Zedboard, dual ARM cores with an FPGA attached(zedboard.org)
zedboard.org
Zedboard, dual ARM cores with an FPGA attached
http://zedboard.org/
2 comments
I've heard this so many times that I've grown tired of it. The "$15 in quantity" claim from Xilinx and other FPGA manufacturers. Yes, it is quite likely that after 2 years your order of 100,000 units (first batch in a larger 500k schedule, of course) will eventually reach that mythical $15.
I've learned to design things using my local distributor as a pricing source. I use Farnell and look at their qty 10 pricing. That, together with the fact that a part is available at all, is a good indicator: I know I can get the pricing down at least 40% if I achieve quantity. And I have my feet firmly planted on earth.
So, that Zynq chip? Farnell doesn't stock it at all, and findchips shows one source, beginning at $105 per chip. Ouch.
I realize hackers enjoy tinkering with FPGAs, but for most practical applications in this power/performance range you'd be much better off with a Freescale i.MX6, those begin at $15 (quantity 10) and $42 will buy you a quad-core. Those chips have loads of peripherals, pretty much anything you can think of, and pretty impressive graphics accelerators. Which means you end up using pre-designed, tested and documented subsystems instead of developing your own logic.
I've learned to design things using my local distributor as a pricing source. I use Farnell and look at their qty 10 pricing. That, together with the fact that a part is available at all, is a good indicator: I know I can get the pricing down at least 40% if I achieve quantity. And I have my feet firmly planted on earth.
So, that Zynq chip? Farnell doesn't stock it at all, and findchips shows one source, beginning at $105 per chip. Ouch.
I realize hackers enjoy tinkering with FPGAs, but for most practical applications in this power/performance range you'd be much better off with a Freescale i.MX6, those begin at $15 (quantity 10) and $42 will buy you a quad-core. Those chips have loads of peripherals, pretty much anything you can think of, and pretty impressive graphics accelerators. Which means you end up using pre-designed, tested and documented subsystems instead of developing your own logic.
Interesting.
Zynq-7020 is supposedly less than $50 in reasonable quantity, if you stay firm with the distributors. Xilinx are known for unpublished discounts, using the discount code Altera :)
BTW Couldn't you make a cheap FPGA into a frame buffer, not use a separate ARM chip?
Zynq-7020 is supposedly less than $50 in reasonable quantity, if you stay firm with the distributors. Xilinx are known for unpublished discounts, using the discount code Altera :)
BTW Couldn't you make a cheap FPGA into a frame buffer, not use a separate ARM chip?
Couldn't you make a cheap FPGA into a frame buffer, not use a separate ARM chip?
Yes you can, integration isn't as nice though. It's pretty reasonable to build a DDR3 memory controller into a Spartan3 or Cyclone4 and the ARM PMI bus. My sketch has the memory slaved to a DMA control dumping RGB888 into a Silicon Labs or TI HDMI 'phy' at the requisite rate to generate a 1080p/60 display. The thought being that the RAM would be 'display preferred' and processor access would be 'when available'. Not as nice as a direct memory interconnect but easy to add to pretty much anyone's standalone ARM chip. I was shooting for something like the Sun3 color framebuffer on the Sun 3/60 which was pretty easy to talk to an pretty much hard to imply I'm violating Patents since those documents are from the 80's.
I couldn't get it much below $15 though, it is a 3 chip solution (RAM / FPGA / Phy). The Zynq absorbs both the RAM and the FPGA so I'm left with the physical interface. And in the Zynq case I actually get two CPU cores so one could conceivably do a full 3D software implementation in one CPU and have the other one run Linux and talk to it via a simple in memory mailbox protocol.
The Zynq is also already soldered together and trying to get the Digilent plugon modules talking to the Spartan 3 dev board i had was looking dicey at best. :-) I can fantasize about Atmel doing a hard core of a simple Frame buffer but its not realistically in the 'likely' category.
Yes you can, integration isn't as nice though. It's pretty reasonable to build a DDR3 memory controller into a Spartan3 or Cyclone4 and the ARM PMI bus. My sketch has the memory slaved to a DMA control dumping RGB888 into a Silicon Labs or TI HDMI 'phy' at the requisite rate to generate a 1080p/60 display. The thought being that the RAM would be 'display preferred' and processor access would be 'when available'. Not as nice as a direct memory interconnect but easy to add to pretty much anyone's standalone ARM chip. I was shooting for something like the Sun3 color framebuffer on the Sun 3/60 which was pretty easy to talk to an pretty much hard to imply I'm violating Patents since those documents are from the 80's.
I couldn't get it much below $15 though, it is a 3 chip solution (RAM / FPGA / Phy). The Zynq absorbs both the RAM and the FPGA so I'm left with the physical interface. And in the Zynq case I actually get two CPU cores so one could conceivably do a full 3D software implementation in one CPU and have the other one run Linux and talk to it via a simple in memory mailbox protocol.
The Zynq is also already soldered together and trying to get the Digilent plugon modules talking to the Spartan 3 dev board i had was looking dicey at best. :-) I can fantasize about Atmel doing a hard core of a simple Frame buffer but its not realistically in the 'likely' category.
Chuck, have you verified that the price is $15? because i heard it's mostly marketing and the real price is much higher.
Well it was a quote I got from the distributor. At a conference the guy talking about the chip gave me that price and I asked the distributor to verify it. They squirmed a bit but eventually said they would (also Arrow btw).
That said, there has apparently been a lot more interest in this chip than Xilinx expected and I got the impression they were worried they had under priced it. But some of the markets they are shooting for won't tolerate a higher price so my hope is that it will hold. At 10K units its only a $150K buy to get the 'good' price, one wonders if you could get enough interest in a kickstarter to meet that quantity goal.
That said, there has apparently been a lot more interest in this chip than Xilinx expected and I got the impression they were worried they had under priced it. But some of the markets they are shooting for won't tolerate a higher price so my hope is that it will hold. At 10K units its only a $150K buy to get the 'good' price, one wonders if you could get enough interest in a kickstarter to meet that quantity goal.
I think the real challenge for a kickstarter would be to offer attractive fpga development route. Currently xilinx prices their fpga-c dev tool at around $5K. My guess is that's the real limit , not the price of the zedboard.
Xilinx's official dev boards are all pretty expensive, but they tend to be all-singing all-dancing affairs, with largest-in-class FPGAs, tons of peripherals and interfaces, and enough little blinky lights to blind a goat. Third-party dev boards (like the ones from Digilent: http://digilentinc.com/) tend to be much more reasonable.
I'd be interested in knowing how fast data can be transferred between the fpga fabric and the processor. Any idea?
I believe the 2.4GB/sec is the spec rate for this chip. Their demo board does three 1080p feeds mixed into a 1080p output which suggests it is pretty usable.
One of the key differences in this chip versus other architectures is that ARM system boots first and has some ability to do reconfiguration. (all of the key subsystems for the ARM core to boot are hard blocks.)
I don't hold out a lot of hope for open floorplanning tools from Xilinx but they do have a Linux toolchain so hopefully it will be possible to do native development.
One of the key differences in this chip versus other architectures is that ARM system boots first and has some ability to do reconfiguration. (all of the key subsystems for the ARM core to boot are hard blocks.)
I don't hold out a lot of hope for open floorplanning tools from Xilinx but they do have a Linux toolchain so hopefully it will be possible to do native development.
You should take a closer look at Xilinx's tech docs (there are a lot): http://www.xilinx.com/products/silicon-devices/soc/zynq-7000... - it uses ARM AMBA for interconnect.
FWIW, Parallella uses a very similar chip (ZYNQ7010 vs 7020) and lists the interconnect bandwidth as 2.4GB/s so I'd expect it'd be around there.
FWIW, Parallella uses a very similar chip (ZYNQ7010 vs 7020) and lists the interconnect bandwidth as 2.4GB/s so I'd expect it'd be around there.
Xilinx' block diagram shows the FPGA having direct access to both the peripheral bus and the memory bus, including the cache coherency port, so I'd guess you can transfer data to the ARM cores fairly fast (as fast as you can get something from dram into L2 cache).
Prior to finding out about this chip I was sort of half-heartedly working on a simple 1080p framebuffer to attach over the ARM PMI bus to an Atmel SAM9 part but this is much nicer being fully integrated. My plan now is to build a system out of it and then spin a board without the extra bits to get the board price down around $99 rather than $400. Very interesting architecture.