Author Topic: has anyone any experience of using and interfacing with FGPA /VHDL  (Read 975 times)

mikeburr

  • Member
  • **
  • Posts: 96
im interested in developing some low level maths using [specifically the Xilinx chips 7000 or 7100 range if i can afford it ! PCIE ] and would like to know if any body has interfaced with this chipset ... saw the CUDA posts as they offer some thing slightly different they are interesting too particularly the deep learning aspect   regards  mike b

K_F

  • Member
  • *****
  • Posts: 1609
  • Anybody out there?
I've gone into the whole 'research' on whether to use FPGA or just buy a big fat gpu card.

It's definitely worth it if you can afford the FPGA route... I drool over the Stratix 10... but don't have 8000 euros (Xilinx is also $$$) for the plugin-in card.
FPGA's are nice in that you can cook your own algorithms with accompanied 'power' throughput, and you have 'absolute control' over the hardware.

GPU cards offer the best bang for buck, but the Deep Learning Cards (multiple GPUs) are stupidly expensive as well, even AMDs lot.
The prices are elevated by the BitCoin (and other digital currency) charade, and maybe with these currencies 'collapsing' there might be a glut of GPU cards at low prices.

Added to this Nvidia and AMD go out their way, big time, to prevent you from programming GPU cards in Asm - They just don't support this.
GPU Asm has been achieved with limited success, but the Asm structure changes with each new GPU, so PTX (Nvidia Psuedo Assembler) is the way to go here.

Also I'm not really in favour of AI libraries you get with 'Deep Learning'. One has to rely on someone else's idea of what is best... nope, nada..zip ;)

I haven't looked a lot at AMDs SDK as yet, but it looks like they follow a similar idea to NVidia.




'Sire, Sire!... the peasants are Revolting !!!'
'Yes, they are.. aren't they....'

mikeburr

  • Member
  • **
  • Posts: 96
many thanks K_F ... got a very similar impression
"Added to this Nvidia and AMD go out their way, big time, to prevent you from programming GPU cards in Asm - They just don't support this.
GPU Asm has been achieved with limited success, but the Asm structure changes with each new GPU, so PTX (Nvidia Psuedo Assembler) is the way to go here."  ...
you only have to read LiaoMi and Biteriders interchange to see how awkward it is .. having written that microsoft dont exactly expose their assembler and if it wasnt for the dedication of the people who run and contribute so much to this web site it would be extremely hard to achieve anything
...
i have a very good Chinese friend and asked him to look up some prices of the mid range zynq 7045 and 7100 cards on PCIe .. by passing him some searches on ebay here in the UK .. it turns out that the equivalent  7100 card is about 2/3 price there £500 compared to £850 here and about $1000 in the states ... eeeeeeeeekkkkkk   .. other cards with less power are similarly cheaper the 7045 looks a decent bet at a more reasonable price ... if i can get hold of one ill put details on here of what happens as its bound to be traumatic ..
 incidentally i noted with interest that music companies are now using these processors to generate waveforms and otherwise utilising the low latency and concurrency they provide ... im personally investigating some aspects of number theory with what i have here ...because of the limitations of bit length [32 bit  and 64] i decided to build a software "core"  which kind of emulates some aspects of the chip sort of on the fly .. this is hampered a bit in Windows as it doesnt like memory manipulation with its version of mutli threading [ which i suspect was an after thought in the design and having worked briefly with Fred Brooks who was culpable for one of IBM's i can sort of see why if he represented the culture though i think im being very unfair (he is sadly recently deceased) ]   
 

K_F

  • Member
  • *****
  • Posts: 1609
  • Anybody out there?
I'd be inclined to ask your Chinese 'friend' on the exact chips (their ID designation) used in their production models.

The ID designation usually determines operating temperature range as well as their speed ranges.
In Other Words (IOW), the reliability of the system/equipment to perform to 'spec' set by the original manufacturer.
Invariably you'll find the cheaper products 'sub-spec'.

This is not to say the Chinese produce sub-spec... they produce good stuff.
The difference is in the price, and you must be aware of this.
Their 'in-spec' stuff is just as pricey as the west. ;)
'Sire, Sire!... the peasants are Revolting !!!'
'Yes, they are.. aren't they....'

K_F

  • Member
  • *****
  • Posts: 1609
  • Anybody out there?
I'm also looking at the cheaper DSP modules, and their scalability.

There are smaller FPGA (cheaper) modules available that you can build/scale up as you can afford.
I'm also looking at transferring Masm to a core FPGA system, with a simple link interface from a PC.

My impression is that MS has lost the plot, and fallen out with Intel, so in time MS will be dumped.
'Sire, Sire!... the peasants are Revolting !!!'
'Yes, they are.. aren't they....'

mikeburr

  • Member
  • **
  • Posts: 96
thanks for that info .."Invariably you'll find the cheaper products 'sub-spec'." in my naivety i assumed everything with the same number on it would be the same ... i agree with you regarding MS in fact most of the scientific stuff seems to be done on linux now  as the posix threading system has always been designed into the core of it .. in assembler terms that might mean using QT and Fasm [ boo hiss ] as NASM is a bit clunky which is why i tried to build effectively  Turing style
 processing on my 32 bit machine [ havent finished yet but the bits ive done so far have proved to be excellent and much more flexible than i thought ] i was hoping to transfer this to FPGA and get concurrency that way as out lined ] .. what is the expandable system youve looked at  as that sounds to have potential and are increase i scale lost by complexity of communication significantly ??? regards mikeb