HP Labs talks about "The Machine" - a new way to think about and design computing architecture

[Ilie Pandia]

In this 30 minute presentation Martin Think from HP Labs will discuss "The Machine".

"The Machine" is a project they are working on to address the current limitations of speed and size in the IT industry by rethinking the computer architecture from the ground up.

It's a bit geeky, but entertaining and informative.

Of course we can speculate just how far the "black projects" have gone if this is now out in the public domain and what the implications are for everything being stored in this huge non-volatile memory?

They want to make this open sourced and free... so that is good news as it has the potential to empower a lot of people that lack the processing power to do their research (just to give an example).

Enjoy!

Source: https://youtube.com/watch?v=JzbMSR9vA-c

[Aurelius]

The Dr. PP "chip" + an OS + photonics


... currently all already in use (for many years now) by the "money pits".

[ThePythonicCow]

Posted by Ilie Pandia (here)
In this 30 minute presentation Martin Think from HP Labs will discuss "The Machine".

Awesome - thanks - I would have missed this otherwise.

When I was working at Silicon Graphics (SGI) up until 2008, our large machines were single-system-image (not network cluster) computers with approximately 10 terabytes of main memory and 1000 CPUs. They used rather expensive copper interconnects between racks, Intel CPU's and volatile DRAM memory, with expensive hierarchical storage options for the clustered disk drives (rotating magnetic disks.)

("Volatile" means they lose their contents on power off.)

This HP effort replaces copper with fiberoptic for the cpu-to-memory interconnects, and replaces volatile DRAM with non-volatile memristors for the memory. HP has been leading the effort to make memristors a commercially viable replacement for all of Flash, SSD, DRAM and SRAM, as explained here.

The change of copper to fiber for the interconnect enables expanding the large system memory from the 10 terabytes I knew, to the 160 petabytes he spoke of. That's going from about 44 bits of main memory address space to 58 bits, a 16,000 times increase, and doing it with cheaper technology. That would put a rather unbearable strain on the non-volatile storage (aka disk) subsystem, but for their other change, from volatile DRAM to non-volatile memristor. Loading 160 petabytes from conventional (for the present) disks would take ... like forever. An enterprise fibre channel (used to connect computers to disks) might run at up to 16 gigabits per second, or, figuring a rough 10 bits per byte over a comm line (with error checking and framing) 1.6 gigabytes per second. Dividing 160 petabytes by 1.6 gigabytes per second yields 100 million seconds, or about 3.17 years.

Yup - over three years to load that memory from disk, or to store it back out to disk, using just one Fibrechannel.

Having non-volatile memristor memory that doesn't need to be continually shuffled back and forth between memory and disk gets rather important.

This changes the software stack ... dramatically. Every single operating system I've worked on since the late 1970's had fundamental assumptions that get thrown out the window with this.

But the advantages are sufficient that this locomotive barrelling down the tracks at us will not be stopped.

Almost makes me want to get back in the saddle again.

(If anyone from HP is reading this, search LKML for pj AT sgi.com to find some of what I was working on, in particular cpusets, the last few years of my time at SGI.)

[gripreaper]

Wow, I would have missed this too. Thanks Ilie. I too am speculating how far advanced this technology already is in the covert world, considering it is now going mainstream.

The implications of this architecture (as Paul points out) completely makes current architecture obsolete, and makes data fully integrated and super fast.

I'm also speculating on how us commoners can take advantage (as investors and or computer users) of this paradigm shift. I will be checking in on this thread for updates from time to time.

[Operator]

It somehow reminds me a bit of the power they demonstrated using transputer workstations more than 25 years
ago, while I was working for the Dutch navy.

I've always been anticipating when that technology would appear in the public sector worldwide. But that never
happened. Quad core processors (which by far are still no transputers) are more or less the best of multi-core
processors available to the public. But a real good OS that can dynamically configure the topology of the processors
is still missing ... I wonder what will get out of this. They have learned by now not to put too much computing
power in the hands of the public.