huge news because it's a straight-up unicorn. They managed to improve memory utilization by ~40x without any impact on compute, nor any loss of function. It's brilliant, but such opportunities are rare. You could never "optimize" Win11 to run in 640k of RAM.

You can't just keep running a file through pkzip over and over until it's zero bytes

 

[dalton108]

Generally, optimization is done after you get it working. It's pointless to optimize code that you're not even sure you're gonna keep.

The DeepSeek revelation was huge news because it's a straight-up unicorn. They managed to improve memory utilization by ~40x without any impact on compute, nor any loss of function. It's brilliant, but such opportunities are rare. You could never "optimize" Win11 to run in 640k of RAM.

You can't just keep running a file through pkzip over and over until it's zero bytes

What about WinAmp? I hear It really whips the llama’s ass!

 

[CyberTexas]

Generally, optimization is done after you get it working. It's pointless to optimize code that you're not even sure you're gonna keep.

The DeepSeek revelation was huge news because it's a straight-up unicorn. They managed to improve memory utilization by ~40x without any impact on compute, nor any loss of function. It's brilliant, but such opportunities are rare. You could never "optimize" Win11 to run in 640k of RAM.

You can't just keep running a file through pkzip over and over until it's zero bytes

Maybe pointless at beginning, but when your software is 7.xx GB and your hardware limit is 8 GB, then maybe it is not a bad time to start thinking about optimizing, especially when it may prevent your $100k+ paying customers from using your latest software.

Like someone above said, FSD Is ever so evolving. So when will you get to optimizing? 2050?

 

[REM]

Maybe pointless at beginning, but when your software is 7.xx GB and your hardware limit is 8 GB, then maybe it is not a bad time to start thinking about optimizing, especially when it may prevent your $100k+ paying customers from using your latest software.

Like someone above said, FSD Is ever so evolving. So when will you get to optimizing? 2050?

The fact that you think it isn't already extremely optimized is dumbfounding to me ?

Sponsored