Interesting article about nVidia and ATI at Anandtech
http://www.anandtech.com/video/showdoc.html?i=1711
http://www.anandtech.com/video/showdoc.html?i=1711
NV30 shown running in hardware....but only at a few KHz...
- Thread starter McElvis
- Start date
It's not software, Sabastien... FPGAs *are* hardware. They're not hardwired silicon chips, but it IS the actual chip layout being run on them. In hardware.
Still a simulation though, so saying the NV30 being shown 'in hardware' is a bit of a stretch. That I'll agree to.
*G*
Still a simulation though, so saying the NV30 being shown 'in hardware' is a bit of a stretch. That I'll agree to.
*G*
Prometheus
Newcomer
Maybe someone should mention that Anand visited nvidia a month ago!! 
Prometheus
Newcomer
This is the Nv30 second version!!!
Mephisto said:Grall said:
Hey, my JAVA program runs on a Pentium 4, it runs in hardware, cool heh?
Hey wait a min.... Windows runs in hardware too..
Seriously though they are emulating hardware using software.. That was my interpritation of the IKOS box.. of course they are using hardware of sorts ... just not the nv30.
No, they are emulating hardware using hardware.
An FPGA is a chip that can reconfigure its "silicon" at runtime. If FGPA's could run at 300Mhz, and were cheap and easy to mass produce, then no one would bother fab'ing chips.
You'd simply buy an FPGA on a card for your computer with the most amount of transistors and fastest clock and then "software upgrade" your GPU by downing "tapeouts" (configurations) to the FPGA.
ATI and NVidia would then be software companies, selling you a GPU on a DVD-ROM.
Alas, FPGA's will never be as efficient and fast as directly fabbed designs, so I doubt we'll see that anytime soon.
But they are great for emulation/testing purposes.
An FPGA is a chip that can reconfigure its "silicon" at runtime. If FGPA's could run at 300Mhz, and were cheap and easy to mass produce, then no one would bother fab'ing chips.
You'd simply buy an FPGA on a card for your computer with the most amount of transistors and fastest clock and then "software upgrade" your GPU by downing "tapeouts" (configurations) to the FPGA.
ATI and NVidia would then be software companies, selling you a GPU on a DVD-ROM.
Alas, FPGA's will never be as efficient and fast as directly fabbed designs, so I doubt we'll see that anytime soon.
But they are great for emulation/testing purposes.
But aren’t you really making more of a technical argument when you say that the nv30 is running in hardware? I mean if you simply put an FPGA card in a machine it does not equate to running the nv30 by any stretch of the imagination. I mean wouldn't you have to input the chip configuration into the FPGA card in someway? So the FPGA card is the method of choice of IHVs for the emulation of hardware that is not existent yet via the input of software that allows the FPGA card to emulate the algorithms of a particular hardware design?
Galilee said:
I understand that they are using the coding language HDL to input the chips configuartion into the FPGA card. But where is the nv30 in .13um hardware? At any rate to say the nv30 is running in hardware is deceptive to say the least. On that note I think I do understand what the matter here is. They could have been running the nv30 in FPGA cards for months now as it is an intrical part of the design process but that does not make one conclude that indeed that the nv30 final hardware is "just around the corner" if anything it says the nv30 is still in the design process. But as Prometheus stated this vist that Anand made was a while back so who knows. Interesting article BTW, too bad there were not any pics of ATIs complex and stuff.
Er, anyone who's done Electrical Engineering or anything like it will know that "emulating" a piece of hardware via an FPGA is completely different to emulating it in software. When you're running that hardware, you're RUNNING THAT CHIP, but in a piece of configurable hardware. At that moment, the hardware is wired so it's identical to the chip that you're going to make, so it's not really emulating it. It is for all intents and purposes the chip you plan on fabbing, but at a much lower clock speed. It is hardware, not software.
I can't comment on your particular case, but it was possible to take a Xilinx design and produce a hardwired version that was much cheaper to produce than the 'parent' FPGA. In terms of a per-unit cost it wasn't as cheap as a standard ASIC but the set-up costs were much lower.pcchen said:I think there is a FPGA chip (did Xilinx produces other type of chips?) on my soundcard
The "few kHz" surprised me. They must be deliberately using exactly the same logic as is intended for the final chip. I would have thought it'd be easy enough to make it run at least at several MHz by inserting more register stages.Title said:
That's not really true. Apart from the convenience of interfacing drivers etc., you should be getting exactly the same logical results whether you use a fully "software-based" simulator, an FPGA simulator, or final hardware. (Of course you can't allow for process related issues.).mech said:
It's never interested me much, but arent these things just netlist simulators? It's not like they try to include timing I imagine ... or that they can include things such as signalling and noise levels AFAICS.
When a FPGA is doing things you cannot simulate in software it is most likely not working like you want it to (Except for the few who are trying to use them for analog signal processing through GA optimized programming.)
Marco
PS. well what do you know, I see IKOS's NSIM system (uses dedicated hardware, but not FPGA based though) does include timing verification ... but still you could do that in software too.
When a FPGA is doing things you cannot simulate in software it is most likely not working like you want it to
(Except for the few who are trying to use them for analog signal processing through GA optimized programming.)Marco
PS. well what do you know, I see IKOS's NSIM system (uses dedicated hardware, but not FPGA based though) does include timing verification ... but still you could do that in software too.
Simon F said:
I was suprised also, (I think). Most modern micro processor/controller cores can run 20-30mhz in an FPGA. On the other hand, it could be that the IKOS, which is essentially an array of FPGAs, has much greater delays in them due to interconnecting the FPGAs, which make timing that much worse.
But a few kHz just sounds pokey.
For the record, if it was a purely a software simulation of the gates, it most likely it would be running less than 1Hz, or several seconds per clock. (Of course, depending on the processor thrown at it).
Anybody else suprised that NVIDIA had their own FIB? Or the sheer amount of servers? Egad. Makes the company I work for look like DIY hobbyists in comparison.
And yes, Marco, they're just functional emulators. They won't ferret out any process related issues (like crosstalk, etc). But they're dandy for making sure you've designed the right thing, and probably for writing drivers, etc.
