Automotive A100 SXM2 for FSD? (NVIDIA DRIVE A100)

Notice: Page may contain affiliate links for which we may earn a small commission through services like Amazon Affiliates or Skimlinks.

Leiko

New Member
Aug 15, 2021
26
6
3
TMUs should be a direct relationship to SMs. 4x SMs.

4*96 = 384 TMUs.

ROPs seems a little more in flux. not sure these matter too much on a GPU that doesnt render graphics anyway.
If you have any idea about the 900 int8 TOPS, I'm interested. They said on the official page that this card was a A100 but "in an automotive grade form factor", maybe they buffed int8 for the computer vision tasks ?

1737302153784.png
 

gsrcrxsi

Active Member
Dec 12, 2018
401
138
43
almost certain that the "automotive-grade" is referring to the fact that it's covered in a conformal coating to resist corrosion.

i'd need some kind of app to test the int8 performance in exactly the way they claim.
 

gsrcrxsi

Active Member
Dec 12, 2018
401
138
43
this is a PG199 (pre production drive a100 prod) gpu-z, courtesy of a Chinese friend:
View attachment 41425

And these are the official specs from when they were on Nvidia's website (they deleted the page since):
View attachment 41426
I'm not sure how to differentiate what is "Production" and what is "pre-production" or "ES/QS".

my unit came in an Nvidia branded box. the GPU itself has a matching serial to the box. does this mean it's "Production"?

yet, the die heat spreader markings indicate "QS", nvidia-smi labels it as "PG199-PROD", and the VBIOS revision is exactly the same VBIOS as your "pre-production" screen shot. 92.00.79.00.01

a lot of conflicting information.
 

Leiko

New Member
Aug 15, 2021
26
6
3
I'm not sure how to differentiate what is "Production" and what is "pre-production" or "ES/QS".

my unit came in an Nvidia branded box. the GPU itself has a matching serial to the box. does this mean it's "Production"?

yet, the die heat spreader markings indicate "QS", nvidia-smi labels it as "PG199-PROD", and the VBIOS revision is exactly the same VBIOS as your "pre-production" screen shot. 92.00.79.00.01

a lot of conflicting information.
My friend told me the ones which says QS are production. The es/qs are pg199 written on the die and in the name in nvidia-smi
 

Leiko

New Member
Aug 15, 2021
26
6
3
almost certain that the "automotive-grade" is referring to the fact that it's covered in a conformal coating to resist corrosion.

i'd need some kind of app to test the int8 performance in exactly the way they claim.
I’m trying to write a cublasLt benchmark right now but its not going well x)
 

gsrcrxsi

Active Member
Dec 12, 2018
401
138
43
There's a big chance they were from different cars, and differ a little.
Mine was brand new. Never installed in any car. Unless there’s different production/pipelines for different SKUs intended for different vehicles.

I kind of got the impression that these things were never intended to be installed in production vehicles though. Only to be used as a developement platform for the various manufacturers to play with, then their final design actually intended to be used for cars would be a more custom/proprietary ECM type product that would plug into the vehicle much like other electronics modules attach to vehicles. (No SXM2 interface, all soldered to a single board, automotive connectors, etc)
 

CyklonDX

Well-Known Member
Nov 8, 2022
1,413
485
83
Unless there’s different production/pipelines for different SKUs intended for different vehicles.
thats the idea.

Only to be used as a developement platform for the various manufacturers to play with, then their final design actually intended to be used for cars would be a more custom/proprietary ECM type product that would plug into the vehicle much like other electronics modules attach to vehicles.
I do recall seeing gpu car part number some time ago mercedes while looking for a100s and such but that was long time ago - i can't find them anymore i guess mechanics would know - as they have software with lists of all car parts and steps.


I do recall seeing nvidia using arm cortex setup later down the road.
 
  • Like
Reactions: gsrcrxsi

Leiko

New Member
Aug 15, 2021
26
6
3
Mine was brand new. Never installed in any car. Unless there’s different production/pipelines for different SKUs intended for different vehicles.

I kind of got the impression that these things were never intended to be installed in production vehicles though. Only to be used as a developement platform for the various manufacturers to play with, then their final design actually intended to be used for cars would be a more custom/proprietary ECM type product that would plug into the vehicle much like other electronics modules attach to vehicles. (No SXM2 interface, all soldered to a single board, automotive connectors, etc)
If you got from the same ebay vendor as me, he explained why they had this many for sale: a major car brand purchased a lot of them for a projet but that same project was scrapped halfway through. They paid these hardware liquidators to sell all of the cards and recover some of the invested money.
 
  • Like
Reactions: CyklonDX

gsrcrxsi

Active Member
Dec 12, 2018
401
138
43
So it looks like we all have QS then. ES/CS/QS all have the same bios revision. Production has a different bios.

I wonder if it’s possible to flash the production bios to the QS module. Would that enable power limiting?

are the boost clocks different on the Production version vs QS? Are there any other differences between QS and Production in specs or performance?
 
Last edited:

MilkyWeight

New Member
Mar 15, 2024
12
1
3
This is the whole contraption for now: Drive A100.

Unfortunately, the power consumption regularly spikes to 400W, and it makes plenty of coil whine, maybe because the Chinese adapter doesn't have the NVLink populated, which has a bunch of grounds. I'm also worried about the power connectors, perhaps I should add one more.

This power consumption explains why I killed some of the servers.
What kind of screws did you use to attach it to the adapter board?
 

MilkyWeight

New Member
Mar 15, 2024
12
1
3
nvlink doesn't work by default for sure. I KNOW that some people on this planet have it working. I ordered an aom-sxmv and got 4 of these cards sitting on my desk, will be posting updates here once everything arrives.
Where did you learn about this? Did they modify the open GPU kernel modules to make it work? Or did they swap the firmware with nvflashk into A100 40GB?
 
Last edited:

generalheed

New Member
Feb 1, 2025
2
0
1
windows Is a pain to get working but I'll post a screenshot when mine work. I do have a screenshot from the PG199 if that interests you.
I've been curious about getting one of these and an adapter to try out for some AI stuff as well as other things too. I'm wondering though, how does Windows support work? What did you have to do to get it working?

Also will any SXM2 to PCIe adapter work for this? I've been eyeing this one on ebay: SXM2 To PCIE Adapter For Nvidia Tesla V100 A100 SXM2 GPU Computing Graphics | eBay

That listing says that adapter works for the A100 but I can't find much info on what the A100 automative actually works on. But if that adapter can work, that'd be really awesome to make a slim dual slot A100.
 

MilkyWeight

New Member
Mar 15, 2024
12
1
3
I've been curious about getting one of these and an adapter to try out for some AI stuff as well as other things too. I'm wondering though, how does Windows support work? What did you have to do to get it working?

Also will any SXM2 to PCIe adapter work for this? I've been eyeing this one on ebay: SXM2 To PCIE Adapter For Nvidia Tesla V100 A100 SXM2 GPU Computing Graphics | eBay

That listing says that adapter works for the A100 but I can't find much info on what the A100 automative actually works on. But if that adapter can work, that'd be really awesome to make a slim dual slot A100.
yes the A100 will work. Depending on your system you may have unreliability in the card being recognized/falling off the bus.

you can run some boot commands to make it reliable.

cooling is going to be the tricky challenge. Even a 4U cooler isn’t enough. You also can’t control its power usage either.

probably need to get the water cooling. It runs hot.
 

generalheed

New Member
Feb 1, 2025
2
0
1
Thoese adapters are totally overpriced. ~80-100 €/$ worth it.
Unfortunately that's about the cheapest I can find on eBay, no one is selling that version cheaper. There is this other version that's about $100 cheaper: SXM2 to PCIE Adapter For Nvidia Tesla V100 A100 SXM2 GPU Computing Graphics | eBay

But this one is quite different and doesn't come with an enclosure, heatsink and fan. The heatsinks I've seen other people here use are those giant blocks. I was hoping the other adapter I was looking at that comes with the heatsink and fan in dual slot form factor would be enough to cool an A100.

yes the A100 will work. Depending on your system you may have unreliability in the card being recognized/falling off the bus.

you can run some boot commands to make it reliable.

cooling is going to be the tricky challenge. Even a 4U cooler isn’t enough. You also can’t control its power usage either.

probably need to get the water cooling. It runs hot.
What would I have to watch out for on my system that would cause the card to be unreliable/fall of the bus? Is it my motherboard, OS or the adapter itself? What boot commands are there that would make it more reliable on a Windows system?

And yeah I'm a bit nervous about the cooling as well. The adapter I was looking at initially has a heatsink, fan and enclosure to make the whole thing look like a typical dual slot GPU. But I wonder how well that'd be able to cool this A100. Supposedly it's sufficient for the V100 and it's listed as supporting the A100 so I'm not sure. Haven't found anyone that's actually tested that specific adapter with this particular A100 yet.
 

MilkyWeight

New Member
Mar 15, 2024
12
1
3
Unfortunately that's about the cheapest I can find on eBay, no one is selling that version cheaper. There is this other version that's about $100 cheaper: SXM2 to PCIE Adapter For Nvidia Tesla V100 A100 SXM2 GPU Computing Graphics | eBay

But this one is quite different and doesn't come with an enclosure, heatsink and fan. The heatsinks I've seen other people here use are those giant blocks. I was hoping the other adapter I was looking at that comes with the heatsink and fan in dual slot form factor would be enough to cool an A100.



What would I have to watch out for on my system that would cause the card to be unreliable/fall of the bus? Is it my motherboard, OS or the adapter itself? What boot commands are there that would make it more reliable on a Windows system?

And yeah I'm a bit nervous about the cooling as well. The adapter I was looking at initially has a heatsink, fan and enclosure to make the whole thing look like a typical dual slot GPU. But I wonder how well that'd be able to cool this A100. Supposedly it's sufficient for the V100 and it's listed as supporting the A100 so I'm not sure. Haven't found anyone that's actually tested that specific adapter with this particular A100 yet.
I’ve tested it. It works.
Just get the water block. Bykski or whatever it’s called.

these commands try them in order adding one by one and testing
GRUB_CMDLINE_LINUX_DEFAULT="quiet splash pci=realloc pci=nocrs pci=nommconf"
 

gsrcrxsi

Active Member
Dec 12, 2018
401
138
43
I haven’t seen the heatsink completely that comes with those eBay adapters. But it looks like a solid block of copper with fins. Totally insufficient to cool the A100 IMO. Maybe barely enough to cool a power limited V100.

I’m using a waterblock on my A100 and those 3U heatsinks on my V100s.