varelse

77 ms·

511.
by varelse 10y ago
Crazy idea: buy the rights to sell AMD's Vega GPUs fabbed out of Intel and use Intel's resources to build top-notch math and AI libraries for them. Stupid idea: Keep insisting that x86-compatibility is the killer feature for winni
512.
by varelse 10y ago
Sure, great, but if they could build this to support Deep Learning and renegade HPC, I suspect they could sell 10,000+ units/year on the down-low. That would be $10M in revenue, maybe twice that if it also worked with Macs. NVIDIA wi
513.
by varelse 10y ago
Tell that to the creators of NCCL and DSSTNE, who have both achieved ~12.5 GB/s O(n) collectives on 16x PCIE Gen 3 across an arbitrary number of GPUs given they are connected across a network of PCIE switches with no intervening CPU in
514.
by varelse 10y ago
Except that if you want O(n) performance, the easiest topology is a ring connecting all the GPUs without having to pass through any CPUs. Passing through 1 CPU will incur a ~20% bandwidth penalty, and passing PCIE data directly between 2 C
515.
by varelse 10y ago
The right way to build this sort of thing is with a hub of 8796 PCIE switches for each group of 4 GPUs such that they could all form a continuous 16x PCIE bidirectional ring suitable for O(n) collectives like gather/reduce/allGath
516.
by varelse 10y ago
Ya know, it's not like this was some dude who asked a co-worker out, got turned down, and that was the end of it. Nope, this is full-on sexual harassment enabled by HR that went so far as to ambiguously punish the victim. Every time I
517.
by varelse 10y ago
The thing is that it didn't seem like that at the time. Someone high up went around the company to see if there was a fit for me, claimed there wasn't, and then said I could either give up on the technology or go work somewhere e
518.
by varelse 10y ago
That sounds right. I was on a team leaking personnel managed by a manager who clearly wanted to be doing other things with his other team and who forgot I existed for the 3 out of the 4 months I was there until I reminded him I existed whe
519.
by varelse 10y ago
Google is awesome if you can avoid the allocation problem, I agree. I was not given such options.
520.
by varelse 10y ago
Sure, that would be great, but as a result of leaving, I have been blacklisted by HR from ever returning. And the one guy that tried to bring me back in a year ago met in the lobby of his building and then walked me out to the marsh behind
521.
by varelse 10y ago
I really like the first part of the paper and it describes a lot of their best practices, but the last 1/3 or so of the paper is more about Google's culture, no? IMO that opens them up for the snark here. Google has some amazing
522.
by varelse 10y ago
Except no, I saw stuff like that too when I was there along with childish crap like throwing gum in the urinals and googlers acting like entitled jerks towards the cleaning staff. I met some great people at Google, but I'd be remiss no
523.
by varelse 10y ago
Not one word about what I consider the most toxic aspect of working at Google: blind allocation. Unless one absolutely does not care what one wishes to work on, joining Google is throwing your future into the Hogwarts hat of an ill-defined
524.
by varelse 10y ago
It makes for the same sort of easily repeatable distractoid like bleating that Obama did the same thing(tm) in 2011. http://foreignpolicy.com/2017/01/30/sorry-mr-president-the-o...
525.
by varelse 10y ago
According to this study of homeless people in nearby Santa Cruz ( https://blog.civinomics.com/2015/04/07/the-reality-of-homele... ), ~1/3 the climate, ~1/5 each the social programs and lifestyle, foll
526.
by varelse 10y ago
When in Austin, I've talked to a lot of former Uber drivers who have moved to other ridesharing services like Wingz. I have yet to find one that doesn't prefer their current service over Uber.
527.
by varelse 10y ago
Do these attempts make use of the documented idioscynracies of Ms Pacman? http://gameinternals.com/post/2072558330/understanding-pac-m...
528.
by varelse 10y ago
It's also inferring the game state from looking at the screen rather than being spoonfed the data. One could conceivably perform a depth-limited search on the actual game state if it were available and then use an AlphaGo-like DNN to p
529.
by varelse 10y ago
Conventional AI can clean up at Pacman if it can do tree search on the game state, source: CS188 out of Berkeley. http://ai.berkeley.edu/lecture_videos.html
530.
by varelse 10y ago
I can only respond with "Opportunity is missed by most people because it is dressed in overalls and looks like work" which is my favorite Edison quote. A 9 to 5 effort IMO will have disappointing 9 to 5 results. If you offset tha
531.
by varelse 10y ago
IMO the travesty is crippling FP16 on consumer GPUs for exactly the reasons you cite. And given that AMD's Vega GPUs apparently will possess full speed FP16, it appears to be yet another in a long line of arbitrary performance cripp
532.
by varelse 10y ago
Rather than attempt to out these downlow Ge Force deep learners, why don't you ask yourself why you can only buy Titan X Pascal from NVIDIA itself.
533.
by varelse 10y ago
Virtualization? Don't care, in fact, virtualization is what disabled P2P copies and created craptastic upload/download perf on AWS until the P2 instance. DMA/CUDA Direct? Say Hello to P2P and staged MPI transfers, faster, c
534.
by varelse 10y ago
Informal anecdotal knowledge: GM wants 10-20 watts for the entire self-driving system, sensors and all.
535.
by varelse 10y ago
Seems indeed, and in fact an A/C will eat a kilowatt or more, but... The components of this system need to generate next to zero heat and they need to be placed in all sorts of inconvenient locations. That's causing automakers to
536.
by varelse 10y ago
NVIDIA's Drive PX 2 has too high power consumption and too low perf/W for the moment. They're winning this space because they are there more than that they are the best possible solution. And they may continue to win because
537.
by varelse 10y ago
100 of which GPU? And how were they connected to each other? They almost never give enough details to really understand these bold claims. If this is 100 K40s in a 10 Gb/s datacenter, this is old hat. If this is 100 Titan XPs in Bi
538.
by varelse 10y ago
It's more like Rock, Paper, Scissors, Data Scientist, except: NFL quarterback crushes puny data scientist Movie star steals data scientist's girlfriend Stock picker acqui-hires data scientist's company on the downturn for pen
539.
by varelse 10y ago
Physical space on the motherboard IMO is the true limiting factor here. And this is what made Google's TPU interesting IMO. I don't think it demonstrated world-beating performance, but it likely fit a strict power and size env
540.
by varelse 10y ago
I don't use NCCL, I wrote my own ring collectives after reinventing them in 2014 (I felt really clever for a whole 20 minutes before a Google search popped my balloon) to avoid working with a vendor that was trying to gouge us on serve

More ›