Greg Hogan

@claudeai has reached AGI

27 tinybox pro servers up and running! 216 GPUs doing some heavy weight lifting :)

@PigeonInvestor service.tesla.com/en-US/vehicle-…

@gregjhogan How did you get the pinout of the radar? It is CAN o ETH?

2021 (or older) Tesla Model 3 with radar is the best openpilot vehicle you can buy if you are willing to tear apart your car to get to connector X121 (pins 6/7) on the autopilot computer to tap the radar bus, but unfortunately it isn't that easy to get to.

@ellenjxu_ openpilot experimental mode is not great yet, but hopefully that changes soon! This makes a big difference for chill mode, where vision only ACC is quite bad.

@gregjhogan Wow will have to try this out! How does it compare with experimental?

Code to make it work here, basically just copied in the old model s raven (continental) radar interface. github.com/gregjhogan/ope…

Also did they really completely stop using the radar on equipped vehicles? The autopilot computer gets mad when it is disconnected which is a little strange.

@timzaman @__tinygrad__ We optimize for $/flop and low maintenance. We still have some room for improvement, but liquid cooling would be in the wrong direction.

@gregjhogan @__tinygrad__ You can put new massive 200mm fans on? i personally put a heat exchanger across the entire front and liquid cool everything, leading to superb cooling performance

There's been a lot of requests for an 8x 4090 tinybox in a 4U form factor. We have internal prototypes, and can produce for $40,000. 2x GENOA CPU, 384GB RAM Looking for 20 non-refundable 5% down ($2k) preorders with a ship date of December to make this happen. Who's in?

@Liquidator_1 I used this dump the flash on that ECU github.com/gregjhogan/ren… Which also has this CAN shell that could be used to dump the flash github.com/gregjhogan/can…

@gregjhogan Hey Greg, would a similar approach be possible on the following rack? #post-518734" target="_blank" rel="nofollow noopener">t6forum.com/threads/t6-1-s… thank you very much :)

#carhacking doesn't get any easier than every pin you would want has a pad, and read memory is not disabled :)

@Tesla It is the autosar crc8 of a byte from a lookup table indexed by the current counter, and the initialization value is the autosar crc8 of the data bytes in the CAN message.

@Tesla openpilot support needs the checksum algorithm for the SCCM stalk CAN messages and it is something non-standard ... no problem, easy to get from the firmware! But why so weird?!

I guess I won't be building this 100 gpu machine anytime soon

I tried setting up a machine with 40 nvidia gpus but only 32 show up :( Looks like some driver modifications are required to exceed 32, but those changes would break most existing software that uses the driver.

@__tinygrad__ It seems like there should be some way to get this working on older GPUs, too, based on the fact that this exists: github.com/terminatorul/N…

Source is here. With some cleanups, it might even be upstreamable. It relies on large BAR support. There's a decent writeup of how it works in the README. github.com/tinygrad/open-…

We added P2P support to 4090 by modifying NVIDIA's driver. Works with tinygrad and nccl (aka torch). 14.7 GB/s AllReduce on tinybox green!

@yacineMTB With 8 GPUs in a server it makes a huge difference for algorithms like all reduce. And perhaps you already know it is purely a software limitation, which is even more infuriating! twitter.com/gregjhogan/sta…

actually it doesn't really matter according to lambda labs blog post lambdalabs.com/blog/nvidia-rt… but also i presume i can just work around it with architecture

it's come to my attention that multi RTX 4090 systems don't support p2p. i've decided that this doesn't matter for me. my goal is to build a big system which I can both run inference for dingboard but also train models for experimentation with spare compute. video related

lol, shouldn't have removed the emmc chip to dump it, you can use the CORE_UART port to get in

most of the bytes in the dump are 0xCC 😢

How to dump the firmware on a Unitree Go2: - plug usb-c port into computer - hold button circled in red below - power it up - run: rkdeveloptool rl 0 200000000 output