D
Deleted member 887
Guest
Since the famous Linux code review which revealed the t239 "has eight cores in a single cluster", there hasn't been any significant reference to t239 or Tegra239.
In the last message from the review, the Nvidia dev replied:
Apparently, that V2 is yet to be seen.
So, excessive detail here, just FYI.more likely that the t239 broke off into its own private branch. probably to hide it, but also probably because it's also a semi-custom part and anything else needs to be privatized
First, at least some of T239 driver development was moved to its own separate repository in 2021. Presumably, the reason this came to the mainline kernel was because it contained updates to the Orin drivers that were already upstreamed, and neither the developer nor the review have dropped anything new on LKML, so I think it's reasonable to assume that whatever internally caused the push upstream wasn't a high priority, or we'd see other work from them. It might be related to the attention on it, but who knows.
What can't happen is new code names in those mainline patches, because the code has to actually work. Nvidia has a Linux driver called
host1x
. host1x
is, essentially, the generic Tegra driver that gives the Linux kernel access to the CPU, the GPU and any other totally custom hardware blocks (the OFA, the DLA, etc) that a Tegra SOC might have, and those know what components to expose by matching on the name of the chip. Rather than manually update Windows, Linux, QNX (and presumably, Switch) drivers separately, Nvidia uses a shared driver core ("fabrics"), some of which is autogenerated from a database of info about the chips ("the control backbone"). You can see some of this in their L4T code.
In short, patches to mainline Linux can't use a different code name. However, in places that Nvidia totally controls, they can do some fuckery, and do. Nvidia's open source video drivers (separate from the mainline Linux Tegra drivers), are pulled from their (larger, more robust) internal driver code. It appears that this process is automated. When Nvidia has a specific chip they don't want details leaking about, the process strips that code from the open source version and leaves a dangling reference to a "new" code name for the GPU.
It's very subtle. T239's fake chip name is probably... T239D. You can see T234D hanging around in the driver before Orin released.
We know from "The Leak" that Nvidia uses a mechanism that prevents binaries shipping with REDACTED code names in them, but that doesn't prevent a random developer with a deadline and a stack of git commits to upstream from firing off a total business as usual email containing a reference to a chip that is already in public documentation. I'm willing to bet both Nvidia's and Nintendo's lawyers made sure they made their money's worth afterward, anyway, to check and see if a contract was violated.
We do know that Nvidia assigns different codenames to the chip they're building and the overall customer project they might be building the chip for. T210 was the code name for the chip in the Switch, but the Switch project at Nvidia was "Odin". I imagine that name is even more tied down, but could possibly have slipped out somewhere without us noticing