IBM Power10 Coming to Market: E1080 for ‘Frictionless Hybrid Cloud Experiences’ | Hacker News Reader

IBM Power10 Coming to Market: E1080 for ‘Frictionless Hybrid Cloud Experiences’

anandtech.com

53 pointsrbanffy32 comments

Don't forget: https://www.talospace.com/2021/02/a-better-theory-on-why-the...

What a privilege it is to contribute to P10. OMI is mind-bogglingly fast and will be an even greater advantage during the enterprise DDR5 era.

Would love to see one of these in a Raptor Talos machine.

I was thinking the same, I dunno if they sold well enough to justify a new version, but I hope so. I’ve been tempted several times, but it seems like quite a lot of money for a Linux box that probably has quite a lot of subtle incompatibility quirks.

The problem currently isn't how many sold, but issues regarding IP that RCS (reasonably) does not want to compromise on. Issues that block, for now, a new POWER10 system unless you still want blobs, just not shipped on-board - and want to pay ridiculous prices for unique memory modules.

Ahh, that’s a shame. Blobs would kill the whole thing dead for a large proportion of their customer base. Let’s hope they find a way to resolve it.

There seems to only be two blobs in POWER10. The memory controller and something about I/O.

The question is what to do about it. Maybe an open source firmware rewrite can happen. Or maybe IBM needs to make a cheaper desktop version first with a more open DDR4/5 controller?

That advanced RAM module isn't needed for desktop level workloads anyway.

Either way, 2022 is out. Maybe 2023 will have thing line up better?

In an ideal world, IBM would help cultivate the ecosystem Raptor is creating.

I'd say that releasing all the other firmware of POWER10 as open source is a big deal still.

They didn't reach Raptor's high-standard of open firmware, but POWER10 is probably one of the most open CPUs out there in the modern marketplace.

The OMI issue is probably solvable with an open controller. Not a trivial undertaking but OMI, at least, is documented.

The on-chip PPE I/O controller is a bigger problem. I suspect, but don't know, that it's the PCIe interface. If so, I can't think of an easy way of getting around it.

> Or maybe IBM needs to make a cheaper desktop version first

I doubt that will happen. For that to happen, there would need to be a market for low-end POWER servers. Low-end means low-margin and competition from x86 boxes. There may be space for a low-spec IBM machine (a POWER-based descendant of the AS/400), and I imagine the IBM i subsidizes a lot of the development of the AIX-based POWER boxes (because they are the same hardware).

I really think IBM should spend some money on entry-level boxes for their exclusive platforms (P, I and Z). I don't see I or Z going anywhere - they are good enough and too expensive/risky to migrate away from, but P and AIX are too close to generic x86 hardware running Linux for them to feel too comfortable. Right now, it's hard to justify even suggesting a green field project using anything that's not a commodity platform. Being less expensive doesn't help when the minimum sticker price is more than 100K. Chances are development would start on generic boxes or cloud and stay there. IBM is the only company that can make POWER competitive with x86 at an entry-level and a single-core SMT8 part would probably be a justifiable expense for many R&D departments.

> with a more open DDR4/5 controller?

My impression is that the Power Axon interface would be able to control DDR4/5 memory.

There is only one blob that can't be recreated from source that is needed for a working system, and that's OMI-DDR4 interface. There's talk of reversing it, but nobody has the time or resources. So long as said chip is the only source for making a motherboard that doesn't require OMI memory modules (which have the same chip anyway), there's a problem.

EDIT: Hah, I missed the PPE discussion recently. There's now some ppl looking towards reversing that too, but again, not RCS itself AFAIK nobody has seen non-OMI POWER10 cpu on the road map.

What is the intended usage for that? As in used in Consumer PC ?

Would microWatts [1] fit that purpose? Not only do you have an Open ISA as in OpenPower, you also have an Open implementation of that ISA. AFAIK RISC-V doesn't have anything similar, only open source design in embedded usage.

Otherwise a low cost POWER10 ( or now Power10 without the capital... and I hate it ) wont make much sense. You are talking about ~30mm2+ per core design. It is huge.

[1] https://github.com/antonblanchard/microwatt

Same usage as the current Talos II: an expensive, but high-performance, open source computer system.

18-core POWER9 was competitive with AMD Zen1's Threadripper. Zen3 Threadripper is getting quite a bit faster now, so an upgrade to POWER10 would once again be cost-competitive at the $5000 price-point.

Here is a relevant article

https://www.phoronix.com/scan.php?page=news_item&px=IBM-POWE...

Such a beast. I would love to play with one of these. Other than some really serious CFD or EM field simulations I don't know if I could keep it busy :-)

So how do these stack up against running that same workload (simulations) on a powerful GPU? Often a sim requires running the same operation over thousands of different cells: it seems really suited to the “warp” style of parallelism, where all cores in a set run the same instruction in lockstep. Or SIMD, but that’s usually a bit more effort to port. LLVM has a spirv backend these days, so does POWER actually make sense for these workloads outside of places where it’s difficult to port your codebase?

Generally it depends on the working set. Massively parallel architectures with shared visibility into a coherent memory space can often more efficiently compute volumetric problems. Doing the same problems on GPUs involve a lot of data shuffling and you end up on the wrong side of Amdahl's law.

I still don't get what hybrid cloud is and how much of it is marketing and how much is actually useful tools, services and/or infrastructure.

Basically having mixed deployments across cloud environments, with services that ease application management and deployment across them.

Sure one can do that today, but consulting shops and cloud vendors need new products for keeping the board happy with exponential growth.

It's mostly around federation and abstraction of the lock in's of a particular cloud provider.

For example, I might want to run my own hardware for a certain reason and then also run workloads on AWS, but I don't want to have to manage different sets of API's , auth methods etc.

The other commenters mention multi-cloud use cases but another case is prem-cloud hybrid. You may want some of your business on prem for whatever reason but be able to move workloads to various clouds and back, for price, load, or whatever reason.

So if that’s the set-up (on-prem + a cloud provider), what’s being sold as a hybrid cloud solution - some services that allow you to manage them better or something else? What are some examples?

Chips that are architecturally similar / behave like the cloud platforms, and they are claiming it is designed to allow a higher container density (i.e. run more containers at once, where some of those containers might have very little activity).

AWS is probably further along than the others throwing offerings at the wall. They let you do things like have a single api that can manage your stuff on your prem and in their cloud. It can look like containers, like VM's, like a VPN, all sorts of options. Sometimes it sticks to the wall.

https://aws.amazon.com/hybrid

AWS Outpost and GCP Anthos are examples of hybrid cloud services.

At first cloud providers created tools to make it easier to migrate/connect your on prem to your cloud environments. I’m guessing what usually happened was that some workloads would get migrated but others would remain on prem (for various reasons) so cloud providers started building products that made “your Data Center in our cloud”. Ie deploy cloud-like services within your Data Center so your developers have access to the same interface regardless of where the workload is deployed.

It’s about putting things in the cheapest place with the lowest risk.

Cloud was a great option in the beginning because oh the elasticity, less cost and effort to maintain (in comparison to on prem). But, as companies as AWS, Google, MS and IBM worked to lock their clients in, costs have risen and the cloud economic benefit faded, some companies tried multicloud, that worked for a while and now some companies are actually moving some of their infrastructure back to on prem, so there is the hybrid cloud case. What I've seen is companies covering their base demand with on prem and leave the cloud to deal with the elastic part of their demand. If costs change sides, they can also use more cloud and less on prem and vice versa.

I agree to some point with the other commenters. But the reality is that almost every enterprise of any size will be hybrid cloud for quite some time if not forever. It’s impossible to lift and shift everything at once which means you need the infrastructure to connect those environments regardless.

On top of that you’ve still go some applications that run on platforms that just aren’t compatible and will need to be either rebuilt or left to run on-premise indefinitely.

Additionally, getting back to their other comments. I think there is some logic around data being viewed as “heavy” given the egress costs on most clouds. Having one place which you can easily upload data to any new environment that pops up for cheaper seems like a decent idea. Then again, a lot of money related things can be solved by simple contract negotiation, so maybe it’s not worth it.

> you need the infrastructure to connect those environments regardless

What does that infrastructure look like? What are some examples?

One example would be an on-prem object storage device that's capable of replicating data to AWS S3.

One example would be to use “interconnects” which link your data center to the cloud through a high capacity (maybe low latency?) “dedicated” line. So basically you have private IP connectivity between your DC and a virtual private cloud (VPC) so your workloads “think” they’re on the same private IP network. I thought of interconnects as beefy VPNs (but this may not be accurate, just a helpful mental aid).

Note: I’m using mostly GCP terminology, https://cloud.google.com/hybrid-connectivity

Reply on news.ycombinator.com