Hacker News Reader

Comment by rektide | Hacker News Reader

There's a "Ryzen AI" on these AMD cores that is really a modest sized fpga, supposedly. Linux has been landing some support for interfacing with this- calling it XDNA (on a new "CDX" bus) - but no signs yet that users can program it.

Would be a fun thing to get exposed to users. Fingers crossed that actually happens; I could definitely see this being very vendored up, that it ends up only really running a couple predefined ML workloads.

View on HN

Am I reading this right?

* There's a decent built-in gpu

* Or I can plug an external gpu into the back

* Or I can rip out the mainboard on an old Framework and pop this one in

* Or I can pop this mainboard into a Cooler Master case or some rando 3D printed case and have a PC which can sit on my desktop, underneath my TV, etc. (Granted this thing probably won't have the horsepower of a real desktop but for a lot of workloads it's probably be fine)

The original Framework had a great philosophy with limited execution. This seems like a whole new level of utility, and perhaps just the beginning of what an ecosystem built around the Framework could become. The question you usually have to ask about any product is why would I buy this? But in this case I'm struggling to come up with good reasons NOT to buy. I already have a laptop, a desktop, and a living room PC. If I can potentially refresh the Framework to fulfill ANY of those roles what excuse do I have to not to own one?

Yes, decent iGPU. But no on an external GPU in the back since this is for the 13" which has no modular bay for internal GPU (They mean a USB4 -> PCIe dock for the external GPU).

There IS a 16" coming that will let you plugin a dedicated GPU.

I will probably order this (or it's next version) one for my machine in 18 - 30 months since mine is only ~6 months old.

edit: and apologies if I mis-understood or am being pedantic

You can plug in an eGPU with this.

From the post:

> If you want to use an eGPU, you can do that too! This is because the Framework Laptop 13 with Ryzen™ 7040 Series processors has two fully capable USB4 ports, with the back left and back right Expansion Cards slots.

Right, that's exactly what they said

> no on an external GPU in the back since this is for the 13" which has no modular bay for internal GPU (They mean a USB4 -> PCIe dock for the external GPU).

Unless the Mobo doesn't support it, you should be able to use external GPU's even with Ryzen

This makes a lot of sense, I was just waiting for their acquisition of Xilinx to materialize in their consumer products in some way. This seems a great way to unlock e.g. some ML potential. The possibilities seem endless.

I think AMD already acquired Xilinx?

Yes. But the question remains of when & how we see Xilinx technologies/capabilities boost AMD's core businesses.

Exactly. Just like we can see Nvidia’s acquisition of Mellanox starting to make sense now that we see Nvidia shipping entire clusters of hundreds or thousands of GPUs connected through InfiniBand, I’m waiting for what AMD’s acquisition of Xilinx has to offer.

You're already seeing it. Xilinx led the industry in chip interconnect technology, SERDES, MCMs, die stacking... they've been making chiplet designs since 2011. They were the ones inventing all the interconnect and packaging technology AMD is using in Zen, Epyc, and now Radeon products.

This sketchy-looking link is a marketing slide for Xilinx chip-to-chip interconnect tech from 2019. It's illustrative of the kind of IP AMD was shopping for. https://146a55aca6f00848c565-a7635525d40ac1c70300198708936b4...

Having fpgas on-board feels like something that would have been a game-changer even pre-ai. Imagine if your Ryzen chip could do hardware encode/decode of h266 the moment the spec was finalized.

I can't find evidence it really made it out to consumers, & even less evidence there was any followup, but Intel did have a Xeon + Arria 10 GX FPGA chip, back in 2018: Xeon Gold 6138P with Arria 10 FPGA. https://www.anandtech.com/show/12773/intel-shows-xeon-scalab...

At the hyperscaler level there's been a lot of adoption of SmartNICs, which are often FPGA based. That makes me feel like there's a pretty big market here already for integrated FPGAs. I think the other requirement is that these FPGAs need to be made broadly available if we're going to see a real shift. Selling a couple many-thousand-dollar high end chips isn't going to create new audiences clamboring for these things, which is required for the tech to actually get interesting & well adopted, instead of existing only in a small but important niche.

Follow-on question, what makes FPGAs so dang expensive?

Size, speed, features. You can have a small FPGA for about 2$. You will however be limited in the complexity of the design you can implement because it has limited gates and pins. And limited speed. Like normal processors if you want faster, bigger and more features the cost shoots up. Combine that with the limited market for big FPGA's and they can be very high in price. Mind you that many also have things like a multi core ARM processor on board since many designs need a processor to control the circuit. Same with many interfaces. Would be a shame to use the limited gates in the FPGA for standard things like that.

Large die area and low production volumes.

Bunnie Huang already built himself a laptop with on-board FPGA back in 2012.

[1] https://www.bunniestudios.com/blog/?p=2686

[2] https://www.bunniestudios.com/blog/?p=3597

Source (CES 2023): https://www.youtube.com/live/OMxU4BDIm4M?feature=share&t=137...

Textual rehash: https://www.tomshardware.com/news/amd-brings-chiplets-zen-4-...

Keep an eye on the ONNX runtime[0]. Microsoft has gotten all sorts of weird AI accelerators to work, from OpenVINO to the Rockchip NPU.

[0] https://onnxruntime.ai/

Reply on news.ycombinator.com