jhj

16 ms·

91.
by jhj 8y ago
I've thought of that, but the problem is that it needs to linearly interpolate between the more accurate values, and depending upon how finely grained the linear interpolation is, you would need a pretty big fixed point multiplier to d
92.
by jhj 8y ago
Multiplication or division by a power of 2 can be done by bit shift assuming binary numbers represent base-2 numbers; i.e. not a beta-expansion https://en.wikipedia.org/wiki/Non-integer_representation where binary numb
93.
by jhj 8y ago
At least 69% more multiply-add flops at the same power iso-process is nothing to sneeze at (we're largely power/heat bound at this point), and unlike normal floating point (IEEE or posit or whatever), multiplication, division&#x2F
94.
by jhj 8y ago
The ELMA LUT problem is in the log -> linear approximation to perform sums in the linear domain. This avoids the issue that LNS implementations have had in the past, which is in trying to keep the sum in the log domain, requiring an even
95.
by jhj 8y ago
Weyerhaeuser owns something in the neighborhood of 12 million acres of timberland in the US, though I don't know what their largest contiguous parcel is. They are probably the largest private landowner in the US (corporate, not individ
96.
by jhj 8y ago
I worked on a big commercial MMO engine for about a decade. In our implementation, the player runs ahead on the client (client autonomous) but server verified (actions replayed on the server). The new authoritative server position for the p
97.
by jhj 8y ago
When as a NYC renter it is impossible to live in near half of the rental housing stock at any price due to de facto permanent tenancy, this only pushes up the price of the housing that is possible to rent (the exempt buildings). Instead you
98.
by jhj 8y ago
Also, many days the system is outright not running at all, at least as far as displaying arrival times in the station, even on the L. This seems to be the case many weekends for some reason.
99.
by jhj 8y ago
Rather than these, the original Winograd algorithm ("A New Algorithm for Inner Product", 1968) is more useful in practice, if not for matrix multiplication in software on the typical CPU architecture. This decomposition is directl
100.
by jhj 9y ago
Yep, also SGD is not necessarily implied either (for example, full batch gradient descent).
101.
by jhj 9y ago
Minima/maxima in high-dimensional spaces are not much of an issue; the bigger issue are saddle points. cf https://arxiv.org/pdf/1406.2572.pdf This is presuming that the problem has a continuous formulation anyways
102.
by jhj 9y ago
If you were to pick only one and don't mind a flight and a (super long) drive from Phoenix, Albuquerque or El Paso, Gila NF in New Mexico is the place to go. As I recall, driving along the northern fringe on route 59 I experienced prob
103.
by jhj 9y ago
This map provides a pretty good guide, and I've found it to be fairly accurate as to orders of magnitude even if its data is dated: http://darksitefinder.com/maps/world.html To get truly dark skies you'll hav
104.
by jhj 9y ago
If you have “generalized multicore computing”, you still need to generate a code schedule for your grid / hypercube / what have you of CPUs or PEs, unless you want to program all that stuff by hand. Compilers 30+ years on still ar
105.
by jhj 9y ago
Prices include wages too. If all wages and prices adjusted instantaneously, then there is no real difference between inflation or deflation. However, people are mentally resistant to being paid less (nominally) for a job, even though in inf
106.
by jhj 9y ago
Usually a Morton ordering is used for things like improving average memory locality of N-dimensional data (e.g., loop iteration order, data layout, ...). But it is average locality that is being improved, because of huge jumps between many
107.
by jhj 9y ago
A GPU database isn't that useful, because the arithmetic intensity (ops/byte) is relatively low. Cross-sectional memory bandwidth is what really matters; you can get similar effects with a cluster of CPU machines provisioned appro
108.
by jhj 9y ago
You can express some forms of #2 or #3 well on a GPU. It depends upon how wide the graph of tasks is (maximum number of concurrent tasks possible in the graph). On Nvidia GPUs, 16 to 32 warps per SM x 60 SMs on a P100 gives a lot of hardwar
109.
by jhj 10y ago
ConvNets can handle sequences just as well (cf. TDNNs), you just treat time as another dimension. Whether you want to call this a "frame" or not is subject to question. https://arxiv.org/pdf/1611.02344.pdf Wi
110.
by jhj 10y ago
Remove is supported for some indices. The indexing scheme (IVFPQ, IMI, etc.) is based on the statistics (distribution) of the underlying data. If the statistics of the new data are not changing much, then it is reasonable to just remove the
111.
by jhj 10y ago
Faiss query, at least for IVFPQ which is the main emphasis of the library, is based on real vectors as input but against quantized database vectors; the so called “ADC” (asymmetric distance computation), so you don't have to query base
112.
by jhj 10y ago
Hi, I’m the author of the Faiss GPU code. The GPU implementation depends upon a lot of particularities in CUDA (warp shuffles and warp ballots, tricks with using register memory), and gets much of its speed in this way. Much of the code wou
113.
by jhj 10y ago
The "3,854 cores" versus "28 cores" is dubious as always. 3,854 I think counts just the individual fp32 ALUs; a true similar comparison would be number of warp schedulers or maximum number of warps resident at once, or e
114.
by jhj 10y ago
It's the memory size here that's likely important, 8 GTX 1080s have a third of the total global memory size as 8 K80s. And the aggregate memory bandwidth is way more important than flops, since almost certainly this workload isn&#
115.
by jhj 10y ago
Eh, the money is going in and eventually coming out over lengthy timescales. If active management means dipping into a smaller number of tradable assets, then the effect would only be worse. If they have to split it between a large number o
116.
by jhj 10y ago
im2col + sgemm on CPU as in Caffe for instance is really slow; you are heavily penalized for extra memory traffic and the sgemm tile sizes are probably not well tuned for the problem size at hand. At the roofline of performance, the differe
117.
by jhj 10y ago
> they speed up both training and evaluation by a factor of 100 The factor is more like 5-10x for truly optimized CPU versus GPU in fp32 (e.g., nnpack on CPU versus cuDNN).
118.
by jhj 10y ago
The numbers for GPU databases look “good” because you can get pretty high cross-sectional bandwidth to a reasonably large memory from 8 GPUs in one box, and advertise blazing speed from that. But it’s just a trick. The only thing that matte
119.
by jhj 10y ago
I don't think that everyone agrees on that in NYC, especially not the City Council or any of the people in local government with the power to change it. Much of the city is still single-family housing or multiple occupancy dwellings ty
120.
by jhj 10y ago
It depends upon the op / byte loaded intensity. Nvidia packs their GPUs with a lot of float32 (or float64) units because some problems (e.g., convolution, or more typical HPC problems like PDEs, which will probably be done in float64)

More ›