Power9 is one of the highest instructions-per-clock architectures in the world.

Power9 just has relatively weak vector units: only 128-bit. (Even AMD Threadripper has support for 256-bit. Power9 128-bit is executed per super-slice, so its really 2x64-bit, to compare against AMD's 2x128-bit support) Otherwise, its throughput is downright insane. 6-issues per clock (Skylake only has 6 if its in the uOp cache: otherwise the decoder only pushes 4 uOps), 10MB of L3 per pair of cores. That's 20MB of L3 on the 4-core, and 110MB of L3 cache on the 22-core.

-----

As far as price/performance goes, the 18-core / 72-thread chip seems best. But the 4-core / 16-thread or 8-core / 32-thread chips aren't bad, as far as I can tell.

The main issue is that most compute-heavy code these days use those vector units (2x128 on AMD, 2x256 on Intel). But if you got a big GPU for SIMD / vectorized ops, then that disadvantage kinda goes away. GPUs are way fatter than Intel or AMD.

The other issue is that AMD is offering cores at stupidly low prices. AMD Threadripper and EPYC are the machines to look at for raw CPU power these days. But I'd argue that the 18-core Power9 ($1,050) is at a good price/performance level. If you can find any application that benefits from the stupid-high 90MB of L3 cache on the 18-core, then you're more or less golden.

I'd bet that databases love that L3 cache.

thanks for the detailed info.

phoronix.com just published some Power9 vs. Xeon vs Epyc benchmarks[1]. It is actually not on par with Xeon/Epyc. also note that the motherboard used by those Power9 has a price tag of $2499, when a dual socket LGA3647 motherboard from a first tier vendor can be purchased at $400 delivered.

[1] https://www.phoronix.com/scan.php?page=article&item=power9-x...

A lot of the "weakness" there is due to the weak 4x64-bit vector units equipped on the Power9. Again: 4x64-bit vector units (Power9) simply has less throughput than 4x128-bit (EPYC) units, which has less throughput than 3x512-bit (Xeon Gold) vector units.

On the few tests which don't have anything to do with vector units, Power9 does incredibly well. See 7-Zip Compression, Stockfish, Fhourstones, LLVM compiles. Power9 is very, very far ahead of the game here.

When vector units (ie: OpenMP or intrinsics) are involved: such as Parboil, Gimp, x264, Rodinia, etc. etc. the Power9 does very poorly.

-------------

Something else I've noticed: Power9 has a lot of latency behind its units: it is always 2-cycles or 3-cycles of latency (while EPYC or Skylake have many operations with 1-cycle of latency, like add or subtract). Power9 relies upon instruction-level parallelism to get things done.

So there are a few cases like the Rust Mandelbrot benchmark, which surely hits that latency problem, while the Rustlang Prime benchmark shows off the ILP goodness the system offers.

The same "bundle" with the 8-core POWER9 processor [0] is also available for pre-order.

[0]: https://secure.raptorcs.com/content/BK1B02/intro.html

Considering you don't even mention the openness of the raptor products vs. the xeon alternative, you're not the target market.

This is really remarkable "you can audit and modify any portion of the open source firmware on the Blackbird™ mainboard, all the way down to the CPU microcode."

I'd be heading to RISC-V for openness.

You can buy riscv in a year or 2 (for desktop), or wait 5 years for it to become performance competitive. Or you can buy power9 today (or end of 2019 for a cheaper model) and get a system that already exists, has silicon in production, and already appears to be performance competitive with Intel.

Mind, I'm probably waiting for whichever gets <$500 first because I'm not necessarily in a place where I can justify the money even for an open platform, but let's not pretend that there are massive benefits to both.

> Mind, I'm probably waiting for whichever gets <$500 first

This is pretty much my threshold as well, but when I actually think about these numbers and put them in perspective with regards to how much I used to spend on 486/pentium desktop computers when they were modern, and adjust those numbers for inflation, it's downright absurd that I'm not willing to spend what raptorcs is asking for a modern, open platform, in 2018 dollars.

Bear in mind the P9 chip does SMT4, so four cores gives you 16 hardware threads, and 8 cores 32 threads etc.

this will give much more performance and compatibility than risc, at least if it ships soon I guess.

The Power9, like all PowerPC chips, is a RISC based chip. Do you mean than RISC-V?

Or a 4-core ITX ARM board with almost[1] 100% open TianoCore UEFI firmware with ACPI and with x86 emulation for VBIOS [2] that allows video output from the EFI framebuffer… with a DIMM included… for a THIRD of the price of the Blackbird:

https://www.solid-run.com/marvell-armada-family/macchiatobin...

[1] one 12kb blob: https://github.com/MarvellEmbeddedProcessors/binaries-marvel... [2] https://github.com/MarvellEmbeddedProcessors/edk2-open-platf...

Not really blob free then, is it? ;-) Do you get full documentation of the processor internals and board schematics with that kit? How about an owner-controlled secure boot, or is that remaining blob vendor-locked / vendor signature controlled?

The most pertinent point though is that the POWER9 will run circles around the ARM board in terms of performance, guaranteed. We're using the POWER9 as a full desktop replacement, ARM never really worked out in that role due to general performance issues / lagginess vs. the x86 boxes that were being replaced. A ~2GHz embedded ARM core isn't going to match up to a ~4GHz POWER core, at least not favorably -- you do get what you pay for!

EDIT: On further comparison the Blackbird has double the SATA ports, double the SATA speed, two PCIe Gen 4 slots (a x16 and a x8) instead of a single Gen 3 x4, and quadruple the USB 3.0 ports. Really, the Blackbird is in a different class from that ARM system; it's competing against Intel and AMD boards, not against the low end ARM offerings.

Really interested to know your use case for power as a desktop cpu ..

For one, all x86 malware fails instantly. That's why some people I knew who wouldn't be targeted more specifically were still on PowerPC Macs. Their software was fine for their use case. I have one for experiments from 2003 that ran YouTube fine with a bit of lag a year or two ago. Although I used default, there's even still a browser for it:

https://www.floodgap.com/software/tenfourfox/

People who use a lot of terminal apps that don't need a full-on browser or non-portable apps can use it to its fullest. Most importantly besides owner-controlled, the fact that you're getting to use, show off, and develop on a unique system nobody else in your area probably has. Let's not forget the novelty and awing people aspect of tech that might sell some of these.

TenFourFox dev here. I certainly agree that security through obscurity (especially on less common architectures) is underestimated in terms of the protection it offers, though it should never be the only security means, of course. However, I think the biggest practical risk to unusual arches is through cross-platform-capable code. If you'll pardon a minor shameless plug, for the security advice I give Power Mac owners using TenFourFox, see

https://tenfourfox.blogspot.com/2017/11/the-security-blanket...

As you'll see from the article, aside from the OS X-specific exploits on Power Macs, components capable of running platform-independent code such as Java, Flash and Office macros are probably where the biggest risk is. And, of course, web browsers. Unfortunately these are some of the most common types of applications for people to run and very few are maintained on Power Macs anymore.

The good news with Talos and other P9 systems is that they're now running supported and maintained software and most of the applications people want to use "just work," so that problem goes away.

Thanks for chiming in! Great work on TenFourFox, too! I was stunned one person or a small team could even do that given how large Firefox is. I know you probably focus on a subset of it but still.

"components capable of running platform-independent code such as Java, Flash and Office macros are probably where the biggest risk is."

I'll add the risk of those components mostly has to do with their complexity, use of unsafe language, and security not being a concern in design. It's true there's gonna be exploits, esp on legacy systems. The crowd I was talking about was mainly concerned with malware forcing reinstalls, etc. That attackers mostly target high-ROI platforms meant they didnt have that problem any more. Although I suggested Ubuntu, they're Mac people with Mac apps they want to keep.

"The good news with Talos and other P9 systems is that they're now running supported and maintained software and most of the applications people want to use "just work," so that problem goes away."

Exactly. On top of it, many techniques for mitigating vulnerabilities have a performance cost. Esp overflow checking and microkernels. The extra speed of POWER9's might turn that from unbearable to acceptable. For me, Im fine with being stuck at Core Duo 2 performance for most tasks cuz my now-deadish, 9-yr-old laptop was working fine. If I get same performance but more security/control, that's a net gain. If it's faster, too, then that's even better. Similar argument might apply to those of you that port risky PPC software to it.

Thanks! Most of the work these days is maintaining the 32-bit PPC JIT and keeping up with security patches from the ESRs. Unfortunately I can only shove so many new features into old wineskins. :/

But what I learned from Classilla I used to port TenFourFox, and what I learned from TenFourFox I'm using to write a POWER9 JIT for Firefox and keep the build working. So it's all incremental.

I guess situation where you are privacy-conscious enough to inspect all the source code up to and including the CPU microcode? There was post in this thread about FreeBSD being available. If there is a Linux kernel for it, then a distribution might be not too far off e.g. Gentoo (I’m not aware of any POWER9 Linux distribution as off yet).

There's a PPC handbook for Gentoo:

https://wiki.gentoo.org/wiki/Handbook:PPC

Looking around, power9 seems to work fine. For software compatibility it's more a case by case scenario.

Fedora 28 and up boots directly on the Talos II and should boot on this. I use F28 myself, but the Raptor folks use Debian which I'm told also works well.

Looks like Debian supports POWER9 https://wiki.debian.org/PPC64

For me, it's a combination of familiarity with Power ISA since the RS/6000 days, a chance to have as trustworthy and auditable machine as possible (no ME/PSP bullcrap), and a desire to support other architectures than the usual ARM or x86 duopoly. Fortunately, you don't have to make very many compromises. Most stuff just works and I'm delighted with the performance of my Talos. Most of all, I'm very confident it's not doing anything or exposing something that would make me vulnerable or insecure because I can audit it myself and the full documentation is included.

If people want choice in computing, we need to step up and support these alternatives or we'll reap exactly what we've sown.

Of course 4GHz and PCIe x16 are very desirable, but $1000 for just the board+CPU bundle is waaaay too much if you just want to play with a non-x86 desktop.

With AMD, you can fit a whole build (with an 8-core CPU, RAM, SSD and a good GPU) into $1000. Your mainboard is just ridiculously expensive. Yes, I understand that the volume is low and there was a lot of new R&D and whatnot… but it's still just a big PCB. What costs so much? Is PCIe gen 4 (that I don't need) contributing to the cost a lot?

Wow, SFP ports. I'm seriously tempted to pick one of these up... the only downside is that it appears information on the CPU (and specifically the TDM interface protocol) is locked behind a Marvell NDA. That'd limit its potential for expansion, though I guess USB is still an option.