backtop


Print 30 comment(s) - last by saratoga.. on Apr 25 at 8:56 PM


The DRC module plugs directly into a Socket 940 AMD Opteron motherboard
DRC has announced its newest FPGA that drops into AMD's Socket 940

The Register has a fairly in depth look at one start-up's attempt to capitalize on AMD's HyperTransport interface -- a reprogrammable coprocessor that can drop into any Socket 940 socket.  The company, DRC, built its programmable coprocessor on Xilinx Virtex4 field programmable gate array integrated circuits. 

For specialized industries, a dynamic coprocessor is exactly what the doctor ordered; low overhead for extremely specific tasks such as vector math or collision detection.  Companies already pay thousands to millions of dollars to have such overly specific algorithms ported to custom FPGA processors, but the kicker for DRC is that the chip can be integrated into a multi-slot Opteron server running the correct software.

Each series of coprocessors unveiled by the company uses the standard HyperTransport (HT) interface to communicate with the main processor.  The low end coprocessor, the DRC100-L60ES, uses a 200MHz by 8-bit HT link.  DRC's two high end modules, the DRC100-L60 and the DRC110-L160 both use a 400MHz by 16-bit interface instead.  DRC coprocessors range in size from 50,000 to 140,000 programmable gates and all three can utilize 6.4GBps between the Xilinx FPGA and the DDR400 memory bank.

Each DRC module starts at about $4,500.  Competing proprietary systems from SGI and IBM easily cost four times that and generally require additional proprietary hardware and contracts to support.


Comments     Threshold


This article is over a month old, voting and posting comments is disabled

Seperate socket?
By Hypernova on 4/24/2006 4:02:49 AM , Rating: 1
Arent people buying duel sockets for duel CPUs? There should be a dedicated socket for coP's shouldn't there?




RE: Seperate socket?
By Gnarr on 4/24/2006 4:17:43 AM , Rating: 2
The question is availability. This can be used in ALL current opteron servers. That means you could take almost any 3-4 year old dualsocket opteron server and put one dualcore opty and one co-processor, and have all you needed for your task.

Would you really want to buy a new system with a new socket you could only use for a co-processor, when there is already a system available that can do the exact same thing?


RE: Seperate socket?
By chrisd154 on 4/24/2006 4:30:32 AM , Rating: 5
They buy dual sockets for dual CPUs. Two dueling processors would be rather counter-productive to say the least :-)


RE: Seperate socket?
By Griswold on 4/24/2006 5:03:31 AM , Rating: 3
Fight!


RE: Seperate socket?
By Googer on 4/25/2006 1:05:44 AM , Rating: 2
Quad Core has just arrived. So by removing a dual core and adding this fancy co-processor you will have more or just as many cpu's plus this specialized co-processor that is specialized for ISM. In the end you come out with a faster more capable system.

Think Physx, on AGEIA PhysX Chip can do more than four general purpose CPU's.

As a side note, it would be cool if AGEIA took the same concept and applied it to their PhysX Chip. That would allow it to have access to it's own RAM and a direct high speed link to the CPU.


RE: Seperate socket?
By Viditor on 4/24/2006 5:47:20 AM , Rating: 2
quote:
Arent people buying duel sockets for duel CPUs? There should be a dedicated socket for coP's shouldn't there?

Sometimes they are...but what about people who buy a 4P board (or an 8P board) and only need to populate 2 of the sockets with CPUs?
The other point is that using 1 DC Opty and 1 coP may be more efficient for your application...or what about 2 CPUs and 2 coPs?
IMHO, this system is probably the most ideal you can get because it keeps the cost down at the mobo manufacturing level (and should even reduce the cost of larger 4P and 8P boards from increased volume) while allowing users to customise their system for a much wider range of apps.
The fact that it's backwards compatable to any 940 system is icing on the cake!


RE: Seperate socket?
By ceefka on 4/24/2006 8:04:19 AM , Rating: 2
Could this be something for desktops as well? Imagine your PC extremely streamlined for the tasks you do most or simply running all that background stuff like your anti virus and firewall. To keep costs down, they can perhaps think up a solution on a PCI-E card.


RE: Seperate socket?
By Viditor on 4/24/2006 9:41:47 AM , Rating: 2
PCIE is too slow and doesn't connect directly to the cache of the chips like the cHT (coherent HT) links of the sockets do.


RE: Seperate socket?
By Worthalter on 4/24/2006 9:41:48 AM , Rating: 2
quote:
Could this be something for desktops as well? Imagine your PC extremely streamlined for the tasks you do most or simply running all that background stuff like your anti virus and firewall. To keep costs down, they can perhaps think up a solution on a PCI-E card.


sorry, you don't know what you're talking about.


RE: Seperate socket?
By Viditor on 4/24/2006 10:39:21 AM , Rating: 2
quote:
sorry, you don't know what you're talking about

It would probably be more helpful if you explained to him where he is mistaken...just a suggestion for next time m8.

In general, these co-processers (coP) are used for very specific functions. However, what you are thinking about (anti-virus, etc...) are specific apps not functions. It's extraordinarily expensive to make these coPs (~$10k each), so they would obviously not be practical in any desktop/consumer use...

A good example of a coP would be a dedicated Vector Processor
http://en.wikipedia.org/wiki/Vector_processor

In this case, you would be able to add the power of a vector processor to that of a general purpose processor and communicate with the cache of the general processor at HT speeds. This is infinately more efficient and faster than add-in cards or running many more general processers.


RE: Seperate socket?
By Wwhat on 4/25/2006 9:09:31 AM , Rating: 2
Uhm, the nforce chipset has that already, their 'activearmor' is a onchip firewall, alas it has bugs because nvidia doesn't seem to have capable sotwarewriters who also do not communicate with the hardware division during the testing phase so they release chips with bugs like the purevideo part on the orginal GPU and activearmor on their motherboardchipsets, but it seems they managed to now work around the issues via software though, anyway, there's your idea in practise already, not so daft after all then.



RE: Seperate socket?
By peternelson on 4/24/2006 4:57:04 PM , Rating: 2

Virtex 4 on PCI Express?

Already exists. eg Broaddown 4 from www.Enterpoint.co.uk

(among others).


Calc types
By PedroDaGr8 on 4/24/2006 10:37:27 AM , Rating: 2
quote:
Could this be something for desktops as well? Imagine your PC extremely streamlined for the tasks you do most or simply running all that background stuff like your anti virus and firewall. To keep costs down, they can perhaps think up a solution on a PCI-E card.

This is more of intended for specific applications, such as mutliparticle physics or somethign I am familiar with, QUantum Mechanical Modeling. Right now, when I run my calculations, I use around 24 Itanium 2 processors and around 48GB of ram. These applications use a lot of floating point calculations, if I remember correctly.




RE: Calc types
By Operandi on 4/24/2006 1:21:24 PM , Rating: 2
What about physics and AI processing for gaming? That seems like a at least possible application.


RE: Calc types
By saratoga on 4/24/2006 1:33:30 PM , Rating: 2
Not really. These are FPGAs, they're designed for custom applications. This means they're cheaper if you're the only one doing this sort of work. If theres actually a market for something (game accelerators), this is the last thing you want since FPGAs are amazing expensive (per chip) and much, much slower then a conventional part.

The reason they're used is that you can buy one and design your own processor to solve YOUR specific problem. I know labs that use FPGA based custom processors for things like real time processing of sensor data coming in a hundreds to thousands of MB per second that need all of that data crunched in a very specific in just a few milliseconds. These are the people who will want these chips. I doubt most other people will be interested, at least those without a masters in Comp Engineering a lot of Verilog/VHDL programming experience.


RE: Calc types
By peternelson on 4/24/2006 4:46:07 PM , Rating: 2
As I'm sure you are aware, FPGAs are ONLY "much slower" in CLOCK SPEED (compared to Pentium, Opteron etc). However, they have the benefit of MASSIVE PARALLELISM (millions of gates), and amazing internal memory bandwidth, plus customisable operations that are not in the instruction set of x86 processors. Therefore in many technical applications, FPGAs can deliver excellent acceleration, despite the seemingly low clocks.


RE: Calc types
By saratoga on 4/25/2006 8:56:49 PM , Rating: 2
"As I'm sure you are aware, FPGAs are ONLY "much slower" in CLOCK SPEED (compared to Pentium, Opteron etc)."

In what other way can a transistor be slower?

"However, they have the benefit of MASSIVE PARALLELISM (millions of gates), and amazing internal memory bandwidth,"

You're not making sense. FPGAs have far fewer transistors then conventional fabbed chips. Given the same verilog code, the FPGA will always be an order of magnitude slower then what a fab would give you.

The advtange is reprogrammablity and lower setup costs. You can fix your mistakes, and be up and running again in hours, not months if you find a bug. For that you lose speed.

" plus customisable operations that are not in the instruction set of x86 processors."

What the heck are you talking about? FPGAs are a way you can implement processors, including x86 ones, not a rivial to x86.

" Therefore in many technical applications, FPGAs can deliver excellent acceleration, despite the seemingly low clocks."

Ok, but that has nothing to do with whatever you're talking about. The clock speed is just the cost you pay for flexibility.


RE: Calc types
By Wwhat on 4/25/2006 9:26:00 AM , Rating: 2
So why must this be so prohibitively expensive? can we not now in 2006 make FPGAs for a more reasonable price, perhaps it would need some redesign but would it not create its own market once you did?
After all the whole FLASHRAM market andsoforth is moving along at a good pace and prices drop, surely by now we should be able to make this for more retail range prices too.
Imagine selling servers with one onboard for the cost of a normal server, you'd rule the market, and what if you could put it on normal systems, I think soon the linuxcommunity for instance would find many uses for it, and students could learn how to handle the designing for it etc.
It would put the excitemnt and experimentation from the past back in the hands of enthusiast and glasses wearing geeks :}


RE: Calc types
By Zoomer on 4/25/2006 11:23:02 AM , Rating: 2
Volume.


It's the software stupid...
By codeThug on 4/24/2006 10:33:03 AM , Rating: 2
"If DRC takes off, the company plans to bulk up from its current 13-person operation and to tap partners in different verticals to help out with the software work."

Don't get excited, there's no code for it.




RE: It's the software stupid...
By peternelson on 4/24/2006 4:33:21 PM , Rating: 2

There's plenty of "software" for it. The Virtex 4 takes standard HDL eg VHDL or verilog. You can get some at www.opencores or many others including Xilinx's OWN IP for tasks like Fast fourier transform, DES encryption etc.

The only difficuly is you need a synthesis tool (like a software compiler) to produce a bitstream for the Xilinx yourself. To target this particular chip you can't use the free Xilinx tools but need the pay-for license which does not run particularly cheap. Then again if you're paying $4500 for a fpga on a board, you can pay the tools cost.

I disagree that this is useless on PCI-E or even PCI. That is because the application may have little process intercommunication required. eg to run an ECM factorisation algorithm in the virtex, it does not need to talk much to the rest of the system aside from pass the result. Therefore PCI-E is entirely appropriate, and that is why there are many conventional cards with FPGAs onboard. Recently both CRAY and SGI have developed coprocessor modules connecting Virtex fpgas closely. There does exist an HTX (hypertransport slot) standard which IWill implement. However as most opteron boards do not use that, a direct plug in is a good idea. However, I think they would do well to target it at the forthcoming SOCKET F which should be future opterons on ddr2. I hope AMD provide some more news on that real soon.


RE: It's the software stupid...
By peternelson on 4/24/2006 4:41:48 PM , Rating: 2

Just to add a little, I already have and use an array of PCI fpga boards using the Xilinx Spartan 3 which is the baby brother of these Virtex 4 chips (just less logic available and a little slower). PCI is sufficient for many applications. However, I may shortly be migrating to a PCI-Express based Virtex 4 board. I would be unlikely to invest $4500 in a socket that is about to change. If they launched a socket F version and socket F opteron systems took over I would be more likely to invest in that. For some applications the hypertransport would be advantageous over even PCI express.


COP
By exdeath on 4/24/2006 10:35:56 AM , Rating: 2
Now this is what a Physics coprocessor should be like instead of the 133 MB/sec PCI card people are hyping.

Direct HT path to the main CPU, direct access to the full bandwidth memory bus, etc. No wasted bandwidth, just set the data arrays up in memory with the CPU, and send instructions to the physics COP to plow through them.

But we would have to see dual socket motherboards that are worthy of a gaming machine. Most dual CPU boards suck compared to the highest end gaming boards like the A8N32-SLI for gaming performance.




RE: COP
By peternelson on 4/24/2006 4:36:36 PM , Rating: 2
Physics coprocessor SHOULD be on PCI EXPRESS. It doesn't particularly need hypertransport, but it should have full duplex without contention. PCI Express is fine I hope Asus and BFGTech they migrate their physics boards to it soon.


my first thought...
By Saist on 4/24/2006 3:11:38 PM , Rating: 2
my first thought when I read this was Ageai's PhysX chip. That would probably do well in this use.




RE: my first thought...
By peternelson on 4/24/2006 5:05:42 PM , Rating: 2

In all probability a Xilinx FPGA (indeed several of them) or the equivalent Altera solution, was what they will have used to PROTOTYPE the design of the Physx chip. Once you have the design right, unit costs are much cheaper (apart from a one-time NRE nonreturnable engineering cost) by blowing the design in to a custom ASIC (application specific integrated circuit) and indeed that is what they have done. The ASIC based design will run a little faster too. However, ASICs cannot be reconfigured in the way that FPGAs can so they have less flexibility to correct errors or improve the design and lack the ability to optimise particularly for the current job in hand.


Math Processor
By AggressorPrime on 4/24/2006 4:45:51 PM , Rating: 2
I am guessing this is the math processor that AMD was talking about working with the HT bus.




RE: Math Processor
By peternelson on 4/24/2006 4:50:37 PM , Rating: 2
Yes AMD has been talking to Xilinx, and the technology here has been around for a while in some systems.

The OTHER acceleration vendor was CLEARSPEED(.com) CSX600, who are targeting maths applications specifically. Clearspeed is a multicore processor as opposed to a fully customisable array of logic gates. The two kinds of chips are targetted at different applications. In my opinion the Xilinx is more flexible because it can be reconfigured eg to do maths on minute then encryption the next then image processing the next and so on.


Subtitle correction
By peternelson on 4/24/2006 6:49:33 PM , Rating: 2
"DRC has announced its newest FPGA that drops into AMD's Socket 940"

The FPGA is actually a product of the company XILINX.

DRC make a board MODULE which contains an FPGA.

Therefore insert the word "module" after FPGA because DRC do not make FPGAs.




comment bug
By Wwhat on 4/25/2006 9:28:38 AM , Rating: 2
It seems clear what the bug is in this comment system, if someone posts while you are typing a post the cache differs from the present state and you get a 'oops error'.
Fascinating stuff that someone would code a commentsystem that isn't paralel for a big site...
This apart from the missing edit function of course.




"Nowadays, security guys break the Mac every single day. Every single day, they come out with a total exploit, your machine can be taken over totally. I dare anybody to do that once a month on the Windows machine." -- Bill Gates











botimage
Copyright 2014 DailyTech LLC. - RSS Feed | Advertise | About Us | Ethics | FAQ | Terms, Conditions & Privacy Information | Kristopher Kubicki