TR Forums

OK I joked in another thread about there being a market for a "Personal Supercomputer" for consumers starting at $100,000 and lower. Well I googled Personal Supercomputers and I was expecting small clusters of cheap or older computers someone put together just for giggles, not something in a single PC size case.

Turns out, Nvidia does make one, The DGX Station:
https://www.nvidia.com/en-us/data-center/dgx-station/

I can see building a supercomputer cluster out of cheap old hardware just to do it, but I what would I use one for?

Don't worry about it. You're never ever going to build one any way.

boing wrote:

Don't worry about it. You're never ever going to build one any way.

I still want to know what I could use one for. Or build one out of a bunch of Raspberry Pi's since I don't have the space for multiple computers.

whm1974 wrote:

OK I joked in another thread about there being a market for a "Personal Supercomputer" for consumers starting at $100,000 and lower. Well I googled Personal Supercomputers and I was expecting small clusters of cheap or older computers someone put together just for giggles, not something in a single PC size case.

Turns out, Nvidia does make one, The DGX Station:
https://www.nvidia.com/en-us/data-center/dgx-station/

I can see building a supercomputer cluster out of cheap old hardware just to do it, but I what would I use one for?

What would you use one for? Right from the link you posted:

"Designed for your office, NVIDIA DGX Station™ is the world’s first personal supercomputer for leading-edge AI development."

The idea of a personal super computer is not a new thing. SuperMicro and others have such high end systems support multi GPUs for Teslsa and/or Xeon Phi. I've even heard of mixed brand setups with nVidia Tesla + AMD FirePro or nVidia Tesla + Intel Xeon Phi to follow the idea of using-the-best-tool-for-the-job-and-hope-our-code-doesn't-get-eaten-by-interconnect-overhead. Nowadays, you can get a host Xeon Phi system and then add some Tesla cards.

As for uses, they're often the development machines for larger clusters. So instead of heavily investing into racks at first, they get a couple of these beasts to test code on, then decide where to go from there. These workloads currently are mostly deep learning to follow the newest buzzword but scientific and engineering uses have niches filled by these systems. I'd normally say medical imaging and high resolution video editing but they don't scale as well as one would think across multiple video cards (before off getting a fast single card etc.).

Personal "program-trading" with neural networks and genetic algorithms for an ever-changing predictive model. If it does well and you have enough capital for trading it should pay for itself quickly..

Before GPUs were a thing, people made Beowulf clusters as personal supercomputers. Here is a neat one in a single "case":

http://knowm.org/24-core-linux-cluster- ... from-ikea/

You can do the same with a bunch of Pi as has already been mentioned. Physics simulations, whatever.

Scrotos wrote:

Before GPUs were a thing, people made Beowulf clusters as personal supercomputers. Here is a neat one in a single "case":

http://knowm.org/24-core-linux-cluster- ... from-ikea/

You can do the same with a bunch of Pi as has already been mentioned. Physics simulations, whatever.

Something like that should be even easier to build now than it was back then.

You know, I've seen several Beowulf computers but never any Beowulf software. How does one even write something like that? Is Beowulf still viable in the age of distributed systems like Hadoop/Cassandra/Spark/etc?

Yeah, the idea hasn't really changed any. When I was in college I setup a small Beowulf cluster as a Folding@Home farm, and I installed distcc on it for compiling stuff. It worked pretty well for that. At the time what I used for distributed applications was MPI written in C++. It's just way easier these days to buy some time on Azure or EC2 and use whatever language you want instead of building everything yourself and running it directly on the hardware.

Redocbew wrote:

Yeah, the idea hasn't really changed any. When I was in college I setup a small Beowulf cluster as a Folding@Home farm, and I installed distcc on it for compiling stuff. It worked pretty well for that. At the time what I used for distributed applications was MPI written in C++. It's just way easier these days to buy some time on Azure or EC2 and use whatever language you want instead of building everything yourself and running it directly on the hardware.

So I would need one master system for control and at least how many slave computers to be useful enough to bother, At least four?

whm1974 wrote:

So I would need one master system for control

Could probably use your existing desktop (or file server, if you have one) for this.

whm1974 wrote:

and at least how many slave computers to be useful enough to bother, At least four?

Useful enough for what? It depends on what you're using the cluster for.

just brew it! wrote:

whm1974 wrote:

So I would need one master system for control

Could probably use your existing desktop (or file server, if you have one) for this.

whm1974 wrote:

and at least how many slave computers to be useful enough to bother, At least four?

Useful enough for what? It depends on what you're using the cluster for.

You know, I'm wondering if Threadripper or a Core i9 box with high core counts could beat a cluster made from older computers?

whm1974 wrote:

You know, I'm wondering if Threadripper or a Core i9 box with high core counts could beat a cluster made from older computers?

Depends on how old, and how many of them you use in the cluster. If you've got a small number of machines running everything from a single server isn't a problem. As the cluster grows you've got to think about how best to distribute the workload to the nodes in your cluster. Also, if you've got machine of differing speeds that'll complicate things further. In the small cluster I had(which I think was four, maybe five machines in total) most of the nodes were about equal speed-wise except for one machine which was older than the others. Taking it out of the cluster actually sped things up overall.

Redocbew wrote:

whm1974 wrote:

You know, I'm wondering if Threadripper or a Core i9 box with high core counts could beat a cluster made from older computers?

Depends on how old, and how many of them you use in the cluster. If you've got a small number of machines running everything from a single server isn't a problem. As the cluster grows you've got to think about how best to distribute the workload to the nodes in your cluster. Also, if you've got machine of differing speeds that'll complicate things further. In the small cluster I had(which I think was four, maybe five machines in total) most of the nodes were about equal speed-wise except for one machine which was older than the others. Taking it out of the cluster actually sped things up overall.

It did suddenly dawn on me that one of today's HEDT systems would be considered a supercomputer by itself about ten years ago.

whm1974 wrote:

It did suddenly dawn on me that one of today's HEDT systems would be considered a supercomputer by itself about ten years ago.

Well... you'd probably need to go back more than 10, since CPU performance has stagnated lately. But yes, it is very much a sliding scale as technology improves.

In terms of raw compute power, a modern smartphone is roughly on par with a room full of Cray-1 supercomputers from the 1970s.

just brew it! wrote:

In terms of raw compute power, a modern smartphone is roughly on par with a room full of Cray-1 supercomputers from the 1970s.

That's roughly true. Nobody uses the benchmarks that were in vogue back in the day, but the Cray 1 from 1977 is quoted as reaching 160 MFLOPS; the latest Snapdragon phone cores offer something like 3 - 10 GFLOPS, or roughly 20 to 100 times as much... and of course that's on a single core, but there are 3 more high-performance cores and 4 more lower-performance cores. So you really would need a a lot of Cray 1s... but more than a roomful, because each Cray 1 took up most of a room itself:

wikipedia wrote:

The initial model, the Cray-1A, weighed 5.5 tons including the Freon refrigeration system. Configured with 1 million words of main memory, the machine and its power supplies consumed about 115 kW of power; cooling and storage likely more than doubled this figure.

I took a tour of Boeing Computer Services back in the mid-80s, and they had a Cray quietly humming over in one corner of the building. The guy who was giving the tour apologized because they really felt like it belonged in its own darkened room full of CO₂ fog, But we were all computer geeks on the tour and were suitably awed by the thing all the same, even if it looked more like modernist furniture decorated with a few blinkenlightz.

It didn't take Cray long to ramp up the FLOPS -- they were in the GFLOP range by then (1985 or so). But still, those were million-dollar machines using enormous amounts of electricity and heroic cooling systems. And today much more than that fits in your pocket and runs off a battery. To quote Paul Simon "These are days of miracles and wonder"

just brew it! wrote:

whm1974 wrote:

It did suddenly dawn on me that one of today's HEDT systems would be considered a supercomputer by itself about ten years ago.

Well... you'd probably need to go back more than 10, since CPU performance has stagnated lately. But yes, it is very much a sliding scale as technology improves.

In terms of raw compute power, a modern smartphone is roughly on par with a room full of Cray-1 supercomputers from the 1970s.

Did they even have quad core consumer CPUs back then? Sad to say I don't recall.

whm1974 wrote:

just brew it! wrote:

whm1974 wrote:

It did suddenly dawn on me that one of today's HEDT systems would be considered a supercomputer by itself about ten years ago.

Well... you'd probably need to go back more than 10, since CPU performance has stagnated lately. But yes, it is very much a sliding scale as technology improves.

In terms of raw compute power, a modern smartphone is roughly on par with a room full of Cray-1 supercomputers from the 1970s.

Did they even have quad core consumer CPUs back then? Sad to say I don't recall.

If you mean 1970s, then no. If you mean 10 years ago, then yes that's right about when quad core started to become widely available for what most people would consider to be "consumer" platforms (e.g. Intel Core 2 Quad, AMD Phenom X4).

If you look at the June 2007 Top 500 list, the tail-end of the list comes out at around 4TFlops to be considered in the Top 500: https://www.top500.org/list/2007/06/?page=5

Running the right kind of peak-efficiency code (that is what the Top 500 uses via Linpack), a 10 core 7900X using AVX 512 with all cores running at 3.3 GHz (probably doable) can generate a little over 1 Tflop.

So a 10 core chip isn't quite in the top 500 range from 10 years ago, but a dual-socket server system that is using the higher core count Xeon parts would probably be very close. [Edit: A dual-socket 28 core Xeon Gold executing AVX-512 and running all cores at 2.3GHz would be just a hair above 4TFlops (theoretical).]

UberGerbil wrote:

just brew it! wrote:

In terms of raw compute power, a modern smartphone is roughly on par with a room full of Cray-1 supercomputers from the 1970s.

That's roughly true. Nobody uses the benchmarks that were in vogue back in the day, but the Cray 1 from 1977 is quoted as reaching 160 MFLOPS; the latest Snapdragon phone cores offer something like 3 - 10 GFLOPS, or roughly 20 to 100 times as much... and of course that's on a single core, but there are 3 more high-performance cores and 4 more lower-performance cores. So you really would need a a lot of Cray 1s... but more than a roomful, because each Cray 1 took up most of a room itself:

wikipedia wrote:

The initial model, the Cray-1A, weighed 5.5 tons including the Freon refrigeration system. Configured with 1 million words of main memory, the machine and its power supplies consumed about 115 kW of power; cooling and storage likely more than doubled this figure.

I took a tour of Boeing Computer Services back in the mid-80s, and they had a Cray quietly humming over in one corner of the building. The guy who was giving the tour apologized because they really felt like it belonged in its own darkened room full of CO₂ fog, But we were all computer geeks on the tour and were suitably awed by the thing all the same, even if it looked more like modernist furniture decorated with a few blinkenlightz.

It didn't take Cray long to ramp up the FLOPS -- they were in the GFLOP range by then (1985 or so). But still, those were million-dollar machines using enormous amounts of electricity and heroic cooling systems. And today much more than that fits in your pocket and runs off a battery. To quote Paul Simon "These are days of miracles and wonder"

I felt like a had approached the power of our ancient campus IBM 360 with its 512K maximum core allocation when I got my first 7MHz 8086 machine with its 8087 math coprocessor (purchased a bit later) and 640K of ram in 1983. But its not just the compute power of the core. You need the capacity to feed massive amounts of data through the core and then store the results. My little PC had just two 5 inch floppy drives for storage. That IBM 360 had many disk stacks and tape machines to load up data and make backups. Sometimes you submitted just your keypunched deck and sometimes you gave them your data tape for the program to process. Plus, they had those massively fast line printers. Right around that time frame, they upgraded to a sytem 4300 running VMS with terminals. The point is that a supercomputer is not defined solely by its GFLOPS designation but also by its ability to continuously feed data to the core and spool off the results and the whole file system journaled and backed up so it can be restored from any starting point.

just brew it! wrote:

whm1974 wrote:

just brew it! wrote:

Well... you'd probably need to go back more than 10, since CPU performance has stagnated lately. But yes, it is very much a sliding scale as technology improves.

In terms of raw compute power, a modern smartphone is roughly on par with a room full of Cray-1 supercomputers from the 1970s.

Did they even have quad core consumer CPUs back then? Sad to say I don't recall.

If you mean 1970s, then no. If you mean 10 years ago, then yes that's right about when quad core started to become widely available for what most people would consider to be "consumer" platforms (e.g. Intel Core 2 Quad, AMD Phenom X4).

If you told me that ten years ago I could a build a Core 2 Quad system w/ 4GB of memory and still be able to use it in 2017, I would have died laughing at you.

Yup, two of the 3 PCs in my house are Core 2 Quads, the other is a Core i7 920.

I think the point he was making was the stagnation of CPUs. 10 years ago we were going through CPU models pretty quickly and a CPU from 1997 would have been painful to use in 2007.

yeah it seems the only time intel really pushes hard for change is when AMD gives them a reason to do so, otherwise they coast along barely inovating. It is absolutely crazy that you can still be fine with a 10 year old rig at this point, it just shows how badly intel has been milking people. Now that AMD is pushing them again, hopefully we will see some good things out of intel.

Starfalcon wrote:

yeah it seems the only time intel really pushes hard for change is when AMD gives them a reason to do so, otherwise they coast along barely inovating. It is absolutely crazy that you can still be fine with a 10 year old rig at this point, it just shows how badly intel has been milking people. Now that AMD is pushing them again, hopefully we will see some good things out of intel.

Your statement bizarrely implies that Intel is responsible for the fact that some types of software aren't much more demanding today than they were 10 years ago... although that statement isn't exactly true if you ever read TR's reviews and look at their frame time analysis graphs.

chuckula wrote:

Starfalcon wrote:

yeah it seems the only time intel really pushes hard for change is when AMD gives them a reason to do so, otherwise they coast along barely inovating. It is absolutely crazy that you can still be fine with a 10 year old rig at this point, it just shows how badly intel has been milking people. Now that AMD is pushing them again, hopefully we will see some good things out of intel.

Your statement bizarrely implies that Intel is responsible for the fact that some types of software aren't much more demanding today than they were 10 years ago... although that statement isn't exactly true if you ever read TR's reviews and look at their frame time analysis graphs.

It's a little bit of both. If CPU performance had improved at a faster pace, software would've been written to take advantage of it. But without consumers clamoring for faster CPUs (since the software doesn't need them), there's also less incentive to develop them (until someone else comes along and threatens to upset your existing pricing structure).

It's also easier to add cores than boost single-thread performance, and effectively leveraging additional cores is difficult unless you've got a trivially parallel use case (like media encoding or many server workloads).

just brew it! wrote:

chuckula wrote:

Starfalcon wrote:

yeah it seems the only time intel really pushes hard for change is when AMD gives them a reason to do so, otherwise they coast along barely inovating. It is absolutely crazy that you can still be fine with a 10 year old rig at this point, it just shows how badly intel has been milking people. Now that AMD is pushing them again, hopefully we will see some good things out of intel.

Your statement bizarrely implies that Intel is responsible for the fact that some types of software aren't much more demanding today than they were 10 years ago... although that statement isn't exactly true if you ever read TR's reviews and look at their frame time analysis graphs.

It's a little bit of both. If CPU performance had improved at a faster pace, software would've been written to take advantage of it. But without consumers clamoring for faster CPUs (since the software doesn't need them), there's also less incentive to develop them (until someone else comes along and threatens to upset your existing pricing structure).

It's also easier to add cores than boost single-thread performance, and effectively leveraging additional cores is difficult unless you've got a trivially parallel use case (like media encoding or many server workloads).

Yeah all the low hanging fruit has been picked clean for improving single threaded performance. it will most likely take some major breakthroughs to farther large increases in IPC.

Well I guess the new Alienware Area-51 could be used as a Personal Supercomputer:
http://techreport.com/news/32085/area-5 ... -core-cpus

Apple's marketing department already tried that. Didn't work out so well.

Redocbew wrote:

Apple's marketing department already tried that. Didn't work out so well.

You mean their "trash can" workstation they came out with about four years ago? Yeah I wouldn't buy one either.