1 |
On Mon, Feb 15, 2021 at 3:17 AM Walter Dnes <waltdnes@××××××××.org> wrote: |
2 |
> |
3 |
> On Sun, Feb 14, 2021 at 06:09:58PM -0700, Grant Taylor wrote |
4 |
> > On 2/14/21 10:51 AM, Jack wrote: |
5 |
> > > I don't think you can completely get rid of it. |
6 |
> > |
7 |
> > My (long term) desire is to do away with /lib32 and /lib64, ultimately |
8 |
> > only using /lib. Likewise for the other library directories in /usr or |
9 |
> > wherever they are. I don't see a need for the specific bit variants in |
10 |
> > the future. |
11 |
> |
12 |
> How long before we see /lib and /lib64 *AND* /lib128 ? |
13 |
|
14 |
Well, anything is possible, but it seems unlikely. If it happens soon |
15 |
then chances are that multilib will still be a thing and so less stuff |
16 |
will break than when amd64 was introduced. If it happens in a century |
17 |
when we're all running no-multilib then we'll be reinventing the |
18 |
wheel. |
19 |
|
20 |
The main things that drove amd64 though were: |
21 |
* increasing the number of registers available |
22 |
* allowing direct access to >4GB of RAM (or a fraction of this |
23 |
depending on the OS design) |
24 |
|
25 |
I suspect the first is less of a concern these days - compilers |
26 |
generally only need so many registers and when instructions are added |
27 |
that need more register space they tend to come with registers to |
28 |
accommodate them. The second will be a concern when exabyte-scale |
29 |
data structures are common to work with. Note that current processors |
30 |
generally can't handle this much address space, but the amd64 |
31 |
instruction set itself can (I think), so the CPUs can continue to |
32 |
scale up. RAM capacity doesn't really seem to be increasing in recent |
33 |
years - I'm not sure if that is more market-driven or a technological |
34 |
limitation. RAM speed has improved somewhat, especially in niches |
35 |
like GPUs. Computers with 1GB of RAM were a thing in Y2K and today it |
36 |
is pretty uncommon for a standard desktop to have more than 8GB, and |
37 |
if you want to even cram more than about 128GB into a motherboard you |
38 |
start needing more enterprise-grade hardware. That isn't a very large |
39 |
increase in 20 years - doubling every 3 years (in terms of max |
40 |
capacity). We're using 37 bits today (on desktops), so at 3 years per |
41 |
bit that is another 80 years until we exhaust 64 bits, assuming that |
42 |
we continue to grow exponentially at the same rate. Though you do |
43 |
have to think about what use cases actually need that kind of working |
44 |
set. At 64-bit depth 300dpi 3D graphics would require 200MB/in^3, If |
45 |
you had a house-sized VR space (20k ft^3) rendered at that detail |
46 |
you'd need 7TB of RAM to store a frame of video, which is still only |
47 |
50 bits. Maybe if you want a holodeck that 1000 people can play |
48 |
around in at once you'd run into the 64-bit limit (of course you'd |
49 |
have a ton of IO issues to fix long before then). |
50 |
|
51 |
So, that makes me wonder what the practical requirements are in order |
52 |
to implement The Matrix. :) Of course, if you're sticking people in |
53 |
it maybe you can borrow some of their own memory capacity and |
54 |
processing abilities to drive it. Kind of makes you wonder why you'd |
55 |
even need the human brains in the first place if you're able to deal |
56 |
with that kind of data in a simulation... |
57 |
|
58 |
-- |
59 |
Rich |