Nvidia Unveils Drive Thor, One Chip To Rule All Software-Defined Vehicles (techcrunch.com) 56
An anonymous reader quotes a report from TechCrunch: Nvidia is gearing up to deliver Drive Thor, its next-generation automotive-grade chip that the company claims will be able to unify a wide range of in-car technology from automated driving features and driver monitoring systems to streaming Netflix in the back for the kiddos. Thor, which goes into production in 2025, is notable not just because it's a step up from Nvidia's Drive Orin chip. It's also taking Drive Atlan's spot in the lineup. Nvidia is scrapping the Drive Atlan system on chip ahead of schedule for Thor, founder and CEO Jensen Huang said Tuesday at the company's GTC event. Ever in a race to develop bigger and badder chips, Nvidia is opting for Thor, which, at 2,000 teraflops of performance, will deliver twice the compute and throughput, according to the company.
"If we look at a car today, advanced driver assistance systems, parking, driver monitoring, camera mirrors, digital instrument cluster and infotainment are all different computers distributed throughout the vehicle," said Nvidia's vice president of automotive, Danny Shapiro, at a press briefing Monday. "In 2025, these functions will no longer be separate computers. Rather, Drive Thor will enable manufacturers to efficiently consolidate these functions into a single system, reducing overall system cost." One chip to rule them all. One chip to help automakers build software-defined autonomous vehicles. One chip to continuously upgrade over-the-air. Further reading: Nvidia Announces Next-Gen RTX 4090 and RTX 4080 GPUs
"If we look at a car today, advanced driver assistance systems, parking, driver monitoring, camera mirrors, digital instrument cluster and infotainment are all different computers distributed throughout the vehicle," said Nvidia's vice president of automotive, Danny Shapiro, at a press briefing Monday. "In 2025, these functions will no longer be separate computers. Rather, Drive Thor will enable manufacturers to efficiently consolidate these functions into a single system, reducing overall system cost." One chip to rule them all. One chip to help automakers build software-defined autonomous vehicles. One chip to continuously upgrade over-the-air. Further reading: Nvidia Announces Next-Gen RTX 4090 and RTX 4080 GPUs
one chip to hack it all / one chip to lockout repa (Score:2)
one chip to hack it all / one chip to lockout repair of any thing at non dealer places.
Re: (Score:3)
Indeed. Putting critical safety functions on the same chip as the entertainment system is a really bad idea.
Re: (Score:2)
I used to work in the automotive industry and I can tell you no one is going to put actual safety critical stuff on the same piece of hardware there are user accessible USB ports on.
Re: (Score:2)
I'd like to believe that no-one in the auto industry is that stupid. I really would. Unfortunately numerous big name brands have previously shown that money beats quality, security and even safety all the time.
For example, right now they're falling themselves to put controls you might need to use while driving onto touchscreens. That seems to be the new controls madness after they got bored of loading the steering wheel with controls for things you probably have no business using at all while driving.
Meanwh
Re: (Score:2)
Oh? Did you pay attention to systemd? Putting all your eggs into one otherwise incompatible basket is a very popular "architectural" approach to new projects. Redundancy and duplicated hardware does cost money, and time
Re: (Score:2)
"no one is going to put actual safety critical stuff on the same piece of hardware there are user accessible USB ports on."
Jeep did. https://www.wired.com/2015/07/... [wired.com]
From the Pinto to the Jeep and beyond, history teaches us that no-one in the automotive industry is thinking hard about consequences of crappy design or poor security.
Re: (Score:2)
Re: (Score:2)
Also, it's not like current chips aren't hackable either.
Current chips are hackable if you have physical access to the JTAG pins.
With this chip, you may be able to use the game console in the backseat.
One advantage of the current system (where you have lots of small chips around the vehicle) is that each chip contains simpler firmware than one chip consolidating all the functions.
That is a really, really important advantage. Why would anyone give it up?
Re: (Score:2)
Current chips are vulnerable to wireless keyfob attacks, which unlocks the car and disables the alarm and hence allows physical access to the entire car. Immobilizers are also hackable from the ignition.
Please try to understand the following: In this age of hardware-assisted virtualization and secure enclaves, not everything runs in the same memory space. The code
Re: (Score:3)
Please try to understand the following: In this age of hardware-assisted virtualization and secure enclaves, not everything runs in the same memory space. The code for the game console in the backseat won't run in the same memory space as the code for the control systems or compete for CPU resources. I am surprised by how many Slashdotters are completely unaware of modern practices.
Or maybe they've seen the seemingly endless stream of vulnerabilities where supposedly isolated code still manages to access things it shouldn't on modern chips that tried to be a bit too clever in the name of performance. Ask anyone who runs a data centre with big servers hosting virtual machines how that's been working out lately.
Re: (Score:2)
Please try to understand the following: In this age of hardware-assisted virtualization and secure enclaves, not everything runs in the same memory space.
Still prefer discrete machines over a hypervisor running a bunch of VMs for reliability.
Re: one chip to hack it all / one chip to lockout (Score:2)
How many times has SGX been broken open now?
Actual secure processors on the same die have been relatively successful but enclaves have been a failure AFAICS.
Re: (Score:2)
Re: (Score:2)
Nah, there's basically two ways. One which has failed more often than not.
Spectre and its many offshoots are bad enough for, but with the level of complexity of secure enclaves true privilege escalation bugs are more likely than not to be hiding in there as well, SGX had them. Then there's rowhammer and it's offshoots for more fun, a new way for extremely high power density localized workloads to break isolation can lurk way too easily with closely shared resources, caches and memory shouldn't be shared eit
Re: (Score:2)
Since it doesn't go into production for three years, that gives even slow legacy auto companies time to get onboard. Also, since these same companies are clueless about software and automation, I predict that they will be happy to have someone do this for them.
Good move by Nvidia.
Re: (Score:2)
I assure you the automotive industry is very up on software and automation, but at the end user level they rather not take the risk when people's lives, and ultimately the brand reputation / value takes a hit.
That's why transportation usually lags behind, why replace a module that's been doing its one job for the last 15 years for a do-dad that gets nerfed by a new version of bluetoof and fucks with your ABS modules ... even then they still fuck it up (looking at you Boeing and Toyota)
Re: (Score:1)
Posting AC for reasons, but Hyundai and VW both use large boxes with Nvidia GPUs for self driving.
Re: Who is using it? (Score:2)
Unless NVIDIA merges with some major full stack driver automation company, I don't see how they have a future in automotive.
Also 2 petaflops is way more than needed for simple drive assistance and level 4/5 never ever so who would pay for this expensive slab of silicon?
Could we have a separate chip (Score:2)
One question (Score:1)
Does this mean my windshield becomes a first person shooter game?
Re: (Score:3)
Now what did you think that star on the hood is for? It's a crosshair, duh.
One chip to hack it all... (Score:2)
I dislike the whole concept of the vehicles operation being integrated into the entertainment / communication system. It just leaves the vehicle open to being hacked, and when it's hacked the control of the vehicle can be compromised and people can die.
Re: (Score:2)
It's not going to be one big chip running everything. It will be a chip with no video output stage and a whole bunch of cores being used for self driving, and a chip based on the same cores with the video output and less cores doing infotainment and stitching the camera views streamed from that system (to which the cameras will be directly connected) together for top-down view and such. The automakers aren't that crazy. They will want to update the two systems separately (and frequently.)
Re: (Score:2)
It's not going to be one big chip running everything.
That would make sense, but isn't what the TFA says. NVIDIA [nvidia.com] says it will be a single SoC with the capability to partition tasks (how they are partitioned isn't described in the marketing blurb). The also mention certification to ASIL-D which means 1 failure per 10E7 hours, so the hardware should be pretty reliable. Just from a development and qualification perspective, I think they'd want to keep them physically partitioned, but maybe the bean counters say otherwise...
Re: (Score:2)
That would make sense, but isn't what the TFA says.
Welp, that's what nvidia says, but I don't think most of the automakers are going to bite. It doesn't make sense for the automakers to want to put all of that in one package because they're going to want to use the same self-driving hardware over and over again, but they will want to segment the infotainment. They will also want to be able to replace those two things independently. They might get some of the smaller players to fall for it, but I suspect those guys will want something cheaper than what they
Re: (Score:2)
Welp, that's what nvidia says, but I don't think most of the automakers are going to bite.
Fair point! Marketing claims don't make it so.
Re: (Score:2)
The also mention certification to ASIL-D which means 1 failure per 10E7 hours, so the hardware should be pretty reliable.
Just a nitpick, but ASIL-D is 10 FIT, which is a failure every 1e8 hours. But not all car functions need to be ASIL-D. For example, infotainment is not, but steering control should be.
Re: (Score:2)
Chips really do not cost a great deal of money; this is not worth the savings excuse.
CHIP 1:
security, driving. Secure as possible; isolated and hardened and slower. upgrade possible but with some physical action required no wifi auto update. heavily regulated by government. some physical controls also required. I do not want some idiot driving their car with a PS controller; or a fool taking a nap while their falsely labeled auto-pilot kills both of us.
CHIP 2:
I want the fluff, wifi and the apps on whatever
well (Score:2)
there's a reason the "computers" (mostly microcontrollers) are distributed, its modular so you can use the same thing in multiple cars, platforms and brands (especially in the T2 / T3 suppliers we are sneaky little bitches) . Its also easier to replace a self contained (I hate to say this) "smart" module if something gets damaged, and to that point you don't have to worry as much that if you doink your camera module that's talking on a "network" that it takes out your speedo, so it adds a false sense of red
When does the superclocked version come out? (Score:3)
No part of the drive by wire system should every be near an infotainment system. FFS.
Re: (Score:1)
Re:When does the superclocked version come out? (Score:4, Insightful)
Re: (Score:2)
Failures that would prevent you from bringing the vehicle to a controlled stop either have redundancy or are designed robustly to make failure highly improbable. This is born out in accident statistics that show the nut behind the wheel to be the biggest single point failure.
And, coming to a controlled stop in the event of a failure
Re: (Score:2)
That's pretty good
Re: (Score:2)
Re: (Score:3)
the infotainment system will run on the same memory space as the drive-by-wire system. PROTIP: It will not.
At the same time, I never thought that nuclear power plants or electric substations would be accessible from the internet. I thought people would take the sane approach, and always leave an air gap. And yet, here we are: they have not. It's connected to the internet because people want features.
in the current age of virtualization and secure enclaves,
It is theoretically possible to make virtualization secure, so memory in one VM can't access the memory in another VM. In practice, there's no such thing as a "secure enclave" and you're deluding yourself if you thin
Re: When does the superclocked version come out? (Score:2)
I'm amazed in the era of meltdown, rowhammer and the joke that is SGX anyone would trust virtualization for isolation.
Re: (Score:2)
No part of the drive by wire system should every be near an infotainment system. FFS.
Even if it can be made safe, how would this make it past systems engineering? You want your N^2 diagram to be as sparse as possible.
Did the Netflix App update their codecs? Time to requalify the self-driving capability! I can't imagine you'd save enough on hardware to make mingled safety-critical functions and infotainment worthwhile.
trust Nvidia? (Score:2)
Re: (Score:2)
Given Nvidia's approach to partnering why would any car manufacturer ever trust these pricks?
Worked out fine for Microsoft. Similarly, the car companies are big enough and have enough lawyers to get a proper deal out of nvidia.
Re: (Score:2)
Microsoft got royally screwed by Nvidia
They refused to discount the price of the special custom GPU they built for Microsoft, and that somehow was screwing them? By that standard Microsoft is screwing literally everyone by overcharging for their fucked over OS. Oh wait, the USDoJ proved in court that Microsoft IS screwing literally everyone (then declined to prosecute, then Gates Foundation, then Windows 10 spyware. But I'm sure that's a coincidence.)
The chip technlogy i want to see in my next car (Score:2)
Great (Score:2)
Tesla FSD (Score:2)
For comparison, the Tesla FSD chip which was released in 2019 (six years before Thor is released) puts out just under 75 TOPS on the NPUs (neural processing units, half that number on each of the two). If the 2000 teraflops number Nvidia is putting out is the same thing, then that's about 30 times more powerful for something six years later. That sounds about right to me. (And Tesla should have the next gen FSD computer by then, and it will be interesting to compare it with this Nvidia chip.)
Note: Tesla
Re: (Score:2)
If they are so great at knowing "what the auto companies are looking for" .. why didn't they announce any major companies as partners (unheard of Chinese brands looking to experiment don't count). They had Tesla and lost 'em. They had Mercedes and lost them too.
One chip to continuously upgrade (Score:2)
Drivers? (Score:2)
Given Nvidia's apparently inability to build stable drivers, and/or a reliable GPU, why on earth would we want that controlling a car? Oh look, Netflix just crashed while on the freeway. Just hold on a sec while I reboot the steering and braking controls.
It sounds like an overblown comment, but you absolutely know there is some product manager in there, working out if they can combine a bunch of functions to reduce cost, and thus make their quarterly bonus.
1000 Watts Power Consumption? (Score:1)
Nvidia avoids disclosing any info on power consumption again, Tesla HW3 FSD Chip uses 36 watts, while Nvidia current gen Drive uses 1000 watts, how is that competitive?
But what is on the chip? (Score:2)
Is this an ARM CPU like their Denver or licensed stuff? AI accelerator like Hopper? GPU like Ada Lovelace? Or some combo? Or ???