| wolfspraul | but I rather focus on m1 today and fix many many things on that, and then think about the next step. which is bigger/more powerful fpga, altera or -7 series, zynq (?) | 00:00 |
|---|---|---|
| wolfspraul | we need customer demand, better marketing, easier product. otherwise it's just random lab tests, and we are better off with everybody using different devel boards :-) | 00:00 |
| wolfspraul | I think Sebastien extracted an end user classic out on the mailing list - did you see it? | 00:00 |
| wolfspraul | the poor guy who cannot update his m1. "I did it about 10 times. Not sure about reproducible" | 00:01 |
| wolfspraul | do we care about those things? | 00:01 |
| wolfspraul | or we just want to play with some cool features here and there? | 00:01 |
| wolfspraul | there is no economic need or demand to make an lm32 asic, other than us learning how to tape-out and work with tsmc or whoever | 00:02 |
| wolfspraul | there is hardly an economic need to upgrade the fpga, but I hope we can create that momentum... | 00:02 |
| wpwrak | the problem with the lm32 core is that it's incredibly slow | 00:02 |
| wpwrak | in m1, we get away with this because we do very little in the core | 00:02 |
| wolfspraul | in the long run - of course most likely we have to "make chips", that is tape-out stuff. but where and when and which product? too early for me | 00:02 |
| wolfspraul | I fully understand [slow] | 00:03 |
| wolfspraul | we are in this situation because that's what we wanted, no? | 00:03 |
| wpwrak | in a more software-centric device like a nanonote, the slowness would be considerably more noticeable | 00:03 |
| wolfspraul | we look down on such lame companies as ARM and Intel... | 00:03 |
| wolfspraul | :-) | 00:03 |
| wpwrak | it's what fate wants for us. for now :) | 00:03 |
| wolfspraul | I feel good about the direction btw, very good. and of course I like Sebastien's lm32 asic idea. but I try to put it in a realidad perspective. | 00:04 |
| wolfspraul | the best result would be that we learn the process, tape-out, work with a foundry, etc. like an applied school. | 00:04 |
| wolfspraul | other than that there is no need, no demonstrable user demand for the result | 00:04 |
| wolfspraul | as you just said, I dismissed people saying "combine fpga with asic cpu" for years, well. maybe they are not so crazy after all, but it's us? :-) | 00:05 |
| wpwrak | i think for M1 the FPGA-only approach is still tolerable | 00:06 |
| wolfspraul | if we make an asic cpu, the reason is? faster lm32? how much faster? it can then compete with what? who cares, who competes? | 00:06 |
| wpwrak | even here the slowness makes itself felt sometimes | 00:06 |
| wolfspraul | the approach with m1 is to sell it as a product where people don't compare mhz | 00:06 |
| wolfspraul | and I do believe in that approach | 00:06 |
| wolfspraul | yes I know | 00:06 |
| wpwrak | the issue is whether the M1 can keep up with its application | 00:06 |
| wolfspraul | let's also hope that migen brings some speedup for the memory at least | 00:06 |
| wpwrak | now, M1 has basically a single application. this makes the task easier. | 00:07 |
| wpwrak | if NN-Ya-FPGA-only is to be marketed as a "universal" machine, like the Ben, then resource pressure from applications will be much larger | 00:07 |
| wolfspraul | my point is: the 'make asic' idea and direction is right, and is one reason why the zynq chip would be a huge distraction. but I don't think we will have such asics running on our desks this year, next year. | 00:08 |
| wolfspraul | just saying | 00:08 |
| wolfspraul | :-) | 00:08 |
| wolfspraul | proove me wrong | 00:08 |
| wpwrak | didn't you want to venture into making ASICs ? :) | 00:08 |
| wolfspraul | absolutely, I just said I think most computing power comes from the chips anyway | 00:08 |
| wolfspraul | but the discussion earlier was zynq or nozynq | 00:09 |
| wolfspraul | that's a practical question, something we can most likely source in small quantities in early 2013 | 00:09 |
| wolfspraul | for a few USD | 00:09 |
| wpwrak | what i like about sebastien's idea is that it would a) considerably reduce the complexity of the task, and b) preserve de facto all of the flexibility | 00:09 |
| wolfspraul | and solder together in our fabulous open process | 00:09 |
| wolfspraul | yeah, me too | 00:09 |
| wolfspraul | only that it will not happen in reality :-) | 00:09 |
| wpwrak | pity | 00:09 |
| wolfspraul | too early, no demand, no marketing pressure or ideas behind it | 00:10 |
| wolfspraul | just engineers wanting to try something new | 00:10 |
| wolfspraul | a lab experiment | 00:10 |
| wpwrak | zynq seems to lead us back into more proprietary gateware. i do hope we'll have something like llhdl in the end. for me that's the most interesting result of all this. | 00:11 |
| wolfspraul | other than that *of course* milkymist has to be taped-out eventually to be sucessful | 00:11 |
| wolfspraul | what else? the end result of an IC design is to run in an fpga? no way | 00:11 |
| wolfspraul | yes agreed | 00:11 |
| wpwrak | why not ? | 00:11 |
| wolfspraul | and it gets us closer to asic | 00:11 |
| wolfspraul | the "end result" | 00:11 |
| wpwrak | maybe we could even synthesize part of a patch into the FPGA | 00:11 |
| wolfspraul | yes, of course | 00:12 |
| wpwrak | i view the FPGA as a living thing. not just a development platform | 00:12 |
| wolfspraul | but the whole point to work on a free IC design is to eventually make hardwired chips out of that design | 00:12 |
| wolfspraul | must be, I think | 00:12 |
| wpwrak | if you can make use of the extra flexibility (which causes the extra cost) of the fpga, then this doesn't have to be so | 00:12 |
| wolfspraul | yes agree, the fpga reprogrammability is something we haven't even started using yet | 00:13 |
| wpwrak | it's like buying microcontrollers with mask ROM or with flash | 00:13 |
| wolfspraul | I agree with you, we just talk about a wide subject, different perspectives | 00:13 |
| wolfspraul | I am not against the fpga reprogrammability, still writing my thoughts on "lm32 asic idea" | 00:13 |
| wolfspraul | wpwrak: come on Werner. isn't that "I did it 10 times, not sure about reproducible" a classic? | 00:14 |
| wolfspraul | I love it! | 00:14 |
| wolfspraul | it reminds me who we really work for, or I at least | 00:14 |
| wolfspraul | users are so cool, so free | 00:14 |
| wpwrak | ;-) | 00:14 |
| wolfspraul | you saw it? | 00:14 |
| wolfspraul | fantastic | 00:14 |
| wolfspraul | that guy is so right | 00:14 |
| wolfspraul | I do care about that, want to make this stuff work for regular people. if it's just a devel board, there are many other/better devel boards that constantly come out with a stream of 'interesting' features. | 00:15 |
| wolfspraul | and, as lars knows, Xilinx is flooding the world of smart engineers with Zynq devel boards right now | 00:15 |
| wolfspraul | to come to a hacker desk near you... :-) | 00:15 |
| wolfspraul | I wanted to ask a Xilinx rep about artix-7 pricing and ISE webpack support the other day. the reply: "Don't you want to buy a Zynq devel board? we have them now!" | 00:16 |
| wolfspraul | huh? :-) | 00:16 |
| wolfspraul | misunderstood me? :-) | 00:16 |
| wpwrak | so far, we have a fairly spotless record when it comes to fatal flaws. the hard school of openmoko seems to be paying off ;-) | 00:16 |
| wpwrak | they probably get a bonus if they sell you a zynq :) | 00:17 |
| wolfspraul | oh they are pushing it like crazy | 00:17 |
| Action: roh still wants a system-mpu on the next board. and if its only sequencing the power rails ;) | 00:17 | |
| wolfspraul | monthly meeting "how many zynq boards did you sell?" etc. etc. | 00:17 |
| wolfspraul | roh: which chip do you have in mind? | 00:18 |
| wolfspraul | how about multiple lm32 cores btw? | 00:19 |
| wolfspraul | in the fpga | 00:19 |
| roh | wolfspraul: it needs to be seperate of all other cores. something tiny, depending on the number of tasks. in case of only reset line and power rail sequencing even a attiny could to the job | 00:20 |
| wolfspraul | aside from optimizing the timing/speed of the core in the fpga (if that is possible), that looks like another way to get more CPU power out of there, no? don't know how hard it is though | 00:20 |
| wolfspraul | what's the advantage of that? | 00:20 |
| roh | i currently do not believe in any sense of lm32 and linux. the reachable speeds are too small by far | 00:20 |
| roh | when it comes to building something mobile i think there is currently no reasonable way not involving readymade cpu cores or soc. | 00:21 |
| wolfspraul | not sure, depends on what you are building and how it is marketed | 00:22 |
| roh | the advantage of having a system-mpu is that you can fix bugs later in firmware. | 00:22 |
| wolfspraul | there are many ultra-low power fpgas, for example | 00:22 |
| roh | remember the nxp chip on the moko.. or the milkymist reset bug (fixed in hw by changing the reset chip and its powerrail) | 00:23 |
| wolfspraul | if you have the "iphone competitor" in mind, then your options are much less, best would be to just say 0 | 00:23 |
| roh | wolfspraul: i am thinking mostly about an successor to the nanonote we have. and i dont see it having less computing power | 00:23 |
| wolfspraul | ok but adding such chips is again a huge distraction to our tiny community, it can easily stall and fragment the 3 folks we have hacking here and there | 00:23 |
| roh | but that would be the result of using a purely fpga design from my pov | 00:23 |
| wolfspraul | the natural reaction is "let's put this and that chip on it, then it's a great product" | 00:24 |
| wolfspraul | and that ignores entirely the dynamics of hardware and time it takes to stabilize a design etc. | 00:24 |
| wolfspraul | by the time all the bugs are fixed (if they ever are, most projects seem to just give up), the design is outdated and the next set of chips is glorified | 00:24 |
| roh | sure. i know the game now quite a time. thats why i basically dont believe a word most vendors say in any of their marketing | 00:25 |
| wpwrak | roh: you're basically proposing what i suggested for gta03 :) alas, then the other folks turned that little "system manager" into a monster. i think they had something with 40+ pins in the end :-( | 00:25 |
| wolfspraul | to speed up the lm32 we have now, how about the following options: | 00:25 |
| wolfspraul | 1. faster clock, tigher timing constraints in the fpga? possible? | 00:26 |
| wolfspraul | 2. more cores? | 00:26 |
| roh | any speeding up doesnt make it more power efficient in an fpga. and if, only by a small percentage | 00:26 |
| wolfspraul | 3. any improvements from mmu, better caching, faster memory, other architectural improvements? | 00:26 |
| wolfspraul | 4. better optimizing compiler? gcc or llvm? | 00:26 |
| wpwrak | 1. i asked sebastien about that, encouraged by the high clock rates he got for the (non-trivial) memory controller. he sounded as if he didn't think there's much room for improvement. | 00:27 |
| wolfspraul | maybe those angles are all exhausted/impossible/too-hard/others? don't know | 00:27 |
| roh | no multiplication of the performance in sight which makes it 2-4 times faster (get to the +700mhz arm9/11 range) | 00:27 |
| wolfspraul | oh I'm sure we cannot always just point to Sebastien and wait | 00:27 |
| wpwrak | 2. that could be an option. but of course, limited. we already have multiple cores, with a pretty fast AVR dedicated to USB :) | 00:27 |
| wolfspraul | someone has to sit down and try independently, and then I'm sure Sebastien will be interested to discuss results, if there are any interesting ones to discuss | 00:27 |
| wpwrak | 3. faster memory is coming. but that probably won't help the core's slowness much, if at all. | 00:27 |
| wolfspraul | I don't easily believe in "cannot be optimized anymore" | 00:28 |
| wolfspraul | roh: see you just compared to ARM again :-) | 00:28 |
| wolfspraul | that was the main point I tried to get to above | 00:28 |
| wolfspraul | if we do that, we will always be behind | 00:28 |
| wpwrak | 4. difficult :) and again, you can perhaps squeeze out 20-30% on average if you put a lot of work into it. | 00:28 |
| wolfspraul | then it's better to just use ARM | 00:28 |
| roh | i think developing system blocks in fpga(s) to make a foss soc possible is a good thing. and also very flexible. but in the end it limits our system performance massivley if we stay at softcores. | 00:28 |
| wolfspraul | hence the make small dedicated asic idea | 00:29 |
| wolfspraul | the problem with that is - you first have to beat the core running in a 45nm (or next year 28nm) fpga | 00:29 |
| wolfspraul | that's why I as wasking about "are we done optimizing the lm32 core as-is" | 00:29 |
| wolfspraul | the optimization problem will only get bigger | 00:30 |
| wolfspraul | and in the very end, our great asic will be even slower :-) | 00:30 |
| wolfspraul | with our 3 hackers trying their best between social responsibilities :-) | 00:30 |
| wpwrak | you reach a certain point where the thing is just fast enough for the usual tasks | 00:30 |
| wolfspraul | yes | 00:30 |
| wpwrak | you still suffer aging (as software gets more bloated), but that's slower | 00:30 |
| roh | wolfspraul: my point is not arm or not arm. its "get it powerfull enough so we can sell it ourside of the foss enthusiasts's hole" | 00:32 |
| wolfspraul | yeah but we need to be honest about what we compare with, now or later | 00:32 |
| wolfspraul | the m1 is on a marketing path where it tries to have no-one compare it to an ARM, iphone, macbook, etc. | 00:33 |
| roh | wolfspraul: sure. i am. comparing it to 500mhz arm cpus is comparing it to crap from 5 years ago. | 00:33 |
| wolfspraul | so far not successful, but any path that leads us to comparing with those is a guaranteed fail imo | 00:33 |
| wolfspraul | that's why I am not comparing | 00:33 |
| wolfspraul | the moment you compare you are far better off with other options, drop milkymist | 00:33 |
| roh | stuff i buy now is >400mhz, regardless of how cheap, as soon as i can connect serious amouts of ram/flash (non-microcontroller market) | 00:34 |
| wolfspraul | moores law has more life in it, I fully expect chips to still get hundreds of times more powerful, cost is < 1 USD anyway per chip, and will stay there | 00:34 |
| wolfspraul | but what's the point from the perspective of milkymist soc? | 00:35 |
| wolfspraul | you could port flickernoise to the raspberry pi, and maybe even beat the m1 :-) | 00:35 |
| roh | wolfspraul: youve seen the carambola board? | 00:35 |
| roh | even by abusing a dsl-router soc we could make a better nanonote. maybe using a fpga to add a proper gpu | 00:35 |
| wolfspraul | no, googling. thanks! nothing better than new urls :-) | 00:35 |
| wolfspraul | I think we are best off tailor-making small and integrated logic boards that bring out the full power of Milkymist SoC today, and allow for healthy growth (i.e. adding new features) | 00:37 |
| wolfspraul | and that's it, and we take that into as many end-user products (sales) as possible | 00:37 |
| wolfspraul | and I say upfront: comparing megahertz or other performance with other chips will make you think this is prehistoric tech, so you may be better off comparing the latest qualcomm chips with the latest nvidia chips (for example), and can save your time about Milkymist | 00:38 |
| wolfspraul | we need to send people who want to do these comparisons away upfront, to be honest with them. the only way to like the Milkymist SoC is if you free yourself from those comparisons and only think about how much you can do with the Milkymist, not how much you cannot do. | 00:39 |
| wpwrak | look at who compares MHz | 00:39 |
| wolfspraul | then we grow the milkymist, hopefully, and of course that will include hardwired circuitry at some point, hopefully. because that would bring more performance, if enough time and money is invested, at the expense of loosing programmability/updatability for those parts. | 00:40 |
| wpwrak | customers of "universal" platforms do. you probably have at least a rough idea of how many GHz your PC's CPU has | 00:40 |
| wpwrak | now, how about your microwave ? :) | 00:40 |
| wolfspraul | I'm just explaining my thinking, anybody can compare whatever they like. | 00:40 |
| wolfspraul | but people come in good faith, then they compare, then they may be disappointed. so we are better served explaining to them upfront that there's another "megahertz myth" community here. | 00:40 |
| wpwrak | milkymist is by and large a system with a fixed feature set. you therefore don't have to have a lot of reserves to run all the stuff that _may_ come along. you still need a bit, but there's less pressure | 00:41 |
| wolfspraul | yes | 00:41 |
| wpwrak | so for the m1, marketing can solve the issue | 00:42 |
| wpwrak | now, with the hypothetical NN-Ya-FPGA-only, you're a lot closer to the universal platform | 00:42 |
| wolfspraul | no no | 00:42 |
| wpwrak | hence the MHz enter the spotlight | 00:42 |
| wolfspraul | we define the marketing message. marketing is free, the point is to find and create happy customers. | 00:42 |
| wolfspraul | no, they don't | 00:43 |
| wolfspraul | "megahertz myth" wasn't that a great invention? | 00:43 |
| wolfspraul | everybody knows that term nowadays | 00:43 |
| wpwrak | i don't remember hearing that one before | 00:43 |
| wolfspraul | I think it was Jon Rubinstein who created it, or is known to have created it. | 00:43 |
| wpwrak | just the GHz race :) | 00:43 |
| wolfspraul | oh no, that was big in the 90's | 00:43 |
| wolfspraul | at Apple, of course, since they were lagging in mhz | 00:44 |
| wolfspraul | so they thought about what they could do :-) | 00:44 |
| wolfspraul | and they started talking about 'megahertz myth' a lot! systematically | 00:44 |
| wpwrak | ;-) | 00:44 |
| wolfspraul | yep | 00:44 |
| wolfspraul | no need to give up, learn from the pros :-) | 00:45 |
| wolfspraul | http://en.wikipedia.org/wiki/Megahertz_myth | 00:45 |
| cladamw | wpwrak, i saw backlog. thanks for your feedback, I'll take some time to modify. :) | 00:45 |
| wolfspraul | "The term came into use in the context of comparing PowerPC-based Apple Macintosh computers with Intel-based PCs." | 00:45 |
| wolfspraul | and yes it was good old Rubi | 00:46 |
| wpwrak | funny. once upon a time, when i had that SGI at my disposition, the typical "modern" sun had a 68030 at some 20-30 MHz. the SGI has a MIPS at, i think, 12 MHz. yet it was much faster. even though it was RISC and didn't have the mighty instructions of CISC :) | 00:46 |
| wolfspraul | later at Palm, and now at Amazon behind the Kindle strategy | 00:46 |
| wpwrak | cladamw: ah great, you found it :) | 00:46 |
| wpwrak | cladamw: i have a few more comments. still have to type them, though | 00:46 |
| wolfspraul | even if we go asic, our mhz will lag FAR FAR behind anything you could possibly compare with, anything contemporary | 00:47 |
| wolfspraul | and like I said, you actually first have to beat the soft-core in a 45 or 28nm gate structure | 00:47 |
| wpwrak | i don't think we need to match apple and enemies MHz for MHz | 00:47 |
| wolfspraul | with our minuscule engineering resources even the tape-out, testing and communication process with a foundry could get us stuck for 10 years :-) | 00:47 |
| wolfspraul | no my point is: we are free in how we market and position our tech | 00:47 |
| wolfspraul | if we compare, then of course others compare | 00:48 |
| wolfspraul | if we invent smart terms like "megahertz myth" and stick to them, there is not much reason actually for people to not listen | 00:48 |
| wolfspraul | maybe we have a point, maybe not | 00:48 |
| wpwrak | you'd still find it difficult to sell devices that aren't fit for the intended purpose. at least in the long run | 00:48 |
| wolfspraul | but I am sure - comparing with arm/nvidia/qualcomm/samsung and others will go nowhere | 00:49 |
| wpwrak | so ... how would you market the NN-Ya-FPGA-only with its 80 MHZ (?) core ? | 00:49 |
| wpwrak | roh: seems that i found the right question (-:C | 00:54 |
| wolfspraul | nice http://geekculture.net/joyoftech/joyarchives/243.html | 00:54 |
| wolfspraul | megahertz myth cartoon | 00:54 |
| wolfspraul | click on 'next joy' to continue | 00:55 |
| wpwrak | heh, great minds think alike :) | 00:55 |
| wpwrak | the page was loading when i read your message | 00:56 |
| roh | wpwrak: hrhr | 01:14 |
| roh | wolfspraul: mhz is partially a myth. but in generic computing its still a fact. sure. we dont need 2ghz. but 80mhz are just not enough. | 01:14 |
| roh | i just tried ssh-ing onto a dvb-set-top-box. it tool like 20 seconds before it responded again because it was doing fscking crypto math | 01:15 |
| roh | why? because its a dirt cheap mips cpu which is ok to do controlling of data streams which dsps do the work on, but not for something simple as responding to a incoming ssh connection nicely | 01:16 |
| wolfspraul | yes | 01:16 |
| wolfspraul | but "80 mhz" is a number and unit, and "how long it takes to login with ssh" is a use case | 01:16 |
| roh | there are 2 answers to that question: one would be 'then dont do what its not designed to handle' ... which would lead to 'then dont do stuff' transferred onto the nn | 01:17 |
| wolfspraul | I don't know whether your dvb settop is selling well or profitable for the manufacturer, but I would guess that an ssh login is nowhere anywhere on the typical use cases for that product | 01:17 |
| wolfspraul | on the list of ... | 01:18 |
| roh | also its a question of 'computations per timeframe' . sure. if you make an usb stack which can do all the work which currently a cpu does on linux mostly in hardware' .. but it dont see that transferred to all subsystems | 01:18 |
| roh | wolfspraul: it was an example of 'whats in that stuff nowadays' | 01:18 |
| wolfspraul | so my thinking for m1 is quite clear now - one is to continue to focus on its current and intended use, and make that better. there are ample opportunities to do that. | 01:18 |
| roh | in that case.. a 5 year old soc. pre-h264-hd area | 01:18 |
| wolfspraul | and the other one is to build an integrated and simple logic board around the milkymist soc to bring out what it can do today, and create room for growth | 01:19 |
| wolfspraul | and in parallel, work on credible marketing messages and sell that box to users | 01:19 |
| roh | just an example. low end dsl-routers like the tp-link or the carambola soc have 300-400mhz mips cpus. which is ok mostly. just dont do any crypto or computation intensive stuff there. | 01:19 |
| roh | so you can route and switch packets. but crypto they only do sanely with hardware support (most router socs do have that nowadays. even the brcm soc in the original linksys wrt had) | 01:20 |
| roh | but compared to a 'user facing frontend node with display and keyboard' i must say that cpus need more punch. even if its only to drive a serious amount of pixels or decode a audio file for which you dont have a hardware codec | 01:21 |
| roh | ive learnt recently that nowadays even the cheap mp3 players dont do dsps anymore. they have small arm cpus and well enough written code to punch that work through in sw. why? because its cheaper then binding yourself to some dsp/cpu vendor and their toolkit and you can reactor market changes like new codecs faster. | 01:22 |
| wolfspraul | interesting little carambola board | 01:22 |
| wolfspraul | I have *no* idea how that company wants to turn this into a good business though | 01:23 |
| roh | yeah. i thought so. | 01:23 |
| wolfspraul | ralink was acquired by mtk | 01:23 |
| roh | business? well.. all the arduino crowd looking for more punch. | 01:23 |
| wolfspraul | they are making a module for a niche market against a small window of opportunity | 01:23 |
| roh | or serious ethernet connection | 01:23 |
| wolfspraul | ok but those are all tiny phenomena | 01:23 |
| wolfspraul | maybe a small lithuanian company can survive on that, don't know | 01:24 |
| wolfspraul | how much are those boards? | 01:24 |
| roh | sure. but they will get a lot of people like me who would atm do projects re-using routers like the tp-links | 01:24 |
| wolfspraul | people like you are a tiny phenomena :-) | 01:24 |
| wolfspraul | rounding error | 01:24 |
| wolfspraul | how much does this board cost? | 01:24 |
| roh | not because they are like 7E more expensive than the router, but because it dont need to hope for continoued supply of something which is a market accicent not a planned feat | 01:25 |
| roh | carambola was like 22E | 01:25 |
| wolfspraul | well like I said the main chip on that board is already gone as an independent business as well, acquired by mtk (similar to qualcomm's purchase of atheros) | 01:25 |
| roh | and a tp-link costs me 15-23E depending on where and how many | 01:25 |
| wolfspraul | yes | 01:25 |
| wolfspraul | but tp-link has massive (read: tens of millions) scale behind it | 01:25 |
| wolfspraul | and even that was not enough to keep Atheros independent | 01:26 |
| wolfspraul | why do you say 'was' - carambola no more? | 01:26 |
| roh | sure. still they could change the product anytime in a way that it makes it unpractical for me. | 01:26 |
| wolfspraul | I find the pricing crazy aggressive, well. that makes me wonder even more about their business. | 01:26 |
| roh | so carambola is intended to be used in verticals, while abusing openwrt based routers isnt. | 01:26 |
| wolfspraul | but carambola has zero influence over the chips either, for sure | 01:27 |
| wolfspraul | in fact, ralink is already history | 01:27 |
| roh | who's pricing? | 01:27 |
| wolfspraul | carambolas | 01:27 |
| roh | the chips they use are single digit money. the expensive part is sourcing i bet | 01:27 |
| roh | same as tp-links business. there is nothing to make it cheaper anymore. its commodity hardware. | 01:28 |
| wolfspraul | true, many tech companies are on a nice path to reduce prices to 0 | 01:28 |
| wolfspraul | nobody needs them | 01:28 |
| wolfspraul | but... tp-link will increase volume and has a giant global distribution network which I would well imagine is growing | 01:29 |
| wolfspraul | and either you have that or not. imagin you want to get into the AA/AAA batteries business. | 01:29 |
| roh | i think they cannot get much more market share. their spectrum is low-end home cpe stuff. | 01:29 |
| wolfspraul | what do you need so that your batteries are flying around in millions of stores worldwide? | 01:29 |
| wolfspraul | sure, and? | 01:29 |
| roh | high end is with asus and companies like arcadyan or avm which do oem for the big telcos. | 01:30 |
| roh | from my pov the real numbers are done via the oem telco designs. not the independant sales. | 01:31 |
| wolfspraul | and you think those customers will not want to squeeze out every penny? | 01:31 |
| roh | check heise for the arcadyan scandal regarding broken wifi keys (default keys) | 01:31 |
| wolfspraul | their certification cycles may be longer that's why they appear a few generations behind | 01:32 |
| roh | they have 5-6 digits of active devices per model | 01:32 |
| wolfspraul | what you describe as 'high-end' | 01:32 |
| wolfspraul | why did you say carambola 'was'? | 01:32 |
| roh | high end is gigabit ports, integrated vdsl modems etc. many features etc. usually has more recent soc, more ram and flash | 01:32 |
| roh | carambola uses a low-end soc | 01:33 |
| roh | wolfspraul: s/was/is/ . sorry to confuse | 01:33 |
| roh | http://shop.8devices.com/wifi4things/carambola | 01:34 |
| wolfspraul | yeah, impressive | 01:34 |
| wolfspraul | I have no idea how that can compete with the flood of cheap tp-link models though | 01:35 |
| wolfspraul | the only thing that differs are the many pins you can connect to | 01:35 |
| wolfspraul | and the it becomes a 'development board' as they say. maybe sparkfun/adafruit would be the strongest sales channels for that, but even then I doubt you can sell even a few thousand of it. | 01:36 |
| roh | and that its intended to be used in verticals. and the site | 01:36 |
| roh | eh size | 01:36 |
| wolfspraul | then you do the math, and I have no idea, really none, how you can finance anything with those margins | 01:36 |
| wolfspraul | it's the old problem of most 'module' makers | 01:36 |
| roh | i think it has a real chance as kind of a high-end arduino/arduino companion | 01:36 |
| roh | sure. its a market nice | 01:37 |
| roh | eh niece? | 01:37 |
| roh | niche. thats the order of chars i wanted | 01:37 |
| wolfspraul | at adafruit/sparkfun a product with that type of bom would probably cost 40 USD or more | 01:37 |
| wolfspraul | and even then it's a small hobbyist/nerd market | 01:38 |
| GitHub190 | [board-m1] adamwang pushed 1 new commit to master: https://github.com/milkymist/board-m1/commit/9a3c2737e8e95452f27afbcc4bb7bca7202d93d9 | 01:38 |
| GitHub190 | [board-m1/master] make D+, D- lines with more clearance - Adam Wang | 01:38 |
| wolfspraul | the moment any actual vertical business would start using this, they would quickly optimize the board away as their volume goes up | 01:38 |
| roh | maybe. who knows what that lithuanian company does the rest of the day. maybe they do oem dsl routers all day and thats just a 'additional nice product' they could do because they wanted | 01:38 |
| wolfspraul | sure, it's nice. development board with unbelievably low margin. | 01:39 |
| roh | wolfspraul: sure. but some verticals dont exist when there is enough knowledge to do bga boards etc. | 01:39 |
| roh | wolfspraul: often there is 'we need foo, but only 10-100 times' | 01:39 |
| wolfspraul | all module makers I have seen are transient businesses | 01:39 |
| roh | think art projects.. they use a lot of arduinos for that | 01:39 |
| wolfspraul | after 2 years they are either dead or are doing something else, higher margin | 01:39 |
| roh | i think only extremely few people who prototype with arduino or similar boards build their own boards in the end. most just use and adapt some existing board. there is only gain in optimizing that away (and having the risk to fuck it up) when you do real mass-market | 01:40 |
| wolfspraul | sure | 01:41 |
| wolfspraul | which arduinos have you seen growing into vertical industry use btw? | 01:41 |
| wolfspraul | and 2 other questions: how many carambolas have you bought so far? and how many cheap tp-link routers? | 01:41 |
| roh | not industry. vertical sure. projects for artists | 01:42 |
| GitHub8 | [board-m1] adamwang pushed 1 new commit to master: https://github.com/milkymist/board-m1/commit/dbc59e37ae1dce363b74b1a8fdeb016f9c5baf7f | 01:42 |
| GitHub8 | [board-m1/master] make ground symbols downward if there's more space. otherwise upward. - Adam Wang | 01:42 |
| roh | carambolas none (because of paypal and my lazyness) and tp-link 2 | 01:42 |
| wolfspraul | does it grow into commercial/consumer electronics use anywhere? | 01:42 |
| wolfspraul | he | 01:42 |
| wolfspraul | so the small inconvenience of carambolas unusual sales channel has kept you away from them so far :-) | 01:43 |
| roh | but i have a project where we will use 20 or so of the routers soon. but we need a switch and the case so it will be tplink i guess | 01:43 |
| wolfspraul | and above I was talking about tp-links giant global sales network with certainly hundreds of thousands of points-of-sale | 01:43 |
| roh | wolfspraul: you need to get away from the massmarket aspect a bit. sure, thats your pov. | 01:43 |
| wolfspraul | and... more tplink :-) | 01:43 |
| wolfspraul | no no | 01:43 |
| wolfspraul | I am just thinking about what you say | 01:43 |
| wolfspraul | and you admit: so far 0 carambola because the sales channel was too inconvenient | 01:44 |
| wolfspraul | tp-link was easier | 01:44 |
| wolfspraul | and the upcoming project, also tp-link | 01:44 |
| wolfspraul | interesting :-) | 01:44 |
| wolfspraul | maybe you are not alone... | 01:44 |
| roh | but there are many other products which would never work in a mass-market but only their niches. which is nice, because massmarket products are boring and mostly optimized to death and thus less flexible | 01:44 |
| wolfspraul | I absolutely don't just think mass-market, not at all | 01:44 |
| wolfspraul | yes | 01:44 |
| wolfspraul | but we are talking about carambola | 01:44 |
| wolfspraul | and I think hacking tp-link boards looks more worthwhile doing | 01:44 |
| roh | i have one long term project where i will test carambola for, which i had planned with a pcengines board till recently | 01:45 |
| wolfspraul | although even that is such small volume that it will never even make it to the radar at tp-link | 01:45 |
| wolfspraul | (sorry btw this is all off-topic for milkymist) | 01:45 |
| roh | depends on how small i think my userland will be what board i will choose in the end | 01:45 |
| wolfspraul | and then about the arduinos, the last years that I am following I don't see that growing into consumer use | 01:45 |
| wolfspraul | not just 'mass market', that's not what I mean | 01:45 |
| roh | wolfspraul: i think their goal and businessplan never intended to compete with tp-link in the first place | 01:46 |
| wolfspraul | but the tech stays with a user base of 'experts', 'hobbyists', 'artists', 'nerds' | 01:46 |
| roh | fuck customers. they are boring. | 01:46 |
| roh | ;) | 01:46 |
| wolfspraul | yep | 01:46 |
| wolfspraul | that's roughly what the guys at mtk, nvidia, qualcomm, arm, and and and would be saying, just about 'hackers' :-) | 01:46 |
| roh | i dont believe that. they know that those are the ones working for them and having all the creative ideas who make new products. | 01:47 |
| wolfspraul | I am not blaming carambola for anything. I think they know what they are doing. like you said, they may have needed this board for themselves, and then just sell a few as well. that's all. | 01:47 |
| wolfspraul | nah, too small | 01:47 |
| wolfspraul | too few | 01:47 |
| roh | the world IS small. | 01:47 |
| wolfspraul | not bad at all, but just too few | 01:47 |
| wolfspraul | please look at the margins | 01:48 |
| wolfspraul | if something like carambola is priced at 100 USD or more, and they can find customers, that's a different story | 01:48 |
| roh | atleats when it comes to where the tech-savyness starts and ends. the majority of that stuff we are talking about is developed and run by very few people. | 01:48 |
| wolfspraul | it's easy to complain about the big guys being ignorant, when you forget what margins they work with, and are forced to work with by us, the customers | 01:49 |
| roh | i am not talking about making the current nn or a future one a tool for the mass-market. but to find a way to make it more interrresting to the tech-savy customers we have now | 01:49 |
| roh | sure. there are no margins in commodity electronics for the mass market. but did we ever seriously try to target that? i dont think so (and i am glad. customer support is hell there) | 01:50 |
| roh | there are much more gains in the 'knows more what he/she is doing and likes fancy toys' market. | 01:51 |
| wolfspraul | not sure, hackers are cheap :-) | 01:51 |
| wolfspraul | their hacking budgets get cut by their girlfriends | 01:51 |
| wolfspraul | but the arduino still passes, thank god | 01:52 |
| roh | thats why people need to get a bang for their buck. not cheap, but good and fairly prices is the sweet spot. | 01:52 |
| roh | the hackers i know have girlfrieds who also hack *sigh* | 01:53 |
| roh | +n | 01:53 |
| roh | maybe we should make a nn with a minipci-e slot, and not wire pci-e there but only the usb. then atleast a 3g modem would work | 01:54 |
| roh | *sigh*. i think before concepting anything new there we need to fix our mechanical problems. or is using what the nn did still an option? | 01:56 |
| wolfspraul | I agree mechanical needs to be simplified | 01:57 |
| wolfspraul | man I feel bad about spamming the milky channel, taking it to qi-hw | 01:57 |
| roh | sure | 01:58 |
| kristianpaul | keep up with its application, and the history of the less know pdf reader feature :) | 03:04 |
| wpwrak | touch\'e ;-) | 03:06 |
| Fallenou | wolfspra1l: I think the 80 MHz softcore is enough for the M1 product for now (it does it's job with no obvious lag). it's a pretty precise use case though | 13:30 |
| wolfspra1l | yes | 13:30 |
| Fallenou | wolfspra1l: for something like a general purpose computer (like the Ben is AFAIK), it might be really slow | 13:30 |
| wolfspra1l | might :-) | 13:30 |
| Fallenou | high probability that it will :/ | 13:31 |
| Fallenou | Having read lm32 softcore verilog code a bit while working on the MMU, I don't see a lot of room for improvement *BUT* I don't have the verilog expertise of lekernel | 13:31 |
| wolfspra1l | is the cpu pipelined? | 13:32 |
| wolfspra1l | I would think so | 13:32 |
| Fallenou | it's a 6-stages pipeline | 13:32 |
| wolfspra1l | could we go multicore? | 13:32 |
| Fallenou | address/fetch/decode/execute/memory/write back | 13:32 |
| wolfspra1l | how about caching? | 13:32 |
| wolfspra1l | branch prediction? | 13:32 |
| Fallenou | caching is already enabled | 13:32 |
| wolfspra1l | out of order processing? | 13:32 |
| Fallenou | branch prediction is there, but maybe it can be improved, it's a really really prehistoric one | 13:32 |
| Fallenou | multicode : I have no idea if this could be easily done , but yes, why not :) we would have to calculate the FPGA resources to check if it fits inside one FPGA | 13:33 |
| Fallenou | multicore* | 13:33 |
| Fallenou | out of order ... I think it's not implemented right now in LM32 (no sure), seems hard to implement. maybe lekernel out of order DRAM controller will be enough to gain more performances | 13:35 |
| Fallenou | keep in mind that the more gates you add in the LM32 design, the more risk you take to see your "max working frequency" to actually decrease | 13:36 |
| Fallenou | but it does not mean we cannot give it a try :) it just means it's risky if you don't pay attention | 13:36 |
| Fallenou | Hopefully, FPGA technology improvements will give us extra MHz without any LM32 code change, I don't know how much though ... could we reach 200 MHZ ? with which FPGA ? | 13:37 |
| kristianpaul | may altera? may be takinng apart lm32 separate could clock higher than a SoC | 13:42 |
| Fallenou | not so sure which part is slowing down the other: the soc or the lm32 | 13:43 |
| kristianpaul | soc have some extra wiring remenber wishbone and FML | 13:43 |
| Fallenou | sure | 13:44 |
| kristianpaul | and i guess we could get a better timing report, just we/i need to try.. | 13:44 |
| Fallenou | could be interesting to synthetise a migen based (milkymist-ng) soc with just uart+sdram+lm32 to see the max working frequency | 13:45 |
| kristianpaul | i would said that for old soc, yes is interesting | 13:46 |
| kristianpaul | On Arria2 it reaches 175Mz | 13:46 |
| kristianpaul | http://www.ohwr.org/projects/lm32 | 13:46 |
| kristianpaul | But those where most simulations i dont remenber well what Terpstra said | 13:46 |
| kristianpaul | Fallenou: also, once you pass out 100Mhz barrier, i bet interesting behavior pop up :D | 13:47 |
| Fallenou | this demi runs @ 50 MHz : http://www.latticesemi.com/products/intellectualproperty/ipcores/mico32/mico32ddrsdramdemo.cfm | 13:49 |
| Fallenou | http://www.latticesemi.com/products/intellectualproperty/ipcores/mico32/index.cfm at the bottom of the page | 13:50 |
| Fallenou | 115 MHz on LatticeECP3 | 13:50 |
| Fallenou | which is a basic Lattice FPGA, not a big big fpga :) | 13:50 |
| Fallenou | I own a dev board with this fpga | 13:50 |
| Fallenou | $99 dev board | 13:50 |
| kristianpaul | oh | 13:50 |
| kristianpaul | so slow.. | 13:50 |
| kristianpaul | :) and is lattice | 13:51 |
| wolfspra1l | we should not compare what we know runs on our board with all sorts of unverified numbers from web sources | 13:57 |
| wolfspra1l | with our experience, we should not :-) | 13:57 |
| wolfspra1l | there are a hundred ways how you can reach "xxx mhz" | 13:57 |
| wolfspra1l | no | 13:57 |
| wolfspra1l | a thousand | 13:57 |
| wolfspra1l | it's not just good/bad, it's that we don't know the details, then we just carry forward the number and forget all sorts of compromises that may have been made | 13:58 |
| wolfspra1l | like "ok, it froze after about 10 seconds, but until then it ran nicely at xxx mhz" | 13:58 |
| wolfspra1l | or hundreds like that | 13:58 |
| Fallenou | hum it's official Lattice demo, I bet they don't publish demos which freezes after 10 seconds :) | 13:59 |
| Fallenou | *brb* | 13:59 |
| wolfspra1l | oh | 13:59 |
| wolfspra1l | a real company demo | 13:59 |
| wolfspra1l | wow | 13:59 |
| wolfspra1l | then that number must be clean like the pope's shirt | 13:59 |
| wolfspra1l | sorry I take back all I said | 13:59 |
| wolfspra1l | it's a real company demoe | 13:59 |
| wolfspra1l | wow | 13:59 |
| wolfspra1l | I read a report the other day about the "megapixel" ads in smartphones | 14:00 |
| wolfspra1l | the number that is published in advertisement, and any actually different 'pixels' you can determine in the outcome | 14:00 |
| wolfspra1l | oh well | 14:00 |
| wolfspra1l | let's say those numbers have taken on life of their own | 14:00 |
| wolfspra1l | some vendors (ahem, cough, Apple) are publishing mostly actual numbers that can be verified, sometimes what apple publishes is even less than the actual measured performance | 14:01 |
| wolfspra1l | but a lot of the others, oh well | 14:01 |
| wolfspra1l | published number 50% above what can be measured | 14:01 |
| wolfspra1l | 70% | 14:01 |
| wolfspra1l | 100% | 14:01 |
| wolfspra1l | etc | 14:01 |
| wolfspra1l | so back to reality, if this is an "official lattice demo", then FOR SURE you should not believe that number | 14:02 |
| wolfspra1l | "official lattice demo" means that their marketing decided that it had to be that number | 14:02 |
| wolfspra1l | and then engineering had to somehow fudge it up, but marketing would say it anyway | 14:02 |
| wpwrak | ;-) | 14:06 |
| wpwrak | at least if the result is great enough for the competition to hate it, it may get some vetting | 14:06 |
| wpwrak | some of the more outrageous "optimized" benchmarks have been exposed in such situations | 14:07 |
| wolfspra1l | well, it's just the fact | 14:07 |
| wolfspra1l | "official company demo" means, to anyone experienced, that the number will not hold in real life | 14:07 |
| wolfspra1l | OF COURSE NOT | 14:07 |
| wolfspra1l | look, give you an example :-) | 14:07 |
| wolfspra1l | nvidia is a successful semiconductor company | 14:08 |
| wpwrak | but anyway, none of these numbers sound too promising. i don't see anyone claiming more than a 100% increase, relative to what we have now | 14:08 |
| wolfspra1l | we all agree | 14:08 |
| wolfspra1l | the CEO has repeated many times his philosophy, in interviews, etc. | 14:08 |
| wolfspra1l | 1. sell first | 14:08 |
| wolfspra1l | 2. worry how to make it | 14:08 |
| wolfspra1l | he is right! | 14:08 |
| wolfspra1l | *if* a customer comes and puts money on the table, anything can be made | 14:08 |
| wolfspra1l | so first claim something | 14:08 |
| wolfspra1l | see whether money shows up | 14:08 |
| wolfspra1l | *then* talk to your engineers :-) | 14:08 |
| wolfspra1l | that's not a secret or some fraud or anything | 14:09 |
| wolfspra1l | that's how the system works, in the example of a very successful company (nvidia) | 14:09 |
| wolfspra1l | just saying | 14:09 |
| wolfspra1l | Fallenou can think about it if he wants :-) | 14:09 |
| wolfspra1l | sell first, then worry about making it... | 14:09 |
| wpwrak | as long as your engineers can always deliver enough on your promises, no problem | 14:09 |
| wolfspra1l | that's their problem | 14:09 |
| wolfspra1l | that's why they get paid :-) | 14:09 |
| wpwrak | i somehow prefer the engineering-run companies ;) | 14:10 |
| wolfspra1l | the nvidia ceo also believes in hiring the top talent from the whole world because he believes talent is so limited that if you hire them all, your competitors have a problem simply from the fact that all the good guys are gone already | 14:10 |
| wolfspra1l | those things all go together | 14:10 |
| wolfspra1l | so "company demo" does not mean that it actually works | 14:10 |
| wolfspra1l | :-) | 14:10 |
| wolfspra1l | just remember that | 14:10 |
| wolfspra1l | it does not | 14:10 |
| wolfspra1l | not all companies would be as drastic as "sell first, then worry how to make it", but it's a good reminder of a valid strategy of a very successful semiconductor corp | 14:11 |
| wolfspra1l | certainly far more successful than lattice :-) | 14:11 |
| wpwrak | well, different resources, different customers, probably also requirements changing at a different pace | 14:12 |
| wolfspra1l | but you would agree with me that "company demo" is a very good indicator that whatever they demo does not *actually* work :-) | 14:13 |
| wolfspra1l | that's our bottom line | 14:13 |
| wpwrak | and the negotiators may also be from different branches | 14:13 |
| wolfspra1l | that's why it's called a "demo" | 14:13 |
| wolfspra1l | maybe we all want this to work, yeah :-) | 14:13 |
| wolfspra1l | a dream | 14:13 |
| wolfspra1l | company dream | 14:13 |
| wpwrak | demo can be pretty much anything. maybe it works, maybe it's a complete fake. (e.g., a video instead of the machine doing it in real time) | 14:14 |
| wpwrak | maybe it's easy to reproduce (with some tolerance), maybe not. | 14:14 |
| wpwrak | maybe the thing you see being demoed works on the inside like your application, maybe not | 14:15 |
| wpwrak | the only somewhat reliable benchmark is your own application | 14:15 |
| wpwrak | ergo, make sure it's easy to port :) | 14:16 |
| wolfspra1l | sure | 14:19 |
| Fallenou | 16:15 < wolfspra1l> sell first, then worry about making it... < I totally agree. Every company does this :) | 14:36 |
| wolfspra1l | ok then, we are all on the same page about the 'company demo' and can focus on the demonstrable performance on m1 :-) | 14:50 |
| wolfspra1l | not demonstrable, *usable* even :-) | 14:50 |
| Fallenou | sure it's a demo, but they publish documentation and source code, they can't lie about the working frequency :p | 14:59 |
| Fallenou | it's too important a number for them to lie | 14:59 |
| Fallenou | anyway , their number is not crazy, so it is not really hard to trust | 15:00 |
| Fallenou | it must be right (115 MHz) | 15:00 |
| Fallenou | it's not like 200 MHz :) | 15:00 |
| Fallenou | IIRC Milkymist was running at 90 or 100 MHz a few years back :) | 15:00 |
| Fallenou | on Virtex 5 *agreed* | 15:01 |
| Action: Fallenou heading home, bye | 15:15 | |
| wpwrak | maybe it's more than a demo. more like an application example | 15:15 |
| roh | r | 17:19 |
| roh | e | 17:19 |
| hypermodern | At LGM talking about Qi Hardware | 17:23 |
| sb0 | and, as always, I have a good reason not to be at LGM... :( | 17:52 |
| Action: Fallenou is back | 19:23 | |
| qi-bot | The firmware build was successful, see images here: http://fidelio.qi-hardware.com/~xiangfu/build-milkymist/milkymist-firmware-20120504-2123/ | 21:06 |
| qi-bot | The firmware (using branch) build was successful, checkout the VERSIONS for detail, see images here: http://fidelio.qi-hardware.com/~xiangfu/build-milkymist/milkymist-firmware-20120504-2306/ | 22:51 |
| --- Sat May 5 2012 | 00:00 | |
Generated by irclog2html.py 2.9.2 by Marius Gedminas - find it at mg.pov.lt!