Hacker News Reader

Hacker News

Comment by giuliomagnifico | Hacker News Reader

Parent Full thread

giuliomagnifico

291W... whoa!

View on HN

Intel has really lost it. My Ryzen 8 core CPU is easy to cool with a quiet set of fans.

Cooling this for real work will require a custom watercooling set up...and it isn't even as fast as a Ryzen 8 core CPU.

> 291W... whoa!

Truly a rocket. All it needs is a nozzle and some reaction mass. Maybe they should pivot?

"Intel exhaust outside". Do some computations and get thrust for free!

And 104*C without overclocking! Sounds like they’re pushing the envelope.

They should drop in kettle stand as a free gift.

Could be a good upgrade for the KFConsole -- turn the chicken warming chamber into an actual air fryer.

Seriously, it's basically a 700$ barbecue at this point. What's the Tj on this CPU ? 140 ?

ctrl-f 104 didn't bring up anything

Yeah, that’s getting into space heater range.

Edit: Temperature peaks on some workloads up to 100c. Jeez

we simply need a radiator or exchanger design to plug in to our domestic water supply at least I'll be very happy with either free compute or free hot water pov depending.

Good thing nobody actually uses AVX512

AVX-512 software is appearing. For example, for neural net inference: https://NN-512.com

AVX-512 is a pleasure from the programmer's perspective, much nicer than writing software for GPUs or for earlier SIMD instruction sets like AVX2/SSE

Versus a GPU, the AVX-512 machine model is just much simpler, and there are no drivers or proprietary compilers

Versus earlier SIMD instruction sets, AVX-512 provides all the instructions you want (permutations, floating point conversions, etc.) and they're all masked so edge cases are easy. AVX-512 Foundation is a really nice instruction set and it has a bright future

Isn’t the issue with AVX-512 is that different CPUs support different instructions? https://en.m.wikipedia.org/wiki/AVX-512

It might eventually get normalized but atm it’s a bit of a mess.

It's a bit less messy if you ignore the GPU-like Xeon Phi (Knights Landing and Knights Mill) and the failed Cannonlake that Intel definitely would prefer everyone pretend never happened. But that still leaves you with a progression of four levels of expanding support, and the Cooper Lake mess of having BFLOAT16 but not the stuff added in Ice Lake and Tiger Lake.

If you don't cut down the support table to eliminate irrelevant niche products, it's most of the way to being a Gray code.

I think Icelake and Tiger lake not having BF16 is going to be a big issue especially for machine learning applications, I can’t find what Rocket Lake supports yet but I’m guessing it’s going to be similar to Tiger Lake.

There is also the big unknown of what kind of support AMD will bring to the table.

I think Intel is banking on people using OneAPI more than writing AVX-512 code using intrinsics or compiler magic.

At that point compatibility might be not as much of an issue if you ignore potentially drastic differences in performance.

If you limit yourself to the Foundation instruction set (AVX-512F) you get a lot of benefit and maximum compatibility. That's what https://NN-512.com does

Interestingly enough, the AVX512 test was the easy win for Intel.

Nobody uses it because it wasn't available and, when available, wasn't much faster overall. Now it is.

Who both buys consumer desktop CPUs (without ECC) and runs software that can use AVX512? Games won't use avx512 (not really simd math heavy work on the CPU, and also consoles don't have it). Web browsers won't use AVX512 (similar to games - not really SIMD-friendly raw math computation limited). Video & photo editing maybe will, but GPU acceleration is already more widely available & faster than avx512.

What else does that leave?

<-- this engineer's company workstation uses a consumer desktop CPU

Most scientific and engineering computing applications use double precision floats.

How many of those workloads would benefit from it if it were more available?

I'm not sure, but we haven't seen it being widely deployed just yet.

Is it though? The vast majority of PCs even on the enthusiast level aren't going to have coolers that can handle 300w for any sustained period.

From eyeballing the chart, they'd only need to handle the load for 20% of the time compared to non AVX512 CPUs to be equivalent.

Anything more is an advantage.

iirc wasn't until recently available in the HEDT line, but the user base is equipped with scalable xeon boxes not awfully constrained desktop chips.

I'm imagining its availability will have more impact outside the user base that's already well served by Xeon Scalable.

Visicalc didn't make a big impact in mainframes.

37ef_ has covered why AVX512 is plenty in use internally everywhere I know. Shrink wrapped products, not so much.

To be fair that's the peak under AVX-512, and looks like early firmware that might not do all the power management right yet.

Still, I can't get my 5950X to ever come anywhere close to that, even running AVX2 Prime95 with max overclocking tops out at 190-ish IIRC.

4.6 GHz all core with AVX512 surprised me. At least AVX512 downclocking seems to be a Skylake problem.

Ice Lake (which Rocket Lake is a backport of) similarly doesn't downclock much: https://travisdowns.github.io/blog/2020/08/19/icl-avx512-fre...

The 100+C temperatures suggest it probably should downclock a bit, at least on 14nm.

Reply on news.ycombinator.com