TR Forums

With todays CPUs using so little power and graphics cards using so much more then a CPU why cant we have add on CPU cards since a dual slot PCI-E 3.0 graphics slot has a lot of bandwidth to connect one Intel CPU to directly to another. Any big graphics card layout card could hold any CPU lets say a i7-4770k, along with SO-DIMM memory slots or imbedded memory of some kind along with a nice GPU type cooler and PCI_E power connectors would supply plenty of voltage.

I think it would be awesome if it could be done! Imagine your 4770k gtx780 gaming rig with a secondary 4770k add in card for CPU intensive workloads. Any mainstream gaming rig could be turned into a Monster Worksation just by adding a PCI-E card.

Any thought on why this has not been done yet? the closet thing to it i can think of would be when Intel tried making a high end graphics card then turned it into a compute card like nvidias tesla....I just cannot think of the name of it right now!

Or is it not being done so you have to pay a fortune on a dual socket WS board along with super expensive CPUs ??

You mean like this?



CPU cards have been around for a long time. I remember seeing a 386 card that plugged into ye olde ISA bus.

It is certainly possible, but it won't behave like a native multi-core or SMP system due to the lack of cache coherency between the CPU on the PCIe card and the main CPU, and the fact that all I/O would still need to be routed through the main CPU's I/O subsystem. So you'd essentially be forced to use it in a co-processor type configuration, just like with GPU-based computing today. Given these limitations, you might as well just use existing GPU-based computing solutions.

vargis14 wrote:

Any thought on why this has not been done yet? the closet thing to it i can think of would be when Intel tried making a high end graphics card then turned it into a compute card like nvidias tesla....I just cannot think of the name of it right now!

You mean the Xeon Phi? They are starting to do quite well in the supercomputing space now but Intel is introducing a proper socketed version in the next gen as well. As JBI pointed out, the PCIe bottleneck is an issue.

Thanks for the responses...

Yes like that Jihadjoe but that does not look like it would fit into a normal motherboard.

Codevine you nailed it the Xeon Phi was the re-purposed Intel attempt to make a high end GPU.

JBI...about how many GB/sec over 8-16GB/sec would you need in actual bandwidth for the CPUs to work together on a project like video encoding or a benchmark like CINEBENCH?
Also say you added a card like jihadjoe posted that would fit in a regular motherboard would it even work along with the motherboard CPU or would it have to have it's own work assigned to it or would it work like a totally separate computer?
Say i am doing video encoding with CyberLink MediaEspresso 6 or 5.5 with a [email protected] and a GTX 770. With the CPU alone it takes 7 min 35 seconds converting a 3.3GB 720p MKV file (2h:12mins) to Apple Mp4 , with the GTX 770 and hardware acceleration enabled it takes 6 min 45 seconds. taking only 50 seconds off converting that 3.3 GB file. Is the difference that small because of bandwidth between the cpu and GPU or is it just the coding in general?
Do not let me waste much of your valuable time doing whatever big and thorough review that is coming up next
Any other gerbils can waste as much time as you want

I have to say it would be really neat to have any i5 quad core or i7- 2600, 3770 or a 4770 riding shotgun in a socket 2011 platform for the super user...but even in any mainstream socket 1155 or socket 1150 system also. Even if you had to assign its own processes or projects to it. As long as it would work on the same OS and not have to be just a separate system and OS just taking up space in your own tower. I left out thew K models since in my eyes overclocking just would not make sense in this type of application.

codedivine wrote:

vargis14 wrote:

Any thought on why this has not been done yet? the closet thing to it i can think of would be when Intel tried making a high end graphics card then turned it into a compute card like nvidias tesla....I just cannot think of the name of it right now!

You mean the Xeon Phi? They are starting to do quite well in the supercomputing space now but Intel is introducing a proper socketed version in the next gen as well. As JBI pointed out, the PCIe bottleneck is an issue.

A coworker expressed his disappointment that China's newest big iron took the top spot- so I had to see what they put in it. It's all Intel- Ivy CPUs with Xeon Phi co-processors.

vargis14: You mentioned CyberLink, and in the process have answered your own question. The system is capable of much more than that excuse for a software company will ever get out of it.

Airmantharp you can call me Varg:) the 14 is not important.

Anyways You are right I probably did answer my own question. But with the right code what would be needed to bandwidth wise to get it a CPU card to work in a Windows OS to get even 75% of the add on CPUs performance? or get it to work alone on projects you assign to it specifically. Say you want to FOLD at home with a i7 and 8 threads 24/7 but you do not want to use your main CPU..or any other project like encoding a full blueray to MKV but you want your main CPU for gaming or something else?

I guess I am also asking if you can make a dual socket motherboard with a graphics card type CPU card?

vargis14 wrote:

JBI...about how many GB/sec over 8-16GB/sec would you need in actual bandwidth for the CPUs to work together on a project like video encoding or a benchmark like CINEBENCH?

If what you mean is traditional SMP/multi-core (i.e. have the additional CPU cores appear just like the cores on your main CPU), it isn't that simple. This would would likely require significant modification to the CPU, motherboard, BIOS, and OS.

In a normal multi-core (SMP) system, all CPUs share a single memory address space. Making this work efficiently requires both high bandwidth and low latency between the cores. It also requires a mechanism for maintaining cache coherency -- i.e. if CPU A writes to a memory location which is in its internal cache, and CPU B also has a copy of that memory location in *its* cache, then steps must be taken to ensure that both copies stay consistent. Either A needs to flush that location out of its cache to DRAM and inform CPU B that it needs to reload that memory location from DRAM, or it needs to send the modified data directly to CPU B so that B can update its cache. This all needs to take place at the hardware level, with minimal (few tens of nanoseconds at most) latency; desktop motherboards (and most desktop CPUs) aren't going to have support for this.

If you give each CPU its own dedicated RAM, then there's no cache coherency issue. But then the second CPU acts more like a coprocessor, or a completely separate system in its own right.

OTOH if you mean using it as an additional compute resource for CUDA or OpenCL apps, then yeah it should be relatively straightforward to do. But in that case you'd probably get better bang for the buck using a GPU.

When the code is broke, theoreticals need not apply .

CyberLink could make the process far quicker- they could even offload parts of it to the CPU's GPU if they cared to try (instead of using the whole thing, which would compromise quality). But they'd have to be financially motivated to do so, and I don't see that company being financially motivated to do anything except to ensure that they're products lock their customers out at least once a year. I stopped feeding the beast, even though they're the only ones that make decent (yet astoundingly horrific) software Blu-ray player.

Airmantharp wrote:

decent (yet astoundingly horrific)

This combination of words confuses me.

Thank's JBI and Airman.
I guess we will have to wait for the future when everything is a like 5 nanometers or whatever is smaller then a nanometer "quantanometer..yeah i made that up" and a motherboard that comes with something like 10 Mini PCI express type CPU sockets that the user can add the 8 core Mini PCI express CPUs as needed with 100% multithreaded coding that somehow just makes single threaded applications faster with 100% scaling just by adding more and more Octa core CPUs. By this time we will not need heatsinks

just brew it! wrote:

Airmantharp wrote:

decent (yet astoundingly horrific)

This combination of words confuses me.

It's decent because it plays Blu-rays. The Consortium made sure, and continues unabated to make sure, that playing Blu-rays on a PC was damn near impossible. So I give them credit. Otherwise I feel their software is an abomination.

vargis14 wrote:

Thank's JBI and Airman.
I guess we will have to wait for the future when everything is a like 5 nanometers or whatever is smaller then a nanometer "quantanometer..yeah i made that up"

picometer

vargis14 wrote:

and a motherboard that comes with something like 10 Mini PCI express type CPU sockets that the user can add the 8 core Mini PCI express CPUs as needed with 100% multithreaded coding that somehow just makes single threaded applications faster with 100% scaling just by adding more and more Octa core CPUs. By this time we will not need heatsinks

If such a tech comes to pass, systems with multiple CPU sockets will still likely be relegated to the server space, just like they are today. The additional cost and complexity just isn't worth it unless you're running a datacenter or compute farm, and need that flexibility and/or density of compute power.

Picometer....sheesh that sounds like some musicians tool that counts the number of times you picked a guitar or some thing. Thanks for the answer anyways and as always "have nice day"

just brew it! wrote:

vargis14 wrote:

JBI...about how many GB/sec over 8-16GB/sec would you need in actual bandwidth for the CPUs to work together on a project like video encoding or a benchmark like CINEBENCH?

If what you mean is traditional SMP/multi-core (i.e. have the additional CPU cores appear just like the cores on your main CPU), it isn't that simple. This would would likely require significant modification to the CPU, motherboard, BIOS, and OS.

In a normal multi-core (SMP) system, all CPUs share a single memory address space. Making this work efficiently requires both high bandwidth and low latency between the cores. It also requires a mechanism for maintaining cache coherency -- i.e. if CPU A writes to a memory location which is in its internal cache, and CPU B also has a copy of that memory location in *its* cache, then steps must be taken to ensure that both copies stay consistent. Either A needs to flush that location out of its cache to DRAM and inform CPU B that it needs to reload that memory location from DRAM, or it needs to send the modified data directly to CPU B so that B can update its cache. This all needs to take place at the hardware level, with minimal (few tens of nanoseconds at most) latency; desktop motherboards (and most desktop CPUs) aren't going to have support for this.

If you give each CPU its own dedicated RAM, then there's no cache coherency issue. But then the second CPU acts more like a coprocessor, or a completely separate system in its own right.

OTOH if you mean using it as an additional compute resource for CUDA or OpenCL apps, then yeah it should be relatively straightforward to do. But in that case you'd probably get better bang for the buck using a GPU.

This. I actually work for a company that does this, though for an FPGA based coprocessor. The coprocessor memory is part of the system memory map and the the coproecssor execution units are dispatched instructions as they come up in program execution. We spend significant resources maintaining a custom Linux distribution to support the coprocessor and yet large data transfers to coprocessor memory still kill performance. The PCIe bus just can't hold a candle to the inter-processor QPI bus. We overcome that by being massively faster at performing computation on the data.

If you want another general purpose CPU available in your system, buy a multi-socket motherboard.

--SS

SecretSquirrel wrote:

If you want another general purpose CPU available in your system, buy a multi-socket motherboard.

--SS

Someone was going to say it . When Intel went down the QPI road, enabling efficient operation of >2 socket systems with the point-to-point NUMA-aware architecture, multi-socket became as useful as it already was with AMD and Hyper-Transport.

So SS, your co-processor is NUMA-aware? That's pretty badass, even though it gets throttled by the PCIe bus!

Airmantharp wrote:

SecretSquirrel wrote:

If you want another general purpose CPU available in your system, buy a multi-socket motherboard.

--SS

Someone was going to say it . When Intel went down the QPI road, enabling efficient operation of >2 socket systems with the point-to-point NUMA-aware architecture, multi-socket became as useful as it already was with AMD and Hyper-Transport.

So SS, your co-processor is NUMA-aware? That's pretty badass, even though it gets throttled by the PCIe bus!

Yes, it is NUMA aware. In fact, the first generation actually plugged into a CPU socket on the host motherboard, but it is on an FSB, not QPI chipset. The second generation is connected via a PCIe bridge card. The FSB connected version is not nearly so limited in memory bandwidth, but there are other complications there.

--SS

vargis14 wrote:

Picometer....sheesh that sounds like some musicians tool that counts the number of times you picked a guitar or some thing. Thanks for the answer anyways and as always "have nice day"

I thought it sounded more like a device for measuring the quality of salsas.

Given how truly tiny a picometer is, it tends not to come up much in everyday conversation.

vargis14 wrote:

With todays CPUs using so little power and graphics cards using so much more then a CPU why cant we have add on CPU cards since a dual slot PCI-E 3.0 graphics slot has a lot of bandwidth to connect one Intel CPU to directly to another. Any big graphics card layout card could hold any CPU lets say a i7-4770k, along with SO-DIMM memory slots or imbedded memory of some kind along with a nice GPU type cooler and PCI_E power connectors would supply plenty of voltage.

Basically what you're describing is a blade server.

There is an enclosure with a backplane that links a bunch of individual servers together. The problem is they are a bunch of individual servers, and they need to be managed like a bunch of individual servers.

Of course, most software doesn't take up all of the processing power of modern hardware, so we have to find things to do with that power. Like run VMs which share the hardware.

What you really want is probably something like the most recent SPARC processors. Lots of individual cores with lots of hardware threads that can be partitioned for different programs.

Flatland_Spider wrote:

What you really want is probably something like the most recent SPARC processors. Lots of individual cores with lots of hardware threads that can be partitioned for different programs.

Actually, it sounds to me like what he really wants is expandable multi-socket n-way SMP on an affordable consumer platform. Probably not gonna happen any time soon (if ever).

vargis14 wrote:

Picometer....sheesh that sounds like some musicians tool that counts the number of times you picked a guitar or some thing. Thanks for the answer anyways and as always "have nice day"

I wish I could play piconotes on my guitar.

vargis14 wrote:

With todays CPUs using so little power and graphics cards using so much more then a CPU why cant we have add on CPU cards since a dual slot PCI-E 3.0 graphics slot has a lot of bandwidth to connect one Intel CPU to directly to another. Any big graphics card layout card could hold any CPU lets say a i7-4770k, along with SO-DIMM memory slots or imbedded memory of some kind along with a nice GPU type cooler and PCI_E power connectors would supply plenty of voltage.

I think it would be awesome if it could be done! Imagine your 4770k gtx780 gaming rig with a secondary 4770k add in card for CPU intensive workloads. Any mainstream gaming rig could be turned into a Monster Worksation just by adding a PCI-E card.

Any thought on why this has not been done yet? the closet thing to it i can think of would be when Intel tried making a high end graphics card then turned it into a compute card like nvidias tesla....I just cannot think of the name of it right now!

Or is it not being done so you have to pay a fortune on a dual socket WS board along with super expensive CPUs ??

The bandwidth would be woefully insufficient, but I guess there's always the chance of implementing a separate system-on-a-card. That's been done before - I know there are cards used as industrial device controllers that amount to discrete systems on a PCB that can be plugged into a workstation. Another example was the plethora of PC-on-a-cards sold for pre-OS X Macs. And many years ago you could take what amounted to a Celeron system on a PCB, plug it into a musty old motherboard, and it'd effectively override the BIOS and let you use the new hardware while using the old case and motherboard as a kind of husk. But that's not really what you're talking about here.