Looking Ahead at Intel's Xe3 GPU Architecture | Hacker News Reader

Looking Ahead at Intel's Xe3 GPU Architecture

chipsandcheese.com

119 pointsryandotsmith36 comments

I’m really hopeful for the future of Intel’s GPU pipeline. B580 proved there is an audience for their parts.

If a C770 or C970 card can prove to be a contender like B580 is at the respective price point then people will buy them.

From some of the interviews i saw at B580’s launch, it seems like Intel know the shortcomings of the Alchemist and Battlemage architectures, but weren’t able to change them before launch.

Of course these are parts that have been in the works for several years now. They’ve had time to see what Nvidia and AMD are doing with their competitive products

I hope this means they're going to continue to invest in the GPU line. Competition is good.

Looking ahead is challenging. Intel launched the B580 on December 13th, and it sold out within hours. We're still waiting on a restock.

I’ve been tracking it for weeks and it comes back in stock only to sell out minutes later. I’ve had 2 or 3 times where I thought I got one, only to get the order cancelled email.

I can’t tell if there’s just no stock, or huge demand, but either way it seems like a great product. If only you could buy it.

I suspect scalping. If you hit Amazon you'll see a bunch of b580s listed with $100 markups.

Where do they get all this information about things like registers? I thought GPU ISAs were treated like trade secrets?

Intel and AMD GPUs have public documentation and open source drivers.

While the AMD ISA documentation has been slightly improving during the last years, so the document about RDNA 3.5 is noticeably better than older documents, Intel has failed to publish the documentation for the latest Xe versions, even if in the past its GPU documentation was much better than that of AMD.

Intel still has to publish several additional manuals, to cover Xe in Meteor Lake, Xe in Arrow Lake H, Xe 2 in Lunar Lake and Xe 2 in discrete GPUs, and then Xe 3 in Panther Lake, to bring their documentation up to date.

You can extract most of the GPU information from the open-source Linux drivers and LLVM compiling back-end, but this is a poor substitute for actual GPU ISA documentation.

Does "open source drivers" imply documentation of the underlying ISA or just the tooling to shepherd it? How much code would I need to go from bootloader to quake 2 assuming I only needed to play that one piece of software on my one card?

https://www.amd.com/content/dam/amd/en/documents/instinct-te...

The compiler is LLVM

https://github.com/llvm/llvm-project/tree/main/llvm/lib/Targ...

It's all there. For the driver/runtime everything is also open and upstreamed into the Linux kernel (register map and pcie bar and etc). The packet protocol isn't well-documented but documented enough that tinygrad managed to build their own driver from scratch.

Oh wow, that's very impressive documentation. We might actually see an opportunity to move beyond linux and closed drivers in this lifetime.

https://www.intel.com/content/www/us/en/docs/graphics-for-li...

edit: they are scary huge.

That's a privilege that the dominant players in the market can get away with. If you are the #1, everyone will do what they need to to run on your stuff. But the players trying desperately to get programmers (and customers) to pay attention to them need to make things as easy as possible to use.

That didn't answer the question at all.

It answers the second question.

linux and llvm contributions are where I would look.

It seems these chips have all sort of hardcoded and heuristic knowledge literally baked into them. Disappointing.

Why can't AI come up with some kind of fast universal computation machine that doesnt have the need for the siliconized version of ifdefs?

Yep, the crowning achievement of machine complexity in recent human history is easily generalized in computational efficacy by effectively markov chain chatbots that were trained on project Gutenberg, reddit comments, and github python and javascript repos.

theres lots of ai other than llms

Ok fair point.

One such method was an improvement in matrix matrix multiplication in 2022. So I’m sure there is a potential discovery of many efficient numerical algorithms that will be uncovered with machine learning.

https://www.nature.com/articles/s41586-022-05172-4

My parent comment was a knee jerk response to the over confidence and wishful thinking of computer work people to a magic AI box solving all the hard physical problems like its some simple conversation prompt away.

Unfortunately I think the fetishization of prompting is a self-fulfilling prophecy. Over time these models will be able to understand enough of the physical problems to have a dumb prompt yield a reasonable if not optimal solution. But of course that house of cards collapses when fundamental advances appear.

If you're not smart enough to make use of a optimal or even acceptable solution, the quality of the LLM answer doesn't matter.

Human brains have lots of "hard wired" heuristic knowledge literally baked into them. See for example https://www.sciencedirect.com/topics/computer-science/ventra...

Is it surprising that silicon does too?

> Is it surprising that silicon does too?

Given that human brains and silicon has nothing in common, then yes, it is surprising. What makes you think otherwise?

If evolution has worked out that having specifically optimized pathways for specific actions it doesn't seem terrible that human designed machines are similar.

I feel that's assuming a whole lot of unrelated things, such as machines and humans having anything at all in common.

I get that your take qualifies as stoner philosophy, so I won't push too far, but really. Do give some concrete examples to support your ideas rather than GPT fluff.

This is, like, the entirety of computing. Did x86 spring fully-formed from Zeus's forehead or did people bake in their knowledge of what programmers wanted their programs to do?

When you write programs do you not use heuristics? And tests?

It’s a little known fact that x86 is literally Athena! Which I guess means Intel should be casting some lightning bolts towards AMD and others.

Do you mean FPGAs?

Because fixed silicon inevitably has fixed "baked in" choices.

AI doesn't "create", it "modifies" existing data.

It regurgitates whatever was in it's training set.

Don't we all :)

Knowledge is significantly different from a training set. In particular we can hold incomplete knowledge that is eventually resolved through further practice and observation. Also knowledge access is analog and inaccurate to a degree that often introduces emergent behavior into outputs.

If you’re lucky.

Isn't a collage a creative work?

Reply on news.ycombinator.com