Re: [COOT] Hardware recommendations

[mesters]

Hi Qiuye,

2x 12GB RTX-3060 will - for example - fit well with sufficient space 
between cards on the MSI MEG *X570S* Unify-X Max (with two GPU slots at 
16x/8x speed) and the Gigabyte X570S Aorus Master (with three slots at 
8x/8x/4x if all are populated), but as you correctly noted, you will 
need a really big case with plenty of fans to remove the heat such as an 
Corsair 7000D Airflowor or the be quiet! silent base 802 or, for the 
brave, build a water-cooled rig..


All cheaper X570S and all of the X570 boards do not support 16x/8x or 
8x/8x speed (at least I did not find one).


Best,

Jeroen

Am 19.04.22 um 17:41 schrieb Qiuye Li:

Hi Jeroen,
Thanks for the kind suggestions, they definitely help me as well as 
many others a lot!
I recently tested coot0.9 on a 11800H+64G+A2000+1600p mobile 
workstation and am quite OK with the performance. I guess the same gen 
desktop-level hardware and configuration--something around 
5950X+64G+3060 with my existing 27' 2k display--should do the work 
well, and can be picked up today in a local store. We already have 
more than a dozen consumer cards in our workstations so we'll probably 
worry about NVIDIA's policy later. The idea of a dual-GPU machine 
sounds wonderful, I need to dig a bit more myself to see if there 
are decent chassis and board to fit those gamer cards' little hot 
fans, and still be reasonably quiet in an office.


Best,
Qiuye

On Tue, Apr 19, 2022 at 7:47 AM mesters 
 wrote:


Hi Qiuye,

somehow the one-million-dollar/"budget vs. program specs vs.
current/future applications//vs. access to central computing
facilities/" question, is it not?

As Paul indicated, for a personal/office graphics workstation
(maximally 128GB ram memory), a 8-16 core CPU with a passmark
 of >27.500 and
for <600 US$ will do (intel i9-12900, i7-12700, i5-12600k; Ryzen
5950x, 5900x or 5800x). As heat dissipation is a major problem,
AMD is a good choice for the moment (i9-12900k TDP 241 watts vs.
3950x TDP 105 watts). If there is a reasonable chance, the system
is to be upgraded in the future to a small/medium 2x GPU CryoSpark
workstation, plan for a 16-core CPU from the beginning and make
sure to choose a motherboard that offers the possibility of 2x
PCIe 4.0 at *8x/8x* speed mode (MSI MPG Z690 FORCE / Gygabyte
X570*S* Aorus PRO) as many standard boards only offer 16x/4x speed
mode (read the technical description carefully; most if not all
boards <300 US$ offer 16x/4x only).

As for the graphics card, it depends mainly on program specs,
structure-model-size / e.d.-map-size and screen resolution.

For a 4K screen in combination with rendering, you will indeed
require a >4000 cuda-cores GPU card such as a 8GB RTX3060ti (600
€; 4864 cuda-cores; 200 watt and sadly no FE model available) or a
16GB RTX A4000 (1100 €; 6144 cuda-cores; 140 watt). Again, as heat
dissipation is a major problem, the A4000 is a very good pick with
16GB memory! I recommend a professional RTX *A*xx00-series over
the wide-spread 3x00-series consumer cards (*FE* models if at
all!) as for the moment, Nvidia still allows the installation of
the CUDA-package (AlphaFold, etc., etc.) on consumer cards outside
of any central university computing facility without restrictions
(read the small print in the cuda software documention), but I
would not be surprized this might change in the future (recall the
LHR measures taken by the company)! Nvidia can tell from the
IP-adress who-is-who (office, home or central computing facility).

For a 1440p screen and small- to mid-sized models/maps, an A2000
(3328 cuda-cores, 30% more cores than a 2070 super; 650-750€; 70
watt) with 6 or 12GB GDDR is good although not a bargain compared
to the A4000 that offers a much better price/performance ratio.

Paul's dream-machine - an 3990X combined with a Nvidia RTX 3090 Ti
- will cost you, build-it-yourself, at least 8.000 €, which is
just too expensive in my eyes.

A build-it-yourself machine around a Ryzen 5800x, 12GB RTX A2000,
2TB M.2, 12TB HDD and 32GB ram including power-supply and case
will cost about 2600 €, and with a Ryzen 5950x and 16GB RTX A4000
about 3100 €.

A minimal system for about 1100€ can be build around a Ryzen 5600x
or intel i5-12500 (passmark > 20.000), 16Gb ram and a 6GB GTX 1660
super (1408 cuda-cores; max. 1440p resolution; we run AlphaFold v2
on such a card, not superfast but it gets the job done), for
smaller models/maps of course and with slower Coot1 fancy mode I
guess 

It all boiles down to budget in the end I guess, good luck at
taking the right decision!

Jeroen


Am 15.04.22 um 17:32 schrieb Qiuye Li:

Dear all,
We recently centralized our high-end workstations and no longer
have easy access to their 

Re: [COOT] Hardware recommendations

[Qiuye Li]

Hi Jeroen,
Thanks for the kind suggestions, they definitely help me as well as many
others a lot!
I recently tested coot0.9 on a 11800H+64G+A2000+1600p mobile workstation
and am quite OK with the performance. I guess the same gen desktop-level
hardware and configuration--something around 5950X+64G+3060 with my
existing 27' 2k display--should do the work well, and can be picked up
today in a local store. We already have more than a dozen consumer cards in
our workstations so we'll probably worry about NVIDIA's policy later. The
idea of a dual-GPU machine sounds wonderful, I need to dig a bit more
myself to see if there are decent chassis and board to fit those gamer
cards' little hot fans, and still be reasonably quiet in an office.

Best,
Qiuye

On Tue, Apr 19, 2022 at 7:47 AM mesters 
wrote:

> Hi Qiuye,
>
> somehow the one-million-dollar* "budget vs. program specs vs.
> current/future applications** vs. access to central computing facilities*"
> question, is it not?
>
> As Paul indicated, for a personal/office graphics workstation (maximally
> 128GB ram memory), a 8-16 core CPU with a passmark
>  of >27.500 and for <600
> US$ will do (intel i9-12900, i7-12700, i5-12600k; Ryzen 5950x, 5900x or
> 5800x). As heat dissipation is a major problem, AMD is a good choice for
> the moment (i9-12900k TDP 241 watts vs. 3950x TDP 105 watts). If there is a
> reasonable chance, the system is to be upgraded in the future to a
> small/medium 2x GPU CryoSpark workstation, plan for a 16-core CPU from the
> beginning and make sure to choose a motherboard that offers the possibility
> of 2x PCIe 4.0 at *8x/8x* speed mode (MSI MPG Z690 FORCE / Gygabyte X570
> *S* Aorus PRO) as many standard boards only offer 16x/4x speed mode (read
> the technical description carefully; most if not all boards <300 US$ offer
> 16x/4x only).
>
> As for the graphics card, it depends mainly on program specs,
> structure-model-size / e.d.-map-size and screen resolution.
>
> For a 4K screen in combination with rendering, you will indeed require a
> >4000 cuda-cores GPU card such as a 8GB RTX3060ti (600 €; 4864 cuda-cores;
> 200 watt and sadly no FE model available) or a 16GB RTX A4000 (1100 €; 6144
> cuda-cores; 140 watt). Again, as heat dissipation is a major problem, the
> A4000 is a very good pick with 16GB memory! I recommend a professional RTX
> *A*xx00-series over the wide-spread 3x00-series consumer cards (*FE*
> models if at all!) as for the moment, Nvidia still allows the installation
> of the CUDA-package (AlphaFold, etc., etc.) on consumer cards outside of
> any central university computing facility without restrictions (read the
> small print in the cuda software documention), but I would not be surprized
> this might change in the future (recall the LHR measures taken by the
> company)! Nvidia can tell from the IP-adress who-is-who (office, home or
> central computing facility).
>
> For a 1440p screen and small- to mid-sized models/maps, an A2000 (3328
> cuda-cores, 30% more cores than a 2070 super; 650-750€; 70 watt) with 6 or
> 12GB GDDR is good although not a bargain compared to the A4000 that offers
> a much better price/performance ratio.
>
> Paul's dream-machine - an 3990X combined with a Nvidia RTX 3090 Ti - will
> cost you, build-it-yourself, at least 8.000 €, which is just too expensive
> in my eyes.
>
> A build-it-yourself machine around a Ryzen 5800x, 12GB RTX A2000, 2TB M.2,
> 12TB HDD and 32GB ram including power-supply and case will cost about 2600
> €, and with a Ryzen 5950x and 16GB RTX A4000 about 3100 €.
>
> A minimal system for about 1100€ can be build around a Ryzen 5600x or
> intel i5-12500 (passmark > 20.000), 16Gb ram and a 6GB GTX 1660 super (1408
> cuda-cores; max. 1440p resolution; we run AlphaFold v2 on such a card, not
> superfast but it gets the job done), for smaller models/maps of course and
> with slower Coot1 fancy mode I guess 
>
> It all boiles down to budget in the end I guess, good luck at taking the
> right decision!
>
> Jeroen
>
>
> Am 15.04.22 um 17:32 schrieb Qiuye Li:
>
> Dear all,
> We recently centralized our high-end workstations and no longer have easy
> access to their displays, and thus need a modeling computer. I found coot
> 0.9/wincoot 0.9 are a bit laggy on our less frequently updated office
> computers, and the all-new coot 1 probably requires even more hardware
> resources? With this recent launch of coot 1, it is probably a good time to
> kindly ask for hardware suggestions, like CPU, RAM, SSD, and dGPU.
> With coot, I typically visualize a .mrc map using a radius of ~25 A and
> work with real-space refinement/other simple manual adjustments. Other than
> coot, I often have  1-2 active chimera/Phenix windows open, a few SSH
> tunnels, and some web pages. I'll be happy to provide more details if
> needed. Any suggestions are appreciated!
>
> Best,
> Qiuye
>
> --
>
> To unsubscribe from the COOT 

Re: [COOT] Hardware recommendations

[mesters]

Hi Qiuye,

somehow the one-million-dollar/"budget vs. program specs vs. 
current/future applications//vs. access to central computing 
facilities/" question, is it not?


As Paul indicated, for a personal/office graphics workstation (maximally 
128GB ram memory), a 8-16 core CPU with a passmark 
 of >27.500 and for 
<600 US$ will do (intel i9-12900, i7-12700, i5-12600k; Ryzen 5950x, 
5900x or 5800x). As heat dissipation is a major problem, AMD is a good 
choice for the moment (i9-12900k TDP 241 watts vs. 3950x TDP 105 watts). 
If there is a reasonable chance, the system is to be upgraded in the 
future to a small/medium 2x GPU CryoSpark workstation, plan for a 
16-core CPU from the beginning and make sure to choose a motherboard 
that offers the possibility of 2x PCIe 4.0 at *8x/8x* speed mode (MSI 
MPG Z690 FORCE / Gygabyte X570*S* Aorus PRO) as many standard boards 
only offer 16x/4x speed mode (read the technical description carefully; 
most if not all boards <300 US$ offer 16x/4x only).


As for the graphics card, it depends mainly on program specs, 
structure-model-size / e.d.-map-size and screen resolution.


For a 4K screen in combination with rendering, you will indeed require a 
>4000 cuda-cores GPU card such as a 8GB RTX3060ti (600 €; 4864 
cuda-cores; 200 watt and sadly no FE model available) or a 16GB RTX 
A4000 (1100 €; 6144 cuda-cores; 140 watt). Again, as heat dissipation is 
a major problem, the A4000 is a very good pick with 16GB memory! I 
recommend a professional RTX *A*xx00-series over the wide-spread 
3x00-series consumer cards (*FE* models if at all!) as for the moment, 
Nvidia still allows the installation of the CUDA-package (AlphaFold, 
etc., etc.) on consumer cards outside of any central university 
computing facility without restrictions (read the small print in the 
cuda software documention), but I would not be surprized this might 
change in the future (recall the LHR measures taken by the company)! 
Nvidia can tell from the IP-adress who-is-who (office, home or central 
computing facility).


For a 1440p screen and small- to mid-sized models/maps, an A2000 (3328 
cuda-cores, 30% more cores than a 2070 super; 650-750€; 70 watt) with 6 
or 12GB GDDR is good although not a bargain compared to the A4000 that 
offers a much better price/performance ratio.


Paul's dream-machine - an 3990X combined with a Nvidia RTX 3090 Ti - 
will cost you, build-it-yourself, at least 8.000 €, which is just too 
expensive in my eyes.


A build-it-yourself machine around a Ryzen 5800x, 12GB RTX A2000, 2TB 
M.2, 12TB HDD and 32GB ram including power-supply and case will cost 
about 2600 €, and with a Ryzen 5950x and 16GB RTX A4000 about 3100 €.


A minimal system for about 1100€ can be build around a Ryzen 5600x or 
intel i5-12500 (passmark > 20.000), 16Gb ram and a 6GB GTX 1660 super 
(1408 cuda-cores; max. 1440p resolution; we run AlphaFold v2 on such a 
card, not superfast but it gets the job done), for smaller models/maps 
of course and with slower Coot1 fancy mode I guess 


It all boiles down to budget in the end I guess, good luck at taking the 
right decision!


Jeroen


Am 15.04.22 um 17:32 schrieb Qiuye Li:

Dear all,
We recently centralized our high-end workstations and no longer have 
easy access to their displays, and thus need a modeling computer. I 
found coot 0.9/wincoot 0.9 are a bit laggy on our less frequently 
updated office computers, and the all-new coot 1 probably requires 
even more hardware resources? With this recent launch of coot 1, it is 
probably a good time to kindly ask for hardware suggestions, like 
CPU, RAM, SSD, and dGPU.
With coot, I typically visualize a .mrc map using a radius of ~25 A 
and work with real-space refinement/other simple manual adjustments. 
Other than coot, I often have  1-2 active chimera/Phenix windows open, 
a few SSH tunnels, and some web pages. I'll be happy to provide more 
details if needed. Any suggestions are appreciated!


Best,
Qiuye



To unsubscribe from the COOT list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=COOT=1 





--
signature.html *Dr. /math. et dis. nat./ Jeroen R. Mesters
*Deputy, Lecturer, Program Coordinator Infection Biology 

Visiting Professorship (South Bohemian University 
) in Biophysics


*University of Lübeck*
Center for Structural and Cell Biology in Medicine*
Institute of Biochemistry*

Tel  +49 451 3101 3105 (Secretariate 3101)
Fax +49 451 3101 3104 *
*jeroen.mest...@uni-luebeck.de
https://www.biochem.uni-luebeck.de
https://orcid.org/-0001-8532-6699

*Ratzeburger Allee 160
23538 Lübeck, Schleswig-Holstein
Germany*

Re: [COOT] Hardware recommendations

[mesters]

My two cents, if one has access to centralized computing power i.e. CPUs 
and GPUs for doing the heavy calculations, one will hardly need a 
64-core threadripper 3990X (4300 €) in the office especially since most 
programs profit from quality (speed), not quantity (total number of 
cores). A 3960x will do just fine and will provide 75% computing power 
of an 3990x (both are rated at 280 wattswhich sets the total performance 
level for the major part) at just 33% of the costs. Extra money is much 
better spent on other components such as sufficient ram, M.2 SSDs, 
sufficient HDDs, a professional CUDA card such as the 20GB, 7168 
CUDA-core RTX 4500 (2300 €) and most of all, *two* (!) 1440p displays



Happy easter wishes (holidays) to everybody,


Jeroen


Am 17.04.22 um 01:38 schrieb Paul Emsley:

On 15/04/2022 16:32, Qiuye Li wrote:

We recently centralized our high-end workstations and no longer have 
easy access to their displays,


Someone drew the short straw there :-(

I found coot 0.9/wincoot 0.9 are a bit laggy on our less frequently 
updated office computers,


You mean hardware updates? Yes, if you go back a gen or two, you will 
notice this.



and the all-new coot 1 probably requires even more hardware resources?


It was my intention that Coot 1 be faster than 0.9.x in Basic mode - 
Coot 1 is using the GPU in the way it was intended to be used - not 
transferring the map line by line from the CPU to the GPU each frame.


So, with my machine, in the coot spin test, with a sampling rate of 
2.5 and a radius of 99, Coot 0.9.8.1 spins at 5fps, and Coot 1 spins 
at 60fps for all render modes (limited by monitor sync).


Coot 1 has three render modes: Basic, Standard and Fancy. Basic mode 
draws models with lines (like Coot 0.x does), Standard mode uses 
(only) ambient, diffuse and specular lighting, Fancy mode adds 
brightness, gamma, depth blur, SSAO and shadows.


Fancy mode in 4K fullscreen with a ribosome drops the frame rate to 
less than 60 fps on my computer (RTX 2070 Super).


With this recent launch of coot 1, it is probably a good time to 
kindly ask for hardware suggestions, like CPU, RAM, SSD, and dGPU.
With coot, I typically visualize a .mrc map using a radius of ~25 A 
and work with real-space refinement/other simple manual adjustments. 
Other than coot, I often have 1-2 active chimera/Phenix windows open, 
a few SSH tunnels, and some web pages. [] Any suggestions are 
appreciated!


OK, so Coot 1 now takes more than 1 second to open in graphics mode on 
my computers (SSD and m.2 NVMe PCIe 4). Coot 0.9.x was 0.35s. I 
haven't tested it from a hard drive, but I can believe that Coot 1 
would take more than 10 seconds - that could be irritating. So m.2 
NVMe 4 or 5 it is.


Since 2017 any function/algorithm that I've written that takes more 
than 1 second to complete is now written as multi-threaded (and I have 
refactored several older algorithms  to be multi-threaded also). So 
the more of and faster the cores the better. I only have access to 
workstation/Xeon processor at work - they are less speedy (and more 
expensive (reliable though)). Although I haven't tried them myself I'd 
recommend AMD Ryzen 5900X, 5950X or (the new) 5800X3D (the increased 
size of the L3 cache could improve contouring and FFT speed).


Process size: I've seen coot get up to more than 20Gb - typically when 
using map masking for cryo-EM maps. I have 32 Gb RAM on my main 
machine and I haven't see Coot crash due to lack of memory. So I'd 
recommend at least that especially if you have a multi-tabbed browser 
and other graphics going also.


If you have a 4K monitor (if you don't have one, you should get one) 
and want fullscreen 4K Fancy mode goodness (and to make pretty 
screenshots) you'd want something better than my card, so the 3000 
series, RTX 3070 or 3080. I don't follow AMD graphics cards sad to 
say. I intend to make the interface to Blender increasingly powerful 
and easy to use - that also will benefit from big memory and a fast 
graphics card.


So, get yourself a ballin gamer rig, and you'll have a ballin Coot 
workstation :-)


The Best Coot Workstation today would be build around the AMD 
Threadripper 3990X and Nvidia RTX 3090 Ti card and Dell UltraSharp 
U4320Q monitor.


Paul.



To unsubscribe from the COOT list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=COOT=1

This message was issued to members of www.jiscmail.ac.uk/COOT, a 
mailing list hosted by www.jiscmail.ac.uk, terms & conditions are 
available at https://www.jiscmail.ac.uk/policyandsecurity/



--
signature.html *Dr. /math. et dis. nat./ Jeroen R. Mesters
*Deputy, Lecturer, Program Coordinator Infection Biology 

Visiting Professorship (South Bohemian University 
) in Biophysics


*University of 

Re: [COOT] Hardware recommendations

[Paul Emsley]

On 15/04/2022 16:32, Qiuye Li wrote:

We recently centralized our high-end workstations and no longer have 
easy access to their displays,


Someone drew the short straw there :-(

I found coot 0.9/wincoot 0.9 are a bit laggy on our less frequently 
updated office computers,


You mean hardware updates? Yes, if you go back a gen or two, you will 
notice this.



and the all-new coot 1 probably requires even more hardware resources?


It was my intention that Coot 1 be faster than 0.9.x in Basic mode - 
Coot 1 is using the GPU in the way it was intended to be used - not 
transferring the map line by line from the CPU to the GPU each frame.


So, with my machine, in the coot spin test, with a sampling rate of 2.5 
and a radius of 99, Coot 0.9.8.1 spins at 5fps, and Coot 1 spins at 
60fps for all render modes (limited by monitor sync).


Coot 1 has three render modes: Basic, Standard and Fancy. Basic mode 
draws models with lines (like Coot 0.x does), Standard mode uses (only) 
ambient, diffuse and specular lighting, Fancy mode adds brightness, 
gamma, depth blur, SSAO and shadows.


Fancy mode in 4K fullscreen with a ribosome drops the frame rate to less 
than 60 fps on my computer (RTX 2070 Super).


With this recent launch of coot 1, it is probably a good time to 
kindly ask for hardware suggestions, like CPU, RAM, SSD, and dGPU.
With coot, I typically visualize a .mrc map using a radius of ~25 A 
and work with real-space refinement/other simple manual adjustments. 
Other than coot, I often have 1-2 active chimera/Phenix windows open, 
a few SSH tunnels, and some web pages. [] Any suggestions are appreciated!


OK, so Coot 1 now takes more than 1 second to open in graphics mode on 
my computers (SSD and m.2 NVMe PCIe 4). Coot 0.9.x was 0.35s. I haven't 
tested it from a hard drive, but I can believe that Coot 1 would take 
more than 10 seconds - that could be irritating. So m.2 NVMe 4 or 5 it is.


Since 2017 any function/algorithm that I've written that takes more than 
1 second to complete is now written as multi-threaded (and I have 
refactored several older algorithms  to be multi-threaded also). So the 
more of and faster the cores the better. I only have access to 
workstation/Xeon processor at work - they are less speedy (and more 
expensive (reliable though)). Although I haven't tried them myself I'd 
recommend AMD Ryzen 5900X, 5950X or (the new) 5800X3D (the increased 
size of the L3 cache could improve contouring and FFT speed).


Process size: I've seen coot get up to more than 20Gb - typically when 
using map masking for cryo-EM maps. I have 32 Gb RAM on my main machine 
and I haven't see Coot crash due to lack of memory. So I'd recommend at 
least that especially if you have a multi-tabbed browser and other 
graphics going also.


If you have a 4K monitor (if you don't have one, you should get one) and 
want fullscreen 4K Fancy mode goodness (and to make pretty screenshots) 
you'd want something better than my card, so the 3000 series, RTX 3070 
or 3080. I don't follow AMD graphics cards sad to say. I intend to make 
the interface to Blender increasingly powerful and easy to use - that 
also will benefit from big memory and a fast graphics card.


So, get yourself a ballin gamer rig, and you'll have a ballin Coot 
workstation :-)


The Best Coot Workstation today would be build around the AMD 
Threadripper 3990X and Nvidia RTX 3090 Ti card and Dell UltraSharp 
U4320Q monitor.


Paul.



To unsubscribe from the COOT list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=COOT=1

This message was issued to members of www.jiscmail.ac.uk/COOT, a mailing list 
hosted by www.jiscmail.ac.uk, terms & conditions are available at 
https://www.jiscmail.ac.uk/policyandsecurity/

[COOT] Hardware recommendations

[Qiuye Li]

Dear all,
We recently centralized our high-end workstations and no longer have easy
access to their displays, and thus need a modeling computer. I found coot
0.9/wincoot 0.9 are a bit laggy on our less frequently updated office
computers, and the all-new coot 1 probably requires even more hardware
resources? With this recent launch of coot 1, it is probably a good time to
kindly ask for hardware suggestions, like CPU, RAM, SSD, and dGPU.
With coot, I typically visualize a .mrc map using a radius of ~25 A and
work with real-space refinement/other simple manual adjustments. Other than
coot, I often have  1-2 active chimera/Phenix windows open, a few SSH
tunnels, and some web pages. I'll be happy to provide more details if
needed. Any suggestions are appreciated!

Best,
Qiuye



To unsubscribe from the COOT list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=COOT=1

This message was issued to members of www.jiscmail.ac.uk/COOT, a mailing list 
hosted by www.jiscmail.ac.uk, terms & conditions are available at 
https://www.jiscmail.ac.uk/policyandsecurity/