4 ms·
NovuMind is developing a deep learning chip to “do inference efficiently”
- deepnotderp 9y agoInteresting, but inference only and convolutions only is kind of a non-starter for a specialized chip, since the field is simply too dynamic for it to be really useful to focus on only 3x3 convolutions. I'm guessing FWIW, that this guy is implementing the Winograd filtering method or something similar in hardware, hence his restriction to 3x3 kernels.
- Iv 9y agoThere is a slight risk but I think it does make sense. There has been papers on the advantage of 5x5 and 7x7 convolution filters and the consensus seems to be that 3x3 are probably the way to go. If this was out, right now, it would have tons of use. The field moves fast but it is credible that in one year the field will still uses 3x3 and it is certain that being able to do todays applications for a few percents of the energy used now to do them is going to have business applications. Research moves fast but user usage move at a different speed. The first GPU came out when people were not sure if we would be doing triangle-based renderings, realtime raytracing, or something more funky.
- deepnotderp 9y agoThat's a fair enough point, but i don't really see why you need to restrict yourself to only 3x3 when you can support others at a slight penalty. For that reason, i suspect it's a winograd convolution ASIC or some similar algorithm.
- Iv 9y agoMy guess is that they need something out fast. The simpler the chip, the better. Adding features means adding test, creating additional circuits and software, etc... For the first version, especially if you have investors who poured a few dozen millions in, you want something that works ASAP. Then version 2 will receive tons of funds to have all the features one can dream of.
- baybal2 9y agoThe penalty is not that slight. As said in the article, Google TPUs struggle to feed themselves data, while their big convolution units simply ends up being a waste of die area
- 0xbear 9y agoI wish I could bet money on this not happening "by February". Dude is promising perf that's 25% higher than Titan X Pascal in a power envelope that's 1/50th of Titan X Pascal, and he's promising to do it with a few tens of millions of dollars and a handful of people. Yeah, right. Like that's ever going to happen. Sounds like the usual grandiose claims to get clueless moneybags to pony up the cash for series B.
- jamiek88 9y agoYeah, extraordinary claims and all that. Especially as this guy has a credibility problem in the first place because of the ML cheating scandal from a few years ago (1) he was in the middle of. 15 tflops of even 3x3 stacked in 5 watts? 'Taped out by Feb' - I am sceptical to say the least. (1) https://www.enterprisetech.com/2015/06/12/baidu-fires-deep-images-ren-wu/ https://www.enterprisetech.com/2015/06/12/baidu-fires-deep-i...
- rayuela 9y agoOh wow I had not realized that was him. That was pretty bad, what they were doing was outright dishonest. I wouldn't trust them with a dime of my money.
- dnautics 9y agoIt's not impossible. I've seen a company go from zero to tapeout in six months. However, they did wind up getting stuck due to bugs on the motherboard. (the chip did work)
- deepnotderp 9y agoIt's totally possible with a few million dollars and a couple of guys, especially for a chip this simple. See for example the Adapteva Epiphany (one guy on 16nm) and Rex Computing.
- 0xbear 9y agoBut the chip is _not_ simple. You can't get 15TFLOPS with a "simple" chip. Even TPU1 is not simple and it took Google over 2 years to get that from the drawing board to the data centers. Moreover, 15TFLOPS, in the unlikely event that they design and build it, is not going to be 5W. Unless they know something the likes of Intel, NVIDIA etc don't know. The probability of that is not just vanishingly small, it's zero.