A 20nm 32-Core 64MB L3 Cache SPARC M7 Processor | Hacker News Reader

A 20nm 32-Core 64MB L3 Cache SPARC M7 Processor

ieeexplore.ieee.org

70 pointsluu33 comments

I run a community-oriented Sun/SPARC site, since 1997 (and was invited to be part of the private prerelase testing/eval of OpenSolaris) and it's really sad what Oracle has done to the enthusiast community.

The "old" Sun gave me a loaded T1000, gratis, to run the site on (I'd actually asked for one for a review, and they said "Keep it for the site."). The "new" Sun/Oracle gave me the finger.

I had to move everything off the T1K and onto a Debian box hosted at home because I couldn't afford a support contract for it (required to get any patches/updates, even for the OS install that came on it).

I've often likened Oracle's treatment of the hobbyist/enthusiast community as "they shot it in the face and said unless you're going to spend $XXXX with us, we don't care about you. Period."

The "old" Sun recognized that if sysadmin types played with older gear at home and enjoyed it, they were more likely to recommend it for purchases at work, and they encouraged it.

The "old" Sun went out of business following the practices it did. I dislike Oracle as much as anyone else, but Sun isnt a great role model in terms of savvy business practices.

But in this case I think Sun was right. The marginal cost of making a processor is pretty low, I imagine, and it's not like enthusiast sites can be pumped for much money anyway.

Larry's yachts don't manifest themselves.

I was able to, at different times, email both Scott McNealy and Jonathan Schwartz directly, and got personal responses from both of them. In the grand scale of things, I'm a relative nobody, just a dude who runs a web site that talked about their products.

I can't imagine being able to have a 2-3 message conversation with Ellison at all, as I don't spend enough with Oracle to be worth a microsecond of his attention.

If you're interested in the M7 the Hot Chips conference talk was pretty interesting. It starts at the 33 minute mark in this video.

https://www.youtube.com/watch?v=G4f62FJGVcE

If you're interested in CPU architecture the Hot Chips talks are generally worthwhile. They're available for free a few months after the conference.

That can only be yours for just 150,000 downpayment.

SPARC has been pretty long track history and I admit I really used to love Sun Sparc Servers; now I just look at them and think of the cost and cry.

Without access to the full PDF, this blurb from Oracle gives some more information: https://blogs.oracle.com/rajadurai/entry/sparc_m7_chip_32_co...

It's been almost 8 months since this blog post. How's it been since? Is the CPU released already? (whatever that means for that kind of decidedly non-consumer-grade hardware)

EDIT: The CPU is 'scheduled to be available sometime in calendar year 2015' according to http://www.oracle.com/us/corporate/features/sparc-m7/index.h... So, barring rumors and insider information, I gest this is it.

What I want: 64 or more lanes of PCI-E 3.0, 4 or more channels of DDR3-1866 or DDR4-2133 or faster... combined into a chip with few but very fast cores.

My closest option? E5v3-1620 or 1630. It'd be nice if someone started producing non-x86 options for that.

The interesting thing about ARMv8 is that if there's a niche of users with your problem, someone can go ahead and make a SoC. They don't have to ask permission from anyone, just pay the license fees. If you wanted an x86-based SoC, there's no way Intel would make it.

Discussions about ARM vs x86 tend to miss this point. They tend to get stuck on performance or software or process size, missing that the revolutionary thing is the licensing.

[To head off some replies: 64 bit ARMv8 has fast single-thread performance, unlike 32 bit ARMv7]

If you wanted an x86-based SoC, there's no way Intel would make it.

Intel has made several x86 SoCs, and there's also a few others:

http://en.wikipedia.org/wiki/List_of_x86_manufacturers

This one in particular is interesting since it contains most of a standard PC:

http://en.wikipedia.org/wiki/Vortex86

However the current market for x86 SoCs is in industrial/embedded where performance is not a main concern, so these are not particularly fast.

I think they meant 'if you have a custom requirement Intel won't make it, you can only choose between predefined options' as opposed to ARM or others where people will build custom configurations of cores and modules for you.

There are a handful of x86 licensees out there.

It's not remotely the same as AMDs eco system of course, as they are competing in someone else's market. But perhaps it's comparable to making software for Apple or Microsoft's platforms.

(There is an incomplete list here: https://en.wikipedia.org/wiki/List_of_x86_manufacturers )

PowerPC is in the same situation, with 'OpenPOWER' (I think the Open must stand for "we're Open for business"?).

Of course, open instruction sets like RISC-V http://riscv.org/ require no license fee. At lowRISC we're aiming to build a complete open source SoC using RISC-V http://www.lowrisc.org/. I'll note that getting a license for an existing architecture and core implementation isn't as straight-forward as you say. You do need to negotiate and enter an agreement with ARM/MIPS/... (and ARM/Imagination/... need to decide it's worth their while going through that process).

My girlfriend (your typical first time builder: grandmother, artist, refusing my help) selected the E5v3-1650 which really screams at 8 GHz and can be overclocked. The build came in under $2K. She did it all herself and it booted the first time.

> 8 GHz

Is that the overclocked speed? I could only find a maximum of 3.5 GHz in the specs...

He probably meant 3.8GHz, the standard turbo speed of that cpu (speed under high load if cpu is not overheating).

Those are nice chips, yeah. Have you tried out any of the new DDR4 CPUs? I'm very curious as to how the combination of higher memory bandwidth and higher memory latency affects real-world applications. My first guess is that it turns out the same as DDR3-1866. Anyone benched it?

I have DDR3-2133 in my workstation now. The new DDR4-2133 won't be any faster.

But will it be slower for real-world applications? I've seen numbers indicating a ~ 10% worse memory latency. Don't know how that will be felt.

DDR4 in of itself will not perform slower in memory latency in any usefully measurable way. Memory latency in the real world is almost entirely governed by the pre-fetch engine in the CPU (which exists to hide the fact how slow system RAM is to begin with).

I would like to see a test between the Haswell E5v3s that support both DDR3 and DDR4 and see what the real world difference is on that specific implementation (which could not be generalized to other implementations of memory controllers by other vendors).

Man, even at 20nm, the die size must be a bear. Especially the L3$... I know TSMC has a ~600mm^2 max (for remotely reasonable yields), even on the highly mature 28nm process. Thus, I doubt this would be a viable chip for deployment today, rather a proof of concept. Need more details to be sure though.

Cache is repairable and bad cores can be disabled, so yield might not be too bad. I doubt Oracle is concerned about cost, though.

I'll probably get to play with some of these at some point and hope they can make the systems I use a bit faster. The T5s don't cut it against equivalent code running on IBM chips. It's unfortunate, because Solaris is a decent OS.

Oh...AIX isn't the worst thing in the world if you don't try and pretend it's going to be Unix like you want Unix to be and do things the AIX way. In return, Power8 does crush T5 for most things.

If only we'd make a Power 7 or 8 workstation...

Sadly even if the only "unix" thing you want to do is compile a bunch of standard open source packages, you're in for a world of hurt. Need to use xlc on top of that? Better have your seppuku knife ready. At least IBM took our "suggestion" and now maintains modern gdb so at least you aren't stuck with quirky dbx tracking down issues.

There's still some bizarre stuff though. You can set separate text/data page sizes via env variables that differ from the system wide settings. Take a guess which override, if either, will be reported through POSIX calls? Neither! Fun when someone changes how they invoke your app. Just an example...

If you want Linux, it is available for Power.

Take a look at http://www.perzl.org/aix/. Not only for the packages, but the porting patterns in the SRPMs.

Many open source packages don't compile out of the box? Like on many of them don't on *BSD? Dbx, the debugger many of us have been using successfully for the last 2 decades, is "quirky", but gdb isn't? You can tweak weird things in opaque ways? Compilers have differences?

AIX is not Linux, so you don't like it. Exactly my point. If these are your use cases, use Linux; don't bitch about AIX because it isn't. You can even use Power blissfully without ever having to use AIX. But then we'd have to hear about Power sucking when a linux developer baked the assumption that Linux == Intel ISA into some "standard" open source packages.

I use AIX and Solaris pretty much all the time. The pain incurred from building open-source code on Solaris is nothing compared to getting things to work properly on AIX. The dbx quirks on AIX I run into are issues with it not even supporting the code output by IBM's own compiler because the team building AIX is completely separate from the compiler team in Toronto, which is again completely separate from the team that owns dbx. You can imagine what kind of issues you run into when the compiler team implements features that the debugger team has not considered.

You keep missing the point. If your use case is running open source code bases, then don't use AIX. No shit it sucks, since the vast majority of open source developers probably haven't even seen an AIX login screen before, much less made any effort to make their projects seamlessly build under it. Hell, most of them don't even try to make it work under BSD/Solaris. You're using the classic "Ferrari sucks because it isn't build to haul a boat behind it" false argument. Use AIX for DB2, SAP, Tivoli, etc. Big stuff that is actually targeted for AIX.

And, BTW...last time I checked, the people doing the Linux kernel were different than the people doing GCC who are different than the people doing GDB. I don't disagree that the AIX ecosystem is generally more dysfunctional, but the mere presence of separation is yet another straw man. Linus has had some less than flattering things to say about what GCC has been up to in the last year or so.

Does it come with a free spurious patent infringement lawsuit over replicating the API of software its vendor previously released under the GPL?

Reply on news.ycombinator.com