Intel shows off 8-core, 528-thread processor with 1TB/s of co-packaged optics | Hacker News Reader

Intel shows off 8-core, 528-thread processor with 1TB/s of co-packaged optics

theregister.com

84 pointsBender27 comments

Related discussion from a few days ago: https://news.ycombinator.com/item?id=37315802

I wonder how good this sort of chip would be for (parallel capable) compilers. They have lots of branchy cache-missy sort of code and can often scale up to a lot of threads if they're compiling a big problem.

potentially...really good?

however there are lots of pitfalls. if your thread creation and teardown is expensive, then that limits the grain to which you can apply threading (Amdahl). if synchronization is expensive then you have a similar problem (that potentially gets worse with contention).

one of the big problems is exactly _how_ to map threads to workloads. you don't want to leave idle resources on the ground, but its also very counterproductive to generate a massive work queue.

there are also limits - memory transaction concurrency and memory bandwidth, and the costs and contention of going off chip.

cache coherency, while it seems nice from a programming model perspective, can also really limit how much concurrency you can exploit.

I know there are others here that can add to this list.

Yes, all the usual scaling issues, of course it assumes a compiler that doesn't spend all its time contending on one big lock or sitting in cache coherency because it's doing an atomic increment of a single shared counter. But once you have the hardware it becomes reasonable to spend time fixing those issues. They're not fundamentally hard, compilers are mostly semi-functional transforms over large data structures.

In case anyone missed towards the bottom of the article:

> These design considerations drove the team to develop this experimental processor, which TSMC builds using its 7nm FinFET process (in case you didn't know, Intel fabs a lot of non-CPU products at TSMC and has for years) and which features eight cores each with 66 threads.

>custom RISC

Based on RISC-V? Totally bespoke?

Oh great what does this futuristic chip do?

Instead, it was purpose built using a custom RISC architecture for DARPA's Hierarchical Identity Verify Exploit (HIVE) program. That US military initiative seeks to develop a graph analytics processor capable of churning through streaming data 100 times faster than conventional compute architectures, while also consuming less power.

Oh its for the military. Whats Graph Analytics?

Graph analytics deal with how data points connect with other data points in complex systems. The example Intel principal engineer Jason Howard pointed to in his Hot Chips presentation was social networks, which might run graph analytics workloads to understand connections among members.

So its a chip to analyse social networks at 1TB/s

Social networks are an example that is east to describe to people. But that doesn’t mean they’re what the chip would be particularly used for. I don’t think you should read much into the examples that are used here.

The previous thread had some examples. But the workload you should imagine is one where you are doing some kind of big parallel computation where each thread does a lot of unpredictable memory accesses. So various graph operations fit this, as do various database queries.

https://www.darpa.mil/program/hierarchical-identify-verify-e...

https://www.dote.osd.mil/News/What-DOT-Es-Following/Followin...

Darpa themselves say they want it for social media analysis, IP probe detection and infrastructure monitoring

I predict the most interested parities will be in insurance. The quicker an entity can mitigate libel the better, for them.

Is Intel finally back on track? There were some rumors that Nvidia is considering to use Intel fabs.

They are rumoured to have booked the first EXE (latest and greatest) machines from ASML. If that's true, then yeah, even Apple might go back eventually. Just for the fabs, of course. Not use Intel CPUs. I believe this is what Gelsinger meant when he said he would like to win Apple back

I really hope Gelsinger used that as an expression, not a real goal or strategy.

Apple using an Apple chip, given the recent past and circumstances is both a technical and a political move. Going back to Intel ( or AMD ) would be an humiliation and Apple's track-record dealing with inferior products has been basically ignore the problems and put their spin machine to use.

I hope Intel gets better but I also hope they don't get much better than anyone else because the last time that happened their arrogance and anti-competitive moves were so bad even their most "loyal" corporate customers HATED them with a passion.

> Apple's track-record dealing with inferior products has been basically ignore the problems and put their spin machine to use.

Taking a broader view, I'd say that went 'screw this, we'll do it ourselves.'

The article says TSMC built the actual chip part

This intel one is 7nm and apple bought every 3nm capacity this year. I feel like apple is beating intel on chips.

This isn't a production unit, but a research project to play with silicon photonics.

TSMC is beating Intel on chips*

Yeah, also “7nm” and “3nm” aren't actually measuring the same thing (IIRC Intel 7nm is equivalent to TSMC 5nm)

well if the measurements are only marketing words now, then intel should just measure something else and say that they have 2nm

That’s exactly what Intel are planning to do

https://arstechnica.com/gadgets/2021/07/intels-foundry-roadm...

Yeah kinda, they basically did a rebrand recently.

Intel 7nm and TSMC 3nm are roughly equivalent, they both have a transistor density of 200 - 250 MTr/mm2.

No, there is no Intel 7nm anymore, it was renamed to Intel 4. But there is instead an Intel 7, which is the previous Intel 10nm ESF.

This marks the point where they changed their naming scheme, so that Intel 7 is roughly comparable to TSMC N7 (not to TSMC N3, which would be Intel 4).

And this chip is on TSMC 7nm per the article, not Intel 7 or Intel 4 (the former "7nm").

Irrelevant since this is TSMC 7nm.

Apple only had 90% of TSMC's 3nm capacity this year, with Intel contracted for the other 10. They had delays, so TSMC is producing fewer wafers.

Edit - per the article, this chip is actually fabbed by TSMC on their nm process.

That was the rumor, but there's more recently been reports that Apple has taken up all 3nm capacity.

https://www.extremetech.com/computing/apple-bought-all-of-ts...

Reply on news.ycombinator.com