Hacker News Reader

Comment by bullen | Hacker News Reader

I would say Atom is Intels best bet at a ARM competitor, not at the same power but as a slightly beefier alternative.

From what I read they are canceling that line and instead offering low power regular desktop processors.

That's why I bought 4x 8-core Atom servers at 25W, so I don't get GPUs in my low electricity servers / high performance alternative to Raspberry 4.

RISC-V is probably going to take a decade to reach the kind of stability the latest Atom line has in terms of support and performance, if ever; specially on the server-side.

The Xeon line CPUs consume way too much power to be an alternative for home hosting with lead-acid backup.

View on HN

Intel didn’t so much kill the atom line as merge and evolve it.

They are following in ARM’s footsteps by having heterogeneous multi-core CPUs with high performance and low power cores (like the M1 and ARM’s big.little before that)

Intel’s Adler Lake will have a mix of Golden Cove cores for performance and Gracemont cores for power efficiency. Gracemont cores are improved Tremont cores previously found in their Atom line.

Atom is an amazingly power inefficient design.

X86 here is a definite dead weight on its neck. Atoms front-end if few times bigger than the back-end.

Somehow I doubt this wide-sweeping affirmation, are you sure? Atom is really big family by now, starting from Silvermont it vent full out of order with all bells and whistles, I find it hard to believe OO backend is anywhere near the size of x86 insn decoder.

A complicated ISA like the x86, with layer upon layer of additions, costs a lot of silicon that will need power to run.

Let's go see some figures!

"Empirical Study of Power Consumption of x86-64 Instruction Decoder"

https://www.usenix.org/system/files/conference/cooldc16/cool...

From the conclusion:

"The result demonstrates that the decoders consume between 3% and 10% of the total processor package power in our benchmarks. The power consumed by the decoders is small compared with other components such as the L2 cache, which consumed 22% of package power in benchmark #1. We conclude that switching to a different instruction set would save only a small amount of power since the instruction decoder cannot be eliminated completely in modern processors."

1. They are physically huge, and that means more leakage, when there could be none.

2. Inefficient front-end taking toll on everything behind it contributes to the rest of the chip doing nothing except dissipating power.

3. While cores on modern SoCs take much less space than everything else usually, front-ends still eat into the die area.

4. Front-end dictates how the rest of the core behind it is designed. I believe that X86 technical debt absolutely must exert its toll on that, and overall transistor count in the back-end.

It's big, but not overwhelmingly so. To eye is looks like it's about 10% of the area so it jives with the quote provided by minipci1321 which claims ~10% of power.

Here is an annotated die shot of Zen2 that shows it. https://forums.anandtech.com/threads/annotated-hi-res-core-d...

Scaling RISC-V will be orders of magnitude easier than anything of the X86 lineage. It’s a new clean modern design with so much less complexity and legacy baggage. Intel could build crazy fast RISC-V designs within a few years at most if they want to.

A very interesting path back to relevance for Intel would be to design proprietary absolutely crazy fast RISC-V cores and market them in the data center and HPC markets first. The tooling on Linux is about there. Eventually they could get an initially small but very loyal developer laptop market going too. A good RISC-V design could easily compete with the Apple M1.

Get enough momentum and a Windows port could happen. Then you could reboot the PC market.

The automotive and drone niches would be interesting too. There is a market there for low power high performance chips for things like driving assist and autonomy. A new instruction set wouldn’t matter much since the whole software stack is more or less custom anyway.

At the very least it would be a way to hedge in case X86 really does start to enter a death spiral.

If it's so easy to create fast risc-v cores then why hasn't it been done?

Nobody’s done it yet, probably due to lack of perceived demand.

Don’t get me wrong. It’s not so easy that a few college kids could just whip one out. A company like Intel or AMD could definitely do it.

It's not exactly easy -- more of "we know how to do it, given funding".

Designing any fast modern chip -- such as the Apple M1 -- probably still costs around a billion dollars.

In the RISC-V world the company with the most funding so far is SiFive and they've had $190 million.

Intel getting involved could change that very quickly.

> That's why I bought 4x 8-core Atom servers at 25W, so I don't get GPUs in my low electricity servers / high performance alternative to Raspberry 4.

How much did you pay? Last time I checked these were still a couple of times more expensive than Raspberry Pis.

~$500 per board, but then you need to add PicoPSU, passively cooled case, SATA drive, 12V PSU, 2x 16GB RAM, so about $1100 in total

Same Gflops/W but about 4x Watt so for things that cannot be distributed (like a shared MMO world in my case) it's probably as good as it gets...

Those boards have proprietary formats. Are you running them in supermicro chassis?

A2SDi-8C-HLN4F are Mini-ITX so they fit in these: https://streacom.com/products/db1-fanless-mini-itx-case/

But you need to buy this GPU adapter: https://streacom.com/products/db4-gpu-cooling-kit/ (for the small CPU)

And buy M3 female-female spacers that are 50mm long and some M3 thread that you cut at the appropriate length.

So it's a bit fiddly until they release proper small CPU adapters for these.

Also to remove the heatsink that comes on the motherboard you need to run the board without fan until it heats enough to soften the heatpaste. Be careful to not rip the board when removing it, wooden icecream stick to bend might come useful.

Which servers/boards did you buy. I’m interested in one for local dev work as over a year it’s probably cheaper than renting a suitably large VM.

SuperMicro

Thanks will look into it.

Reply on news.ycombinator.com