A brief overview of IBM’s new 7 nm Telum mainframe CPU | Hacker News Reader

A brief overview of IBM’s new 7 nm Telum mainframe CPU

arstechnica.com

80 pointsTomte28 comments

Open articleView on HN

What I thought was interesting was the caching layout:

L1: Typical L1 cache

L2: On-die cache like L1 but more/slower

L3: No such thing. It's just L2 on another core

L4: No such things. It's just L2 on another chip.

Fantastic stuff.

Yeah it remains to be seen how well this will work in practice.

In general I think it's an awesome idea... but years of "that sounds cool and logical" when it comes to cache coherency at those speeds leads my actual reaction to be "yeah... that is going to break in a horrible way or expose nasty side channel attacks".

Yeah I was wondering about side channel attacks, but L3 cache is already shared so it isn't really any different.

> or expose nasty side channel attacks".

Good point, but also keep in mind that unlike desktop/laptop processors these probably aren't going to be running untrusted JS from random web pages or binary "drivers" to watch premier league free, use Photoshop without paying etc.

Side note: Assume you don't have extremely powerful enemies that are targeting you just keeping untrusted code and access off certain machines does wonders on a general basis.

> aren't going to be running untrusted JS from random web pages

But they will be partitioned and run Linux under zVM and KVM. IBM does that with z15’s in their public cloud services.

I’m pretty sure they have lots and lots of mitigation and security contexts flowing along cache data to make sure one context doesn’t expose information about to others.

I wish IBM had a massively lower barrier to entry.

They have a lot of amazing products and tech, but the only way you can use it is to buy a license. When I want to make a simple application, I'm not going to even consider paying for unknown software. But if I can get a version that gives me very good insight into how it works, and a license that would allow me to use the product for a good amount before paying- I'd be much more inclined to adopt it.

POWER CPUs are great, but they're quite hard to get for a developer like myself. I'm sure you could make some cool software on POWER9 or one of these, but they're impossible to get your hands on to play around with.

POWER9 is expensive, POWER10 is unobtainium as far as I can tell and will be for a while.

POWER10 is unlikely to be available in until Q3 or so 2022 I would imagine. With things as they are unobtanium is going to be status quo for some time. :(

You can get Power 9 systems on IBM's cloud. There's some pricing on this page. I don't know enough about clouds to know how it compares to AWS: https://cloud.ibm.com/catalog/services/power-systems-virtual...

For what it's worth, in ~2002 we were developing a product that IBM wanted ported to run as a Linux VM on their hardware, and they just plain out loaned us a smallish mainframe for a year or two. It only fit in the elevator after removing it from the wood crate.

Taking a big step back, it's important to remember that unlike PC processors, mainframe processors do not do I/O. These big frames have dedicated I/O processors called channels. So the mainframe processors are unencumbered and can run at full speed. They issue CCW's (channel command words) which are fielded by a controller and are routed to the storage device.

Do you consider talking with the co-processor to not be IO? If not what what counts as IO?

The co-processor runs channel programs autonomously. So "talking" to it is not necessarily bounded by your I/O volume. It's more like passing a quick message to a separate thread to run some operations, and then the other thread can do quite a lot of work.

One thing to be aware of is that unlike software in the commodity world, the mainframe charges on a "MIPS" basis. Both for the base software, as well as application software from either IBM or 3rd parties.

So there's no magical post-upgrade "increased capacity" for the same overall cost like there is in the Windows/Linux worlds. You may see some uptick in things they don't specifically charge for, like better base I/O.

They are usually sold with unlicensed capacity you can immediately enable (instead of waiting for the new servers to arrive). IIRC that capacity can be enabled during boot to make the startup faster.

What exactly does a mainframe mean today? I’ve logged onto multi user computer systems via a thin client before, but I guess the hardware I was connecting to was probably commodity so not a mainframe? Do these things run Windows/Linux or is this some super proprietary OS2 descendent or what?

The article answers this in the first few paragraphs. Mainframes are extremely reliable and have high I/O bandwidth (and relatively poor CPU performance). You can run Linux on them - Red Hat ship and support RHEL for s390x - but most of the time they run z/OS along with z/VM. https://en.wikipedia.org/wiki/Z/OS

But 5ghz is an incredibly high clock speed. Is the cpu hampered in other ways?

I use z14 at work and it's not especially fast for single thread stuff. For the applications that typically run on these, performance of the entirely separate channel controllers (I/O) is what matters.

We've been having slices of L3$ tied to individual cores for a while now, so this is a bit more precedented than it sounds.

This is quite a bit different. Yes, private cache is not new, whether that's L1, L2, or L3. But for multicore workloads, you HAVE to have some shared level(s) of cache for decent performance. Being able to have lots of fast/local per-core cache, and so have it do double duty as shared LLC is really new AFAIK.

Previously we had L2 caches that was only used by a particular core and slices of L3 which was tied to a particular core but could contain values used by other cores as well. Now we have L2 caches that are tied to a particular core but can contain values used by other cores as well. Especially with the partition into virtual L3 and virtual L4 I think this is an important evolution beyond your old MOESI or whatever but I'd still say that this is more evolutionary than revolutionary.

This seems like a good time to ask if there's any point at all to mainframes besides legacy applications.

If an org has never used a mainframe before, what would be a good reason for them to start?

If you're processing a very large number of transactions of some kind and require easy to validate complete consistency, mainframes are still generally king. Thus their frequent use in finance and finance-like fields like transactional government services. The throughput for transaction processing can be extremely high due to things like offloaded IO, and mainframe operating systems and development tooling are very amenable to an "if in doubt, back out" approach to consistency (e.g. any failure in processing a batch results in the entire batch being reversed and left for an engineer to troubleshoot). IBM also supports a lot of extreme reliability/consistency use cases like consistency checking through mirrored processing on multiple machines.

IBM also likes to argue that, for things like normal Linux VM workloads that require a very high degree of reliability, mainframes can have a lower TCO compared to approaches that rely on complex software for redundancy/failover (mainframes achieve high reliability through hardware-level redundancy, so hardware failures and repairs generally occur without any of the running software ever being interrupted). Judging by the low adoption for this kind of use I'm not sure if it doesn't pan out or it's just too hard of a sell given what's popular in the industry today. There are also more "midcomputer" options for this kind of reliability like HP Tandem and you don't see those as much these days either.

Keep in mind you will not find faster single thread performance anywhere else in the computing world. A free VM on the LinuxONE Community cloud puts many developer laptops to shame.

Meanwhile others say "not especially fast for single thread stuff" https://news.ycombinator.com/item?id=28422687

I'm pretty sure Apple Firestorm and AMD Zen 3 are trading blows for the best single-thread performance right now depending on workload and optimization. Apple no doubt having the best efficiency due to being on TSMC N5.

It depends on the workload, of course. If you are doing things that hit the memory, the massive 960MB L4 of the z15 will help a lot. If all you do fits on the caches of an x86, then the x86 may be a better option. Offloading all the IO to dedicated processors will also be helpful - the z15 feels like it has instantaneous IO. The z15 will probably suck at floating-point SIMD ops (something our laptops are quite good at)

x86 is a sports car, it'll go fast for a while, and tends to overheat. Mainframes are freight trains that'll just keep going, pretty much no matter what.

They're typically used for the kind of old school business critical stuff where downtime (or errors in output) can cost millions per minute.

Imagine hotswapping or upgrading CPUs and RAM without downtime, without having to do any distributed programming.

Imagine CPU offload for SQL queries.

Imagine fast VM-to-VM TCP connections with hardware offload (direct memory handover) and essentially no CPU cost.

Reply on news.ycombinator.com