Hacker Read top | best | new | newcomments | leaders | about | bookmarklet login

>The main problem with the Groq LPUs is, they don't have any HBM on them at all. Just a miniscule (230 MiB) [0] amount of ultra-fast SRAM [...]

IDGAF about any of that, lol. I just want an API endpoint.

480 tokens/sec at $0.27 per million tokens? Sign me in, I don't care about their hardware, at all.



sort by: page size:

> Groq sells their dev kit for $20k even though a single LPU is useless.

I find this a very questionable business decision.


> is there something I'm missing?

Affordable, mature hardware. The Talos systems are starting at 3k for a quad core CPU on a micro ATX board. Same thing happened to MIPS and Sparc. Performance and technical merit mean nothing vs cheap and ubiquitous hardware. It's a lot of cash for a what amounts to an experimental toy. They are also a bit finicky as a friend bought one from another dev that refuses to post for unknown reasons. So there is risk involved too, no one else is making these boards.

The performance gaps and architectural features that made these chips matter 20 years ago have been closed by commodity off the shelf x86 hardware and various Arm CPU's are eating everything.

The only reason Risc-V matters is that no one has to pay for licenses.


> How about providing a single board computer like a Raspberry Pi

The issue is more about connection to external servers than the particular board itself. A product that both runs great out of the box, and is also hackable to the point where it doesn't 100% rely on the original company being around to run is the sweet spot I'm thinking of. I'm mostly just thinking aloud about what features the product needs to have in order to hit that sweet spot.

> That said, I'm not buying any IoT devices, self-hosted or not, in the foreseeable future

That's cool, but this isn't about the average HN user. Consumers want to, and indeed they do, pay for these devices. I just want a nice model where these products can be sold while not leaving the customer stuck if the original company goes under.


> but they are not big enough consumers to benefit economically from owning chips.

AWS Graviton 2 (Arm) would like a word about that with you. https://aws.amazon.com/ec2/graviton/


> they're so ashamed of the pile of spaghetti code they're peddling to their customer that they just can't OpenSource it for fear of losing face and reputation for ever.

LOL no. Everyone knows how this particular sausage is made. It's obvious we are talking of an organically growing decades old codebase. It won't surprise anyone.

To be more on topic, ever since I learned of the PicoEVB a few months ago, I am just giddy of the possibilities, it's amazing we finally got a mobile and PCIe connected FPGA. I wasn't even looking for one because I haven't thought it possible. We truly live in the future.


>The cost is tiny for enormous security.

Security and x86? I would maybe not hate this if there were a stipulation that these companies must move beyond x86 with a completely new architecture, something more reasonable for the modern era. Theres too much legacy crud that just muddies the waters when trying to learn how to program these contraptions. Maybe some way to completely disable speculative / out-of-order execution. What are WE getting out of this? By WE I mean non-giant-chip-conglomerate shareholders/welfare queens.


> An X200 is also an ancient machine. Do you value your own time so little? Clearly you are annoyed by the compile times?

I value my security first and foremost. My X200 has zero binary blobs running on the machine, it also has several other properties that I appreciate.

> You can get a $200 CPU that will blast through a kernel compile in 3 minutes.

That assumes that I even have the money to afford that. Not all of us live in silicon valley.


> Couldn't they have fit that within both budget and size?

Oh they could have. The problem is: it's extremely hard to get access to powerful SoCs - the vendors simply won't work with you and most of the documentation is under NDA.


>I wonder what the thinking was that lead Pano Logic to put expensive FPGAs

Cloud Computing got really hot in 2006, it was perfect time to milk suckers. https://www.crunchbase.com/organization/pano-logic mo was finding bigger idiot before running out of coal.

Reminds me of another 'we will make it up in volume' dotcom scam I-Opener https://en.wikipedia.org/wiki/I-Opener selling ~$500 worth of computer and LCD monitor at $99


> why is it so cheap

Using a high volume Amazon chip. Sadly the software stack is not what this thing deserves. It's a full blown DPU used with something very minimalistic SW wise.


>haven't been running their own LLMs for the last five years

Because the hardware has not existed.

This said by accident I've seen hardware that was brought to a testing company by federal marshals that was massively parallel custom hardware that was likely for signal processing a lot of channels at once. So there is plenty of custom hardware out there, but these items have not been produced at the scale needed (from what anyone can tell) and, again from what we can tell, they don't have the general processing capability that GPU/TPU driven LLMs have.


> However if I was a customer it would have me questioning the capability of and their commitment to the current hardware.

If I was an investor, I would not be very happy if they did no research at all in this area.


>> They are nowhere near as flexible as a microcontroller is, and reprogramming them for a different purpose is a huge investment from a development standpoint.

Isn't that the point of LiteX and Migen though? To reduce the barrier to entry in time and budget? Before the RPi came along SBC's were fairly expensive, and the RPi was only able to offer a product at that price because of scale.

And here we have an FPGA based video processing platform for only $350 produced in low-volume. https://hdmi2usb.tv/numato-opsis/

I'm not saying you're wrong, I just don't think your argument is as much of a slam-dunk that you think it is.


>> The thing about that is they could minimize SKUs even more by providing socketed ram and storage.

Okay, but then they’d lose money by not selling their proprietary parts.


> VThe problem is with building a competitive stack, Huawei can't get access to a 7nm fab in China

yet

> Huawei can build 7nm chips at TSMC, but due to their ARM l

ARM -> RISC, as ARM is not the only game in town. For example, at the moment I am having a lot of fun with a Unielec u7621 board (2 mini PCIe, mSata and USB3 with hardware accelerated NAT to handle 1Gbps) that costed me about $45 : https://openwrt.org/toh/unielec/u7621-06

I would love to see MIPS64 SOCs with a few GBs of RAM to replace my raspberries


> To be fair, POWER is an open standard, and there's absolutely nothing stopping someone like Linode, DigitalOcean, or Hetzner from offering POWER-based systems at a smaller hourly price.

Well besides the fact that 1) IBM is the only party with both the capabilities and interest in making high performance Power based products [1], and 2) evidently does not understand how to (or why to) invest in bringing this to a general audience. I don't really care if they would do this in IBM Cloud or with other cloud infrastructure companies, but they don't seem to be doing either. In addition 3) why would other parties be interested in running Power when they can run amd64 or arm? It certainly doesn't look to have a price advantage...

IBM really needs to shepherd Power well - they're the only ones that can do it. But I can't help but thinking they seem to be leading it to the grave despite apparently very capable engineering.

[1]: And why would anyone but IBM go for Power at this point if they can have ARM too? An open license matters very little when compared to ARM's mindshare and momentum.


>He's reviewing the board as is and not assuming future progress that may or may not take place in some undetermined time.

Absolutely, but he does (for the cryptography example) mention that future chips will do better in this regard.

Why would he do this, and not mention the cryptography engine present in this very SoC?

What's most likely is he managed to overlook this SoC has this hardware.

>I want to buy a board for what it can do now

This board clearly isn't for you.

>kernel updates that are not guaranteed.

Having kernel patches already sent upstream for review[0] is a much better situation than potentially having nothing.

0. https://rvspace.org/en/project/JH7110_Upstream_Plan


> Did you want to buy such technology?

Nope, just merely curious

For most of the use cases I can think of that these would suit, Infiniband seems like it would also work and that's fairly widely available already.

That being said, there's probably use cases this would suit better. They're just not coming to mind easily. :)


> In fact, I'm currently supporting a product shipping with the RPi 3B in quantities large enough to qualify as "real quantities".

How did you get pricing and volume from Broadcom? I'm genuinely interested as that stopped a project I was working on completely cold.

We simply couldn't get them to care about 1K-2K parts.

next

Legal | privacy