RSA-250

[R.D. Silverman]

Originally Posted by R.D. Silverman

Yes, indeed. This raises additional problems for a distributed implementation because one then needs different
software/compiled versions for each (type of) client.

The RSA-240 paper reports:

"After running some experiments, we found that nodes with at least 4 GB of RAM per physical core could smoothly
and efficiently accommodate the following setting:..."

RSA-1024 will require a lot more than 4G/core. And they did use shared memory. AFAIK Greg limits his clients to 1.25G/core.

I just bought a new toy: An i9 with 24 cores and 128G of DDR5. Even it can't handle 5G/core computations.

I realized all of this back in the late 90's when I tried to convince people that MEMORY would be the binding
constraint for RSA-1024. Everyone was focused on computing time. And the machines of that time were single
core 32-bit. They lack the needed address space. I predicted back then that RSA-1024 would not be done
before 2020. I won the bet with xilman. It still seems that it won't be done anytime soon.

Any bets for 2030???

A bit more info....

charybdis reported that RSA-1024 would be

"RSA-1024 is very roughly 500 times more difficult"

Memory requirements, when optimized, scale with the square root of the run time. CADO required
4G/core for RSA-240. 4G/core x sqrt(500) ~ 100 G/core for RSA-1024. Can you say "yech"??

I think this is an overestimate. (maybe by a factor of 2? Who knows?) but ask how many machines have 50Gbytes/core...

Or even 32G/core???

[kruoli]

Regarding memory access patterns locality: would it be feasible to use RAM combined with fast NVMe drives in a hierarchial memory system?

[retina]

Originally Posted by R.D. Silverman

Or even 32G/core???

These exist.

A random link:
https://www.tweaktown.com/news/83600...ram/index.html

4TB/128core = 32G/core

[VBCurtis]

Originally Posted by R.D. Silverman

I just bought a new toy: An i9 with 24 cores and 128G of DDR5. Even it can't handle 5G/core computations.
{.....}

Any bets for 2030???

You can surely handle 5G/core computations, by choosing to use fewer than all 24 cores. If some job needs 6GB/core, then use 20 cores. Heck, that i9 might reach memory bandwidth saturation running 16 cores, and the last 8 are of little speed benefit anyway... or choose a computation method that trades a bit of speed for a bit of memory use, as CADO las already does on memory-constrained machines. When we ran a team CADO job with I=16 siever area, clients with less than 16GB ran on 2/3 the memory footprint as those with 16+GB; the client did this automatically. Slower clients are better than no clients.

On the Speculation side:
I think RSA-896 within 5 years is a favorite to happen, and 1024 in ten years. 2030 seems soon for RSA-1024, it would be a mild surprise.

[R.D. Silverman]

Originally Posted by VBCurtis

You can surely handle 5G/core computations, by choosing to use fewer than all 24 cores.

Well, of course. But the objective is to use all 24. (I give myself an NSS award)

If some job needs 6GB/core, then use 20 cores.

The OS takes at least 10% of memory.... But that is a quibble.

One needs as many cores as one can get for RSA-1024.
(trivially; this observation gets another NSS award)

On the Speculation side:
I think RSA-896 within 5 years is a favorite to happen, and 1024 in ten years. 2030 seems soon for RSA-1024, it would be a mild surprise.

We agree.

[R.D. Silverman]

Originally Posted by retina

These exist.

A random link:
https://www.tweaktown.com/news/83600...ram/index.html

4TB/128core = 32G/core

Price? I did not see one on the link. RSA-1024 would require about 10,000 of them.

I doubt that individuals could afford one. I might pay $10K for such a machine, but I am a small minority.

Also, could a standard U.S. home electrical outlet handle the wattage?

[Uncwilly]

Andrew really has no idea who he is talking to.

[R.D. Silverman]

Originally Posted by R.D. Silverman

Price? I did not see one on the link. RSA-1024 would require about 10,000 of them.

I doubt that individuals could afford one. I might pay $10K for such a machine, but I am a small minority.

Also, could a standard U.S. home electrical outlet handle the wattage?

Let me show my ignorance. I don't know anything about the Ryzen machines. Zilch. Nada. Bupkis.

What OS comes bundled with them? is VS2022 (or VS2019) available? Are they compiled code compatible
with Intel? What software tools are available? (OpenOffice? Adobe? Steam? Itunes? Chrome? DosBox? etc.
ad nauseum)

How do the cores compare with Intel cores vis-a-vis speed?

[xilman]

Originally Posted by R.D. Silverman

I realized all of this back in the late 90's when I tried to convince people that MEMORY would be the binding
constraint for RSA-1024. Everyone was focused on computing time. And the machines of that time were single
core 32-bit. They lack the needed address space. I predicted back then that RSA-1024 would not be done
before 2020. I won the bet with xilman.

Ah that's a good point. Now you have reminded me I remember that I never paid up. I still can't remember the stake or the terms. Something like GBP 20 or USD 20 whichever the more valuable at the end date, to be given to a charity of the winner's choice.

Does that sound familiar?

[retina]

Originally Posted by R.D. Silverman

Let me show my ignorance. I don't know anything about the Ryzen machines. Zilch. Nada. Bupkis.

What OS comes bundled with them? is VS2022 (or VS2019) available? Are they compiled code compatible
with Intel? What software tools are available? (OpenOffice? Adobe? Steam? Itunes? Chrome? DosBox? etc.
ad nauseum)

How do the cores compare with Intel cores vis-a-vis speed?

They are x86-64 cores. The CPU is rated <300W. The system consumption is likely <1kW, but check before you buy it might be 2kW depending on the RAM and how you clock it. OS and software are standard Linux/Win/Whatever.

The CPU alone is ~$5k. So twin CPUs is ~$10k.

[R.D. Silverman]

Originally Posted by xilman

Ah that's a good point. Now you have reminded me I remember that I never paid up. I still can't remember the stake or the terms. Something like GBP 20 or USD 20 whichever the more valuable at the end date, to be given to a charity of the winner's choice.

Does that sound familiar?

The money is the last thing on my mind. Bragging rights are sufficient. I did not mention the bet for the
purpose of getting money. Let's agree to drop it.