Cache splash in Telum means seventh heaven for POWER11? | Hacker News Reader

Cache splash in Telum means seventh heaven for POWER11?

talospace.com

Previous discussion about the cache structure (not in the context of POWER) here: https://news.ycombinator.com/item?id=28401737

I was really surprised by this cache design, in a good way. It's a very interesting innovation, one of those that is simple once you hear it but nobody (as far as I know?) has tried this approach yet. I think it's very clever for a large distributed system like the Z mainframes;

It would be fun to build some Network on Chip FPGA designs using this approach to model how it would perform!

One thing I wish modern multi core chips had was some way to just directly transfer core to core. That would be really interesting, though I suppose it might be really hard to make it play nice with OS abstractions. Might only be useful in specialized HPC applications.

Well.. you can play cacheline ping pong between cores today! No going to DRAM, straight from core to core.

But I'm curious what sort of algorithm you have in mind that needs to spend a lot of time moving data between cores. Maybe some sort of algorithm that needs to alternate short steps of aggregating intermediate data and re-distributing the result to every core, in a loop?

So, the shape of the problem that would fit this solution is not obvious to me :)

Embarassingly parallel algorithms are the ideal. The less communication, the less Amdhal's tax costs you.

Some general purposes CPUs do have stuff like that, but apparently it's rarely used and often buggy due to neither being well tested nor going through the usual cache protocol.

GPUs have this with warp instructions

Everything old is new again. In the 90's Kendall Square Research built modular supercomputers based around using nothing but cache.

https://en.wikipedia.org/wiki/Kendall_Square_Research

This is really awesome, thanks for the link!

Is there any chance this will reach normal mortals? Looking at the Basic Talos II motherboard and CPU bundle, that's already 3480 USD. If you want to get into the powerpc architecture your only option is second hand systems or be lucky enough to have this hardware in your workplace?

T2 Lite and Blackbird are certainly less expensive options. Still, the major cost is the logic board because there are no economies of scale (but the CPU cost, compared core to core and cache to cache, is actually competitive with x86: IBM makes those in rather more volume than Raptor and others manufacture OpenPOWER boards).

POWER in uppercase had always been for clustered mainframes and full depth racked servers

> Looking at the Basic Talos II motherboard and CPU bundle, that's already 3480 USD.

That's quite competitive with AmigaOne once you account for the far better performance you get with the Talos. Run a free AROS operating system on the board and you've got a great, future-proof machine for all sorts of uses.

It's achievable (for most people I suspect frequent HN, anyway) if you put away a couple months of disposable income. If you don't care about playing games on it it's about the same cost as a high end gaming computer, and what games were you gonna play anyway?

If the virtualization story is good, it might still work out. The only concern is the graphics accelerations part. That hardware could be fully handled by the VM though too.

Drivers work, there's just no games. So if you want to train ML models go for it.

Maybe https://cloud.ibm.com/catalog/services/power-systems-virtual... ? The pricing calculator says 0.5 cores and 32GB RAM is ~$310/month, billed in increments of $0.43/hour. (Pricing seems to be, approximately $10 / GB / month for RAM)

Reply on news.ycombinator.com