Comment by aseipp | Hacker News Reader

Comment by aseipp | Hacker News Reader

Parent Full thread

Holy hell, they dropped the price on that thing pretty quick! I think you could originally only get the $700 price tag for Xavier via the "developer discount", which only applied to one unit.

As a word of note, though, the chips in the Tegra machines are, IME, extremely powerful -- but you're stuck with Linux 4 Tegra and Nvidia's Ubuntu-based distro. At least for the TX1 there was some success getting other distros working (without GPU) but I remember several people running into trouble with the TX2. That may or may not be a dealbreaker for some people (it's a sort of franken-Ubuntu and your only hope for problems are the Nvidia forums)

The Solid Run LX2K machine however should (eventually) run any standard AArch64 distro. They also have 16 cores and expandable SATA, as well as an open PCIe 4x slot on these boards, so you can install whatever GPU you like onto it. Not to mention you can plug in a bunch of RAM. You can even buy a slightly different variant of this board with a 100GigE cage on it, still under $1000 USD, which puts it in a very unique position vs any other I've seen.

I think the LX2K boards are (now, at this point in time!) a much better sell if you want an actual "arm workstation" as a developer. But the Xavier will be vastly superior if you want a lower power profile (mobile) or if you're doing something like DL inference.

The PCIe slot is actually 8x. Right now we are targeting AMD Polaris based GPUs for support, just because their power consumption fits well with this SOC. AMD just announced their PRO 3200, which is really a match made in heaven. I just wish they had include SR-IOV on that model so you could get VM acceleration as well.

As for Xavier it is a good platform, but you in many benchmarks the HoneyComb LX2K is way out in front. Then adding in the 10Gbps networking and storage.

You can install whatever GPU you like, but good luck getting aarch64 drivers for it. IMO franken-Ubuntu on Xavier is pretty close to stock, save for CUDA and a few Jetson-specific tools. Certainly close enough to be practical. And you get Pascal-grade CUDA and tensor cores on the Xavier, too. And PyTorch works as well.

Why would that be an issue? People already have Mesa and AMDGPU working fine on non-x86 architectures (people have demoed physical POWER, RISC-V and AArch64 machines, all with various Radeon GPUs and the newer AMDGPU stack) and the support gets better and better every release. But it already works fine.

Also, Franken-Ubuntu I suppose isn't really my complaint, though I still don't like it. It goes beyond that, though -- the Tegra boards have very custom design that means some things are a huge pain in the ass if you want to even remotely treat them like "normal" computers (for example, want to test that fancy driver you wrote on your AArch64 Tegra machine for that M.2 peripherial? well first you gotta reconfigure the BSP firmware to turn off the power rails to various subcomponents so you can reroute them to things like the M.2 E key slots in the TX2. Gotta disable that USB controller via BSP flash if you want a chip in there, etc etc). It also means I am basically stuck on older kernel versions which may mean features I want to use or develop with are basically off limits (eBPF enhancements, io_uring, etc.) I don't just want to test SIMD; I also need to test actual memory/disk/peripherials, etc. These constraints make perfect sense for Tegra devices -- they are still annoying, for me.

I'm not saying the Tegra machines are not nice, or they are not high quality. They absolutely are. If you are just writing code to give the CPUs some work to do, they are great! Flash the default BSP, plug in HDMI, run 'gcc'. They're the highest performing ARM machines for their power profile and size, IME. But the moment you treat them like an "ordinary" computer, all bets are off. These LX2K machines act substantially more like "ordinary" desktop computers, and IMO that's a significant benefit.

Now, if you want to do DL inference or GPU programming, I think CUDA is vastly better (for a number of technical reasons). And the Tegra machines would be a good choice! But if you're just writing ARM code and need a desktop, these machines seem perfectly sufficient and can host open source drivers just fine, and will be far more usable in a number of ways vs a Tegra machine.

(And also: Nvidia also released some press notes last month that they would be bringing the full CUDA stack to ARM for data-center computing, presumably, including the host driver. So it's possible you may be able to stick a Quadro into these things soon enough and use CUDA to your hearts content!)

If what you are doing only revolve around "writing code to give the CPUs some work to do", then a $30 Rasp-Pi will work just as good.

Also 99.999% of people want to "use" software, very often proprietary one, not re-invent the wheel. No matter what, this kind of product will be a niche, by an overwhelmingly large margin.

>> these machines seem perfectly sufficient

That I do agree with. I wanted to buy an ARM machine to do CI for one of my clients, and currently the only real option for that seems to be Cavium ThunderX. And pricing is way out there in the stratosphere for what you get. Instead I could see them deploying a few of these for a fraction of the cost.

Reply on news.ycombinator.com