<freemint> mhh, all benchmarks i can find run on really old hardware
jemk has quit [Ping timeout: 250 seconds]
<freemint> sorry for the naive questions
<kc8apf> GPU is an ASIC designed primarily for floating point.
<freemint> I am aware i was thinking the fact that you complettly preventt pipeline stalls,
<freemint> branches etc would do favor FPGAs over GPUs
<kc8apf> With all the GP-GPU usage, they even have models that follow IEEE754
jemk has joined ##openfpga
jemk is now known as Guest5574
<kc8apf> An ASIC always wins over an FPGA as long as the problem is relevant long enough to offset NRE
<freemint> This seems like a good heuristic
<freemint> Is my impression wrong that high end FPGAs are often one node shrink ahead of the high end GPUs?
<kc8apf> I guess it's really a balance of longevity and volume. Cell phone CPUs are ASICs because the volume is there even with a 1y released cadence
<freemint> I thougt FPGAs could go to new nodes earlier since they are very immune to local defects knocking a certain LUT out.
<kc8apf> Even if FPGAs are at the same node as a GPU, an FPGA fabric is huge compared to an optimized design
<freemint> I am aware that at the same noda an FPGA is solwer unless only the asic parts (DSP blocks( are used
<kc8apf> Slower, greater power consumption, and smaller implemented designs
<freemint> thanks you helped dampen my wild ideas about FPGA accelerators for scientific computing.
<kc8apf> Depends. If the algorithms are still rapidly evolving, building an ASIC would be silly
<kc8apf> Or if there are <10k users of such an ASIC
rohitksingh has quit [Ping timeout: 265 seconds]
<freemint> that number seems what estimated for ICE40 sized chips. Does it hold for bigger chips tooP
<freemint> users or instaces of an asic?
<sorear> GPUs target nodes and voltage/frequency operating points that optimize energy per bit operation, because they are latency insensitive
<freemint> i dound an really up to date paper about this here http://www.bertendsp.com/pdf/whitepaper/BWP001_GPU_vs_FPGA_Performance_Comparison_v1.0.pdf
<freemint> *sort of uptodate
<freemint> Artix7 200T is as fast in FPU as a GT730 (numbers from 2016). They are 3 times as expensive per clock but 10 more enery efficient at flops.
<kc8apf> When to use an ASIC comes down to costs. A set of masks is $1+M. Engineering time can easily be more.
<freemint> I thought you could save a lot by going to older nodes?
<kc8apf> Per-die incremental costs are <$1
<kc8apf> Sure. But then you're comparing what you can do with an FPGA on the latest node versus an ASIC on a much older node
<freemint> ASICS for >11000$ you never stop learning
<sorear> 11k is damn cheap
<freemint> what is the highest you have seen?
<freemint> 11k is probably more than all my networth as a student
somlo has quit [Remote host closed the connection]
<sorear> I've seen slides quoting >100M for tapeouts at leading nodes, but I think that rolls in a fair amount of design work, not just the fab itself
<kc8apf> NRE on things like Google's TPUs is easily $100+M
<kc8apf> Yup
<freemint> Oh yes, sorry i misstyped, FPGAs for >11000$
<kc8apf> Tools and design work is more than fab setup
<sorear> xilinx and probably all other vendors fleece qty 1 buyers
<freemint> I was supprised to see an FPGA which costs more than a decent car. Do you know how high the FPGA prices go?
<kc8apf> Prices on top-end FPGAs are highly variable and never list price
<freemint> ahh
<sorear> if you wanted to build a f1.16xlarge from components at list price, you would spend about $200k
<kc8apf> No one buys qty1 of those via a distributor
<sorear> (nobody pays list price)
<sorear> wait, I think it was 8x$45k
<freemint> Are you aware of any cheap rented access to <= Spartan class FPGAs?
<freemint> Those Amazon FPGA instances are a bit to high class for anything i have in mind. Having FPGA renting like there is for GPUs would be great.
<freemint> Sorry i am tired as hell GN
<GenTooMan> good night rest well
<GenTooMan> I can understand NRE's and that's why those very expensive high end FPGA's are so important before you blow millions on a defective design. :D
<freemint> I was just thinking about calling deutsche Telekom because they have 24/7 hotline for testing and they also offer FPGAs in their cloud ... when it is 3 am there ... .
<freemint> what NRE?
<GenTooMan> Non Reoccurring Engineering Fee for a design of an SoC IC those get expensive fast. You can chip the cost way down but eventually you have to pay I think I saw a few years back someone did a SoC multi-core Adaptiva was the companies name.
<sorear> naive question, but what are you hoping to gain by having rented access to a spartan instead a virtex ultrascale
<sorear> are you trying to use tools that only support spartan?
<sorear> is your budget such that $1/hour for the f1.2xlarge is a lot of money?
<freemint> Let's say some OSHW builds a new softcore and they want to unit test their design with each change they do.
<sorear> unit tests are normally short and run in sim
<freemint> No but f1.2x large is not FPGA which is targetted
<sorear> why are you writing softcores that target specific fpgas
<freemint> Maybe i said that badly. An f1.2x seems a bit overkill
<sorear> in what way are you *hurt* by having extra capacity?
<sorear> 2ghz cpus are overkill for nearly everything, do you go out of your way to get 50mhz software testing machines?
<freemint> Maybe this was just a stupid idea at night i have.
<freemint> If someone has a semi-decent server somewhere and would give access to 4
<freemint> Artix 7 to $person. would that be a good thing or would it be worthless to ##openfpga
<kc8apf> I've asked about this in the past. Everyone wants some sort of CI system for FPGAs. I dunno what that ends up looking like.
<kc8apf> Under what circumstances would a sim be insufficient but additional peripherals aren't required?
<kc8apf> I've got a wide spectrum of FPGAs I could put online. I just have no idea in what configuration they would be useful
<freemint> When you want to test that a Linux boots on your softcore and it can missuse the MMU as it want's without anything breaking? I think sim is >10.000x slower than FPGA
<kc8apf> But that requires some specific, additional peripherals
<freemint> How?
<kc8apf> RAM
<freemint> Right ... maybe having RAM next to the FPGA is worth having
<kc8apf> But it wouldn't be on an iCE40
<freemint> Depends on what board you put there for testing but right i see the problem
<freemint> You could attach dev-boards but ... you find some flaw in that too i guess
<kc8apf> Best I've come up with is having a 2nd FPGA emulating devices
<sorear> f1 *does* give you ram
<kc8apf> Similar to how Glasgow does applets
<kc8apf> Is F1 useful as a CI system?
<freemint> I have no idea but i guess it could be. Does anyone know how fat the FPGAs are in these machines?
azonenberg_work has joined ##openfpga
<kc8apf> They are UltraScale+ with 64GB of RAM attached and x16 PCIe to host
<kc8apf> xcvu9p-flgb2104-2-i
rohitksingh has joined ##openfpga
azonenberg_work has quit [Ping timeout: 250 seconds]
freemint has quit [Ping timeout: 250 seconds]
rohitksingh has quit [Ping timeout: 245 seconds]
rohitksingh has joined ##openfpga
Jybz has joined ##openfpga
Bike has quit [Quit: Lost terminal]
somlo has joined ##openfpga
rohitksingh has quit [Ping timeout: 250 seconds]
rohitksingh has joined ##openfpga
Kitlith has quit [Quit: ZNC 1.6.5 - http://znc.in]
rohitksingh has quit [Ping timeout: 250 seconds]
OmniMancer has joined ##openfpga
Miyu has joined ##openfpga
hackkitten has quit [Ping timeout: 276 seconds]
rombik_su has joined ##openfpga
rombik_su has quit [Quit: Leaving]
Jybz has quit [Ping timeout: 240 seconds]
emeb_mac has quit [Quit: Leaving.]
rohitksingh has joined ##openfpga
Miyu is now known as hackkitten
m4ssi has joined ##openfpga
m4ssi has quit [Quit: Leaving]
rohitksingh has quit [Ping timeout: 245 seconds]
freemint has joined ##openfpga
Zorix has quit [Ping timeout: 264 seconds]
Zorix has joined ##openfpga
<emily> big
monochroma has left ##openfpga [##openfpga]
genii has joined ##openfpga
mkru has joined ##openfpga
carl0s has joined ##openfpga
Guest5574 is now known as jemk
mkru has quit [Remote host closed the connection]
emeb has joined ##openfpga
OmniMancer has quit [Quit: Leaving.]
<kc8apf> $5 GOWIN-based dev board.
<kc8apf> I kinda want to find someone in China to buy 10 on Taobao and mail them to yosys/nexpnr devs
<daveshah> pepijndevos is on it already
<ZirconiumX> Project Apicula
<ZirconiumX> I think?
<carl0s> Are there plans to support gowin devices with open source tools? or there is work already done?
<mwk> it's work in progress
<carl0s> Nice, I think we will get a board at work, but it's a GW1N-9 device.
mumptai has joined ##openfpga
<pepijndevos> ZirconiumX, daveshah yessss. Making good progress too. After staring at binary files for days, I have most of the data to start coding
<ZirconiumX> Yay!
Asu has joined ##openfpga
<carl0s> mwk, is it public yet? I googled for project Apicula and only found a non related one
rohitksingh has joined ##openfpga
rohitksingh has quit [Ping timeout: 265 seconds]
<ZirconiumX> carl0s: mwk is the wrong person to ask, try pepijndevos :P
<pepijndevos> carl0s, nothing public yet
<pepijndevos> If it were, you could not do much with it anyway... just a bunch of fuzzing scripts at the moment
<pepijndevos> I tweet some progress updates some times
<carl0s> ZirconiumX, my bad :P
<carl0s> pepijndevos, same nick on twitter?
<pepijndevos> yep
<carl0s> Cool!
<kc8apf> pepijndevos: even fuzzing scripts and notes are useful
<pepijndevos> kc8apf, for what? Well, I'm just a bit cautious about what to share bc copyright and stuff.
<kc8apf> Getting an understanding of the architecture, naming conventions, quirks
<kc8apf> I can share past discussions this channel and prjxray have had on copyright issues and general RE-tainting concerns
<kc8apf> azonenberg, mithro, etc have visited that from many viewpoints
rombik_su has joined ##openfpga
<pepijndevos> kc8apf, would be interested to learn more about that
<kc8apf> I'm getting on a flight right now. I'll get in touch soon
carl0s has quit [Remote host closed the connection]
mumptai has quit [Quit: Verlassend]
Bird|ghosted has joined ##openfpga
emeb_mac has joined ##openfpga
Bird|otherbox has quit [Ping timeout: 240 seconds]
azonenberg_work has joined ##openfpga
Bird|ghosted has quit [Remote host closed the connection]
Bird|ghosted has joined ##openfpga
rohitksingh has joined ##openfpga
rombik_su has quit [Quit: Leaving]
mumptai has joined ##openfpga
Asu has quit [Ping timeout: 240 seconds]
Asu has joined ##openfpga
genii has quit [Quit: Welcome home, Mitch]
Asu has quit [Remote host closed the connection]
lexano has quit [Ping timeout: 276 seconds]
rohitksingh has quit [Ping timeout: 250 seconds]
rohitksingh has joined ##openfpga
Bike has joined ##openfpga
<TD-Linux> love the fomu led :)
<TD-Linux> oh now slightly disappointed it's rgb and not native aqua
emeb has quit [Quit: Leaving.]