Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

Discussion: HW.4 Suite - Availability, retrofit, suitability etc.

This site may earn commission on affiliate links.
…the fact that HW4 is increment to hw3 doesn’t necessarily make it miles better than its predecessor.
It does actually. They would not make it if it weren’t a huge improvement: interconnect, cooling, raw computing power, upgradability, resolution and the effect that has on maximum speed. All a big deal. It’s such a big deal that Tesla was willing to sacrifice backwards compatibility with millions of cars. And likely they over engineered it because they won’t want to make such a big change again for a while.
 
Last edited:
It does actually. They would not make it if it weren’t a huge improvement: interconnect, cooling, raw computing power, upgradability. All a big deal. It’s such a big deal that Tesla was willing to sacrifice backwards compatibility with millions of cars.
The drop in backwards compatibility in the form factor is more to prepare for future iterations (like HW5) than necessarily HW4 itself being the leap. The major change was the elimination of the GPU daughterboard (which is the MCU part, nothing to do with AP/FSD part), the rest looks like it would still largely fit on a HW3 form factor. So while there is now plenty of room to grow, HW4 doesn't really use the form factor yet to its fullest.

You can see for HW4, the general purpose CPU went from 12 cores (3x4) 2.2GHz to 20 cores (5x4) 2.37 GHz. That's a 1.8x bump, which may seem like a lot, but those are Cortex A72s which are very old cores, about akin to old smart phone level and not really responsible for most of the computing power of the chip.
From here, 4 Cortex-A72 @1300MHz gives 41.6 GFLOPS (FP32), which works out 211 GFLOPs (0.211 TFLOPs) for HW3's CPUs to 379 GFLOPs (0.379 TFLOPs) for HW4.
First OpenCL Encounters on Cortex-A72: Some Benchmarking - CNX Software

The core of the computing power is in the NPUs (or TRIP as described by Tesla). which are measured in TOPs, and the CPUs turn into a rounding error in comparison. Depending on which number you believe, HW3 was either 73 TOPs or 144 TOPs. That went from 2 NPUs to 3 NPUs or a 1.5x bump so either 105 TOPs or 216 TOPs. It's a decent incremental bump (which is speculated to help for the higher resolution cameras), but not exactly a leap (even green noted this).

Some references here and you can see the change from HW2.5 to HW3 was actually a leap (Tesla claimed 7x improvement from 21 TOPs to 144 TOPs).
Tesla's New HW3 Self-Driving Computer — It's A Beast (CleanTechnica Deep Dive) - CleanTechnica
FSD Chip - Tesla - WikiChip

I would say however the HW4 camera quality is a ton better, especially if you rely on dashcam:
 
Last edited:
Is there any information about phantom breaking on HW4?
I've had my HW4 MYLR for exactly 1 month now, I was really worried about phantom breaking, but so far I've not experienced it at all after 2000 miles of driving. I do use Autopilot extensively, even at night too.
After comparing HW3 video footage with HW4 footage, I think reduction in headlight glare might be even more beneficial than increased resolution? It's probably much easier for the computer to process where the nearby vehicles are with reduced glare?
 
Yes, if you are in US, you will get either Fremont or Austin Built MY depends on spec. Both facilites have fully transited to Hw4 MY. I heard some people bought existing inventory and got HW4. You could check as long as the Vin is after PF799 or PA140.
Curious the source of the VIN #s. I missed out on something like this with my Model 3, build date was actually AFTER Tesla's range but still didnt have the right harness for a retrofit.

I am looking at a PF841 - based on your logic should have HW4, then.
 
Now that we can transfer a previous FSD purchase from previous car, I'm even more interested in trading my M3 2019 for a new MY. But the terms of the FSD transfer are super stringent, including language, apparently, that if I refuse delivery, my current FSD purchase on my M3 will evaporate never to be heard from again. So if I show up to pick up my new MY and it is not HW4, I either refuse it, and lose the FSD I bought in 2019 (forever--they will not put it back on the old M3) or I accept an old HW version of the MY. That's a tough risk to take.

Are we *guaranteed* at this point to be getting HW4 on a new MY purchased today? The stakes are high!
Where did you get the terms? I haven't been able to find anything official yet.
 
I did a ton of research in the last couple of days. I didn’t save anything that I could post here. And as you point out, it seems hard to find anything official. But it’s pretty clear that there is some risk involved.

I’ll see if I can backtrack to get you the info I found.
 
Where did you get the terms? I haven't been able to find anything official yet.
Look at post 230 here:

 
  • Like
Reactions: hasselboxster
Curious the source of the VIN #s.

There are a couple of threads on reddit where helpful people with "repairer" access to Tesla's EPC were looking up VINs and confirming what parts they were built with.

Tesla don't always build in VIN order so if you're within the changeover range then you can't guarantee you have the new parts - which likely happened with your Model 3 - but any Model Y built in the USA in the last month or so will have HW4.

This thread says

Austin HW4 switch date: Jun 4 2023. (VIN PA131200 ++ - all VIN after do have HW4)
Fremont HW4 switch date: May 24 2023. (VIN PF789500 ++ - all VIN after do have HW4)

And this thread says

I'm no longer doing lookups for Austin/Fremont VINs. HW4 is rolling out to Austin (VIN A130xxx +) and Fremont (VIN 798xxx +).

I am looking at a PF841 - based on your logic should have HW4, then.

Yes it should. But if you want to be 100% sure you can ask the poster on the 2nd thread to do a full lookup of those VINs for a small donation.
 
Matrix headlights would be nice, wasn't expecting those on an Austin LR build - I'm also waiting for 7SAYGDEE5PA15XXXX here. To be delivered within a week or so.
Matrix headlights are totally useless because there is no software to control the actual matrix part AND there is no mention of when it will be released.

It seems Tesla is to busy making zombie games for the infotainment part, to actually code stuff that matters like matrix headlights...
 
Matrix headlights are totally useless because there is no software to control the actual matrix part AND there is no mention of when it will be released.

It seems Tesla is to busy making zombie games for the infotainment part, to actually code stuff that matters like matrix headlights...

Tesla apparently can't enable the full matrix capability yet because the regulations in the US haven't caught up to the technology. Until the government gets around to specifying how they must/are allowed to work, we're stuck with them emulating "dumb" headlights.
 
  • Like
Reactions: richyrich
Tesla apparently can't enable the full matrix capability yet because the regulations in the US haven't caught up to the technology. Until the government gets around to specifying how they must/are allowed to work, we're stuck with them emulating "dumb" headlights.
That might be an explanation for the US, but here in Europe Matrix headlights have been accepted, so why not enable it for our markets ?

The FSD for cities has been enabled in the US, even though it has apparantly not been approved for use in Europe.