This genuinely surprised me, but I’m less sure than everyone else that this spells the end of Cray. They’re still cranking out monster systems: they have 3 in the top 10 now (and almost certainly more than that since the DoD stopped listing their systems). As everyone has noted, they also signed contracts for new monsters at Argonne and Oak Ridge, which will likely debut at #2 and #1 respectively.
I think it’s very unlikely that the US government lets them go out of business after HPE closes the sale. They’re the best competitor to IBM, and the DoE and DoD are always careful to spread their procurements around to keep more than one company capable of supplying the big defense/weapons supers. This is just totally a guess from having worked in HPC for so long, but I’d be very surprised if this purchase didn’t include some sort of back channel wink and nod by the feds at a promise by HPE to keep building the big computers.
I’ll miss them though. While their systems weren’t always the best, when you got your problems escalated to their R&D group, you got to work with some cool people. I imagine those people will get sucked into HPE and/or get fed up and defect to Intel pretty quickly.
Look at HP's history. They killed the Alpha when the Alpha was the top processor of its time. Even after the decision to kill the Alpha, the sheer momentum of it led to many Top 500 supercomputers such as ASCI Q.
Why did they kill Alpha? Because of a deal with Intel. HP hasn't made money off of shipping Itanic systems. Compared with where they'd be if they never killed Alpha, they've probably lost billions of dollars.
But financially motivated manipulations are more common than drama in a Korean soap. So will HP do the same thing with Cray? We don't know. They could have some back end dealings with Intel again and we could see inferior Intel-based systems instead of AMD-based. It's easily within the realm of possibility.
Alpha is a painful loss for sure, and the end result for Itanic makes HP look bad. But I guess they were looking at the figures to stay at the top with respect to manufacturing costs going forward at the time, which is just brutal when competing with chips produced in much larger quantities such as x86. We're talking tens of billions of dollars here which Alpha had no chance of recouping, ever. IBM/Power has/had similar challenges.
Does Cray manufacture their own silicon? I did glance through https://en.wikipedia.org/wiki/Cray including reading the lead, and did the same Google search... while I'm sure I could dig deeper, maybe it's fun to let people here say what they know :)
Unless you’re TSMC, Samsung, Global Foundries, or Chartered, you’re fabless. Apple, Nvidia, Qualcomm, AMD, and many others are doing ok as fabless companies.
It used to be one of the main things they did, but they spun it out into a separate company about 10 years ago.
They were failing to keep up with Intel in fabrication process because of the massive capital and R&D expenditure required. They decided they would be better off having a choice of fab partners instead.
No, they don't design CPUs. They design ASICs that run inside their network. This is the secret sauce that keeps them competitive in HPC environments where latency/bandwidth matter a lot more.
They used to do the whole shebang, back in the day. But it's been a long, long time since that was the case.
This is almost certainly due to the DoE Frontier contract awarded to Cray and AMD. [0] HPE is "riding the wave" of this major contract.
Cray did something similar with Appro when Appro won a large government supercomputer cluster contract. [1] When the contract was over, former Appro staff at Cray attrited.
Once the Frontier project is finished, there's no telling what HPE will do with Cray.
Seymour, Bill and Dave must be rolling in their graves. [2]
The Seattle office of Cray, which houses executives, HR, and some Chapel/MPI folks, will probably be closed at some point.
The Mall of America and Chippewa Falls Cray offices will probably stay open for some time.
(Full Disclosure: I worked at HP (1995-2002), Cray (2012-2015), and currently work at AMD.)
They do realize that this same action finally pulled SGI under the surface of the water to drown, right?
It would be super awesome if there was a company that had a workable business model where it could successfully do "very large compute system" research and development. Given the way they handled "The Machine" though, suggests to me that this is not that company.
If one looks into Cray's viability as a business, you'll see a pretty much unending string of losses. Though the CEO is the best paid (or pretty darned close) in Seattle, at least on a cash basis. As HPE is close to the #1 HPC vendor anyway, all they're going to get is entree to some of the more esoteric government contracts - which they'll probably wreck by trying to cram their awful consulting services into.
Tell me more...I'm an HPE fan, and live near Ft Collins, get to tour the lab sometimes. I see progress being made but maybe its not in a very public way.
Biggest thing I could add is that the 'Machine' is not a reference architecture or target...its a pipeline of research that (by visual accounts) is continuing...
Well, here is how they announced “The Machine” back in 2014:
This changes everything,” said HP CEO Meg Whitman in introducing the Machine in her HP Discover 2014 keynote in Las Vegas. “We’ve been talking about many of the component technologies for some time,” she noted. “Now we’re bringing them together in a single project to make a revolutionary compute architecture available by the end of the decade.”
Since then, not a lot. HP Labs has done a lot of sexy things, but HPE as a business is big on hype, bad on delivery. Hopefully Cray won’t be the same
I disagree. They announced it as a revelutionary computing system that would change they way all computers would be built. They then chipped away at features until there was litle interesting left... I followed these stories with real interest and was left disappointed.
How does IB compare to Cray’s new interconnect? All supercomputing nodes these days appear to be water cooled, high density servers with a number of Nvidia GPUs connected to a low latency interconnect fabric, along with a Lustre/GPFS storage cluster. The interconnect fabric used to be the secret sauce, and IB has commoditized it. Why buy Cray?
You're correct in that, from the HW side, the "crown jewels" of Cray is the network ASIC's. Everything else is pretty much commodity.
AFAICS, the main difference between the latest Cray network (slinghshot) and IB are:
- Adaptive non-minimal routing, needed to get good performance on their more scalable network topology. IB has basically static ECMP, which works good for fat trees, which scales as O(log N). But for really large machines topologies like Dragonfly which Cray is using become interesting.
- QoS features to isolate jobs from each other.
- Some kind of compatibility with Ethernet. Not sure exactly how this works, but apparently when the NIC communicates with other compute nodes in the cluster it can use the high-performance mode, but if it communicates with some other system, such as Lustre servers, or management systems, or maybe even, say, pulling a container image from an external system, it looks and behaves like an ethernet NIC with the normal Linux TCP/IP stack running on top.
Cray just won a $600M bid for Oak Ridge National Labs' new exaflop HPC cluster using AMD Epyc/Radeon.
Most IB cards come from Intel and Mellanox (recently purchased by NVidia).
IB is a commodity, sure, but it's becoming increasingly tied into the Intel/NVidia ecosystem, and having the ability to go with another interconnect/processor architecture is a powerful strategic advantage.
Likely it's mostly just a customer base grab. I will say there is more to an interconnect than "is the hardware commodity or not" though, you have to be able to show you are able to come up with ways to use it more efficiently than competitor designs.
I don't think that "all" is true, and I'm not sure what "commoditized" means for IB more than any other HPC interconnect. They're all single-vendor as far as I know (as I don't think you can just buy Oracle IB): IB, OPA, Aries (and Slingshot), TOFU, BXI, and maybe others -- I don't know what the Chinese ones actually are.
Yea... except HP is where things go to die. They have a long proud history of acquiring a company and investing some money in it - then deciding it's too expensive, shutting down the department and fire-selling any hardware they have in stock.
Considering how Crysis’ biggest issue on modern hardware is that it doesn’t scale to multi-core, I doubt cramming it onto many more cores would help it, even if it ran
The HP Compaq merger as well as the Agilent spin-off in many ways marked the transition from an engineering company to a bunch of vacuum cleaner salespersons.
PS: Yes it's unfair to blame all this on Compaq, probably more a result of increasingly expensive semiconductor R&D.
My rule of thumb for buying laptops (since the mid 2000s) has been that compaq is the bottom-rung cheap brand that should always be avoided. Not sure their survival has been a good thing.
In 2006, every single one of my coworkers bought a brand-new MacBook, and within a month every single one of my coworkers had a MacBook in the shop.
I bought a Compaq laptop with a 64-bit CPU for under $1000. It ran flawlessly for over a decade, needing only a new battery. I eventually gave it to my parents who still have it.
Brand necrophilia. Compaq consistently built better gear than HP before being absorbed. HP used that brand for their junk as a way of getting back at Compaq.
Wrong. Compaq had much higher DOA and other defects in the mid 90s. They relied on customer institutional memory from the 80s when they really were the best.
Counterpoint: my Compaq Presario 1210 survived for about 20 years before it finally stopped POSTing. Even the original hard drive still worked (albeit with a range of bad sectors around which I had to partition).
The consumer gear was trash. The server lines were an entirely different story. Also worth noting that by the mid 2000s, Compaq was just a branding on the consumer side. The hardware was all the same old HP consumer junk.
Compaq is what ruined HP after they ruined themselves by going from a quality-focused builder to pulling parts out of the seconds and thirds bins to cut costs in the early-mid 90s. They absorbed DEC and ruined it then proceeded to infect HP (with Carly's help) with that culture. Not that HP was blameless either. I was done with HP when we received a $10k LH3 Netserver in '99 or '00 with both of its CPUs dangling inside the chassis from their fan cables. If my memory is right, they'd outsourced most of their building to Ingram Micro by that point.
That was years ago. The HP that exists today is certain a place things go to die. I'm not sure what HP does these days but like IBM it keeps tugging along on its brand name without dying.
No one remembers Palm here ? That WebOS was a fantastic OS and Apple iOS / Google copied it's gestures unceremoniously and think they innovated while still not able to perfect it like Palm.
WebOS tablet, HP touchpad 10 got Android custom ROMs and still thriving. Its a shame that had to die.
Now WebOS lives in LG Smart TV platform. A big shame, this made me remember Tizen (Samsung WatchOS) which was reanimated from MeeGo which was made from MaeMo.
There are plenty of System X HPC installs. They just no longer come from IBM. Even when IBM still owned the line, it was being outsourced. The University of California SRCS system from ~10 years ago was an iDataPlex sold to us by IBM. Most of the boards had Asus marked on them.
Also remember the System X line was sold way after the Think lines.
I honestly didn't realize that Cray was still around.
My dad has told me that when he was in grad school, the Cray was the coolest machine out there, and it felt kind of like magic to him. Do they still do hyper-efficient vector machines?
I work in computational science. Not sure how I feel about this. HP has not had a great track record as of late and I don't know what this means for continuing support of Cray systems already in use.
I agree. I think problems would be more likely if they started a directly competing product with HPC in the name, but an unofficial acronym for the joint name of 2 merged companies? I don't see it.
On a related note, would they be more complimentary or competitive to HPC Server? Both Cray and HPE seems like they're more focused on the hardware and possibly userspace applications than a proprietary OS. I'm pretty sure modern Cray's are Linux-based, and I don't know if HP-UX is suited to running a supercomputer, but even so it doesn't seem like it's core to HP's business - just one of the products in their portfolio that can fit where needed.
If HPC is a generic term for HPCs, then the company wouldn't be able to prevent competitors from marking their competing products as HPCs. So an HPC mark may have limited value. It would be like having a water company called WATER; you could do it, but you'd never be able to prevent other water companies from calling their products water.
Of course they could still call the company HPC, use the mark HPC for products or services other than HPCs, and use the mark CRAY for HPCs.
To be fair, one company was supposed to be called HP Inc, and the other one was supposed to be called Hewlett Packard Enterprise. These are the official names.
Oh, you forgot their software spinoff Micro Focus. 20 years of saying HP <softwarename>, everyone in meetings correct me with Micro Focus....old habits die hard.
I just hope it is not like other companies they bought, rebranded and just sort of update them a little but no real innovation (software mostly i mean).
And this isnt really the original Cray. Its someone who bought Cray and started using their name (although they kept the original location open.) So there is precedent. https://en.wikipedia.org/wiki/Tera_Computer_Company
I haven't been working in supercomputing for 25 years any more and my first reaction to the headline was "Oh, Cray does still exist?". But obviously my intuition wasn't too wrong, the surprise should have more been "Oh, the Cray name does still exist", but that's of course not an equally big surprise.
CrayLink was awesome. I worked for a web hoster in the late 90s that was big on SGI and we'd abuse temporary CrayLinks between two Origin 200s to upgrade storage or spin off sites to a new server with minimal service impact.
My father — who ended up at Unisys after HPE moved away from support (and was previously with Compaq via the Digital acquisition) — says it's unlikely, at least to start.
What is it like to program for a supercomputer? Where can I learn about it? Not having an academic affiliation, is there anywhere I can run small jobs affordably?
MPI is by far the dominant method for communicating among nodes. So learn that. MPI works fine on a multicore machine as well (launch one MPI rank per core), so you can run on your own machine, no need to use the cloud or anything like that.
For using GPU's, there's CUDA, or offloading with OpenMP or OpenACC.
Indeed, OpenMP offload and OpenACC seem so close to each other (Disclaimer, I haven't used either) that it's probably better if the world would converge on one of them.
Programming a supercomputer feel like having to write a program that uses thousands of computers, that’s about it.
Some supercomputer vendors have specialized software running, but others also support open source distributed computing frameworks like Apache Spark.
If you are wondering why we call them supercomputers, it’s because it’s not as easy to just buy thousands of computers. Even easy things like shipping, installation and power consumption becomes a problem on its own. So rather than buying thousands of machines it pays off to buy them through a supercomputer vendor that will optimize space, power consumption, install cluster management software, etc.
So if you want to get started, I would just get a small n-node cluster from any cloud provider and get better at writing distributed code.
I deploy to hundreds of machines at work, but that is just load balanced RPC and message queue processing workloads, interacting with other load balanced RPC and message queue workloads. The little information I can make sense of about “HPC” does not seem familiar through that lens.
Based on a very shallow reading of Infiniband, I think some of the interconnects of supercomputers are essentially DMA-based rather than network-based. So it's like writing something massively multithreaded (your barrier is one step above shared memory) versus multihosted.
HP was once considered one of the best places to work, and known for innovation and making some of the best products.
I understand that HP executives had a belief in management by walking around, and talking to the people on the ground.
I think it was around the time of Fiorina that this changed. There were mergers/acquisitions, changes in management style, and changes in industry.
Today, I still see signs of greatness in some HP products, though I can't forget seeing their brand on corner-cutting consumer PCs, and the playing of games with inkjet consumables is bad for brand goodwill.
I don't know how HPE fits into that history, but I suspect that their market demands they perform well.
I hope HPE keeps the Cray name for at least some purposes, and honors the name with great work befitting it. There's also the great name of HP to honor.
(Story: As a nerdy teen, I once got to go to a Cray division (Cray Research Superservers), to port some software. I'd grown up reading about Cray supercomputers, seeing them on the top supercomputers list, etc., Cray had both technical innovation and style, and there was a huge mystique around them. On-site, typing on a workstation frontend, telneted into a big cabinet across the room that would say Cray on it if the panels were on, was an experience I couldn't process, because I was half-terrified into accomplishing the mission (it turned out to be easy), but there was much gushing in awe to coworkers afterwards.)
The "HP" that was widely considered a great place to work hasn't existed in a long time. They spun off their original field (test & measurement) as Agilent, who then spun it off as Keysight. The two companies that kept the "HP" in their name have pretty much nothing else in common with the once-great company.
The HP(E) that's buying Cray is not the same organization that gave us LEDs, atomic clocks, scientific calculators, inkjet and laser printers, and the RISC CPUs. That HP is gone.
The one in my field of view right now is an HP LaserJet Pro 400 series.
My previous LaserJet, a 5N, I found set out for the trash on the curb late one night, in a sprinkling of rain. So I hauled it home, plugged it in, and proceeded to run it for over a decade, very infrequently putting new toner in it (never even had to replace the rollers).
Eventually, the LaserJet 5N's fan started to fail (all the mechanics and fuser and everything still working fine), which presumably was a replaceable muffin fan, but I also wanted a sleep mode (and preferably to not dim the lights at warmup), so I parted it out (too heavy to ship, but the parts were still marketable), and bought a contemporary LaserJet.
The new LaserJet is not as bulletproof-looking, but is sturdy and has worked like a champ for a few years, for letter and envelopes, and it still respects the toner. My only complaints are that I wish they wouldn't play setup convenience tricks with USB, and that I'm unwilling to give its huge firmware direct network access. (For CUPS drivers, instead of using `hplip`, I now use the simpler "HP LaserJet Series PCL 6 CUPS".)
Those old LaserJets are absolutely bulletproof. My parents are still using the one we bought with our very first 486 clone in the mid 90s. I had to mod it a little to swap the serial port for USB a decade or more ago, but it is still cranking along.
I usually used Ethernet-connected ones, but I'd think RS232 serial at a doable bit rate was viable for most purposes.
Both HP-PCL and PostScript (I wrote code to generate both) can be sufficiently compact. (And you had the trusty built-in fonts, plus sometimes additional fonts in cartridges/cards, so fonts didn't necessarily have to be sent with the print job.)
What could be a problem for connection via RS232 is large images, or an unfortunate setup that rasterizes the whole page off-printer at high dots-per-inch.
I have 2 different HP MF 1212nf (multi-function laserjet) that have run rock solid for 8+ years. The HP laptops designed for business use (e.g. HP Elite X2, my current HP Elitebook 360) have been great as well, slim, and user upgradable (drive, battery). Even a consumer laptop (Compaq) has been solid-ish apart from a prematurely failing cooling fan, which I admit was a pain to replace.
As much as people like to hate on Carly Fiorina, Mark Hurd had positive things to say about the changes she made. He gave her a lot of credit for laying the groundwork for things that eventually worked out.
All currently big companies were once great places to work in. The constant need to grow to satisfy wallstreet (or their own Ego), makes them do shitty things like play games with Iknjet consumables.
I would say that it is more to do with they way human nature works (in some people), than capitalism. Applies in every field of life, not just business.
Capitalism accentuates that underlying nature. It's a lot easier to say no to something clearly unethical/unscrupulous when there isn't a pile of money riding on it.
When there's a heavy incentive to pursue this nature for more fundamental needs like food, family security, etc. then it takes a heck of a lot more will power to not participate.
The definition of a supercomputer has changed over the years, but currently what defines a supercomputer is its interconnect. Most of them have InfiniBand or similar with <1us node-to-node latency. It's necessary for the type of workloads typically run on supercomputers, e.g. scientific applications. The fastest you can get on EC2 right now is ~15us, and that requires specialized hardware and software.
It's funny how some companies just seem to acquire old giants/pioneers as some sort of trophy case to the point where they have the DNA of most of the industry, but frequently little of the talent.
I think of companies like Chrysler who through AMC acquired pretty much all of the extinct mid-century American car companies and now has branched all over the globe. How much of that old history is even tracable.
I think it’s very unlikely that the US government lets them go out of business after HPE closes the sale. They’re the best competitor to IBM, and the DoE and DoD are always careful to spread their procurements around to keep more than one company capable of supplying the big defense/weapons supers. This is just totally a guess from having worked in HPC for so long, but I’d be very surprised if this purchase didn’t include some sort of back channel wink and nod by the feds at a promise by HPE to keep building the big computers.
I’ll miss them though. While their systems weren’t always the best, when you got your problems escalated to their R&D group, you got to work with some cool people. I imagine those people will get sucked into HPE and/or get fed up and defect to Intel pretty quickly.