Intel’s Ponte Vecchio: Chiplets Gone Crazy
chipsandcheese.com
chipsandcheese.com
I suppose one unknown catalyst with intel is what happens in taiwan/china. If things get crazy over there, suddenly intel seems alot more valuable as the 'US' chip maker (they produce roughly 75% in the US iirc). If the gov starts to even more heaivly subsidize non-reliance on asia, intel could find major gains if TSMC/samsung get shut out.
I mean, just look at the market caps- Intel is worth 6x less than nvidia despite historically having the same or greater gross revenue (not counting the most recent quarter of course).
Intel is also well known for inking long-term deals with major discounts for big customers (Google, Facebook, etc.) that can commit to purchasing large amounts of hardware, whereas Nvidia doesn't really have the same reputation. It's conceivable that Intel could use this strategy to help bootstrap their server GPU business. The Googles and Facebooks of the world are going to have to evaluate this in the context of how much additional engineering work it is to support and debug multiple GPU architectures for their ML stack, but thinking long-term/strategically these companies should be highly motivated to negotiate these kinds of contracts to avoid lock-in and get better discounts.
How many of us learned ML using Compute Sticks, OpenVINO and OneAPI or another of their libraries or frameworks, or their great documentation? It's like they didn't really believe in it outside of research.
What irony is it when a bedrock of "AI" fails to dream?
Language models in particular are very similar architectures and effectively a lot of dot products. And running them on GPU's is arguably overkill. Look at llama.cpp for the way the industry is going. I want a fast parallel quantized dot product instruction on a CPU, and I want the memory bandwidth to keep it loaded up. Intel should be able to deliver that, with none of the horrible baggage that comes from CUDA and nvidia drivers.
I'm thinking about the widespread confusion they caused by having different CPUs support different subsets of the AVX-512 ISA.
Now while NVidia is making standard C++ run on CUDA, Intel is still having SYSCL and oneAPI extensions.
Similarly with Python and RAPIDS framework.
Intel and AMD have to up their game for the same kind of developer experience.
SYCL isn't pure C++, meaning writing STL code that goes into the GPU, like CUDA allows for, nor requires the hardware to follow C++ memory model.
Nope. Intel doesn't get "It's the software, stupid."
Intel is congenitally unable to pay software people more than their engineers--and they treat their engineers like crap, mostly. And they're going to keep getting beaten black and blue by nVidia for that.
The more disappointing thing for me is that they bought like 5 AI startups pretty early on and have basically just shut most of them down. Maybe that was always the plan? See which ones develop the best and consider the rest to be acqui-hires? But I think it's more likely just fallout from Intel's era of flailing around and acquiring random crap.
For a company that makes a truckload of money from selling CPUs it's unforgivable their tools are not
a) free
b) top-notch
I know one is limited by how good Windows is when you ship a tool for Windows, but your description is quite horrifying.
They just bad at it.
Some of the wafer fabs are in the US, but most of assembly gets done in Malaysia
https://en.wikipedia.org/wiki/List_of_Intel_manufacturing_si...
I've been hearing this for years though.
Err, is the ISA undocumented/impossible to inspect in the execution pipeline? Seems like an important thing to verify/fix for a hardware benchmark...
> We weren’t able to get to the bottom of this because we don’t have the profiling tools necessary to get disassembly from the GPU.
I mean OpenVINO's perf is up there on Intel CPUs and it's a great optimising compiler, I've thrown a lot of weird stuff in there and it didn't crap out with complaints about unsupported layers or unsupported combination of layers. It also has an OK batching story (as opposed to TVM last time I checked...) if you're ready to perform some network surgery.
I also feel it's very bad at reporting errors, and stepping through with gdb is one of the worst experiences... BUT but yeah most of the code is available now.
Now if they could stop moving shit around, and renaming stuff, it'd be great. Hoping they settle on 'OneAPI' for some time.
[0]: RDNA2 ISA: https://www.amd.com/content/dam/amd/en/documents/radeon-tech...
This sounds vaguely interesting but I am not holding my breath.
Intel's strategy here is baffling to me. Rather than keep trying to improve their existing line of coprocessors (and most critically, keep accumulating key talent), they kill off the program, scatter their talent to the four winds, wait a couple years, and then launch another substandard product.
Just look at how they dragged their feet in transitioning to EUV because it was too expensive. This contributed to large delays in their 10 and 7 nm processes and a total loss in their process leadership.
And look at how many billions they poured into making a 5G modem only to give up and sell their IP to Apple.
Or how they dragged their feet in getting into mobile, then came out with Atom way too late to be successful in the market. They essentially gave the market to ARM.
Optane is another recent example. Cool technology, but if a product is not a smashing success right away, Intel throws in the towel.
There’s no real long term vision that I can see. No resilience to challenges or ability to solve truly difficult problems.
They also had the best ARM chips for years with StrongARM/Xscale (using their own cores). Which they killed because obviously Atom was going to be much better and lock in everyone into x86...
Wasn't the actual (partial) reason that they didnt have a place to actually create them since Micron sold the fab?
https://www.extremetech.com/computing/320932-micron-ends-3d-...
I’m not saying that Optane was a hill they needed to die on, but it’s just another example of their failed leadership and decision making.
Look at how AMD is pursuing and largely succeeding with their vision of using chiplets in their CPUs and GPUs to enable significantly higher core counts at a lower cost.
Or how Nvidia is innovating with massive AI supercomputers, ray tracing, and DLSS.
What is Intel’s vision? In what way are they inventing the next generation of computing? It seems to me that their company objective is just to play catch up with AMD and Nvidia.
And TSMC. Intel really wants to win in both the fab game and the chip game.
They're probably just doing what their shareholders want. Unfortunately, shareholders are shortsighted and risk-averse, contrary to the common rhetoric of being risk takers to justify eliteness.
Surely leadership could be embroiled in lawsuits were they to actually care more about the company than their weak, whimsical, and often incompetent shareholders. Kind of a sad irony actually.
If I was a betting man, I would bet that this project is dead inside 36 months. And if I was a GPU designer, I'd accordingly not touch Intel with a barge pole. They've painted themselves into a corner.
I personally know GPU experts who left Intel for Nvidia because of this. I can't imagine they would consider going back at this point.
You see similar behavior in many failing companies, as well as third world countries. You can't admit faults to iterate, and you need grand new initiatives.
But Intel has been doing GPUs for a very long time however and it doesn't realistically seem like they are going to stop anytime soon. Discrete-class and Datacenter class GPUs are new for them, but hyperscaler space is a place that's staying hot and one they're familiar with. Nvidia literally can't sell H100s fast enough. So, I suspect they'll probably remain in the "GPU accelerator" race for quite a while yet, actually.
cough An expensive lesson, I’m sure.
Being dethroned and free cash flow negative is rather bad I am told.