Building a cheap and second hand FP64 Monster -- Gridcoin

View this thread on: d.buzz | hive.blog | peakd.com | ecency.com
·@jedigeiss·
0.000 HBD
Building a cheap and second hand FP64 Monster -- Gridcoin
![Gridcoin](https://image.ibb.co/ikzPGa/DQm_Nse_JCMZiju_MSo_Qchf1_Pm6o3_Sqf8wxu_H3_Nu4uvue6c1pd_1680x8400.png)

## Hi all, ##  
<br>

Have you ever wanted to build a __very capable computer__ to __support science__ that is on the other hand side quite __cheap to buy__ and __re uses hardware__ that has been on the market for a long time ? There is a __good opportunity__ out there for exactly doing that and additionally you can even __earn gridcoins__ for that computation in exchange. 
In this __series of articles__ i will exactly do that, from the __basics__ and the __shopping list__, the __building and installing__ on to the __crunching__ :)
<br>
But lets start with the __basics__ first.

## The Basics ##

The key to all this is called __FP64__. GPUs are generally __very good at doing math__, the precision within that mathematics can be set by the needs of the programmer. FP64 hereby specifies double precision, FP32 being the single and FP16 half precision.<br> __Why is this important ?__<br>

![ FP64](https://image.ibb.co/hLKkmv/Single_precision_and_double_precision_numbers.png)

__Game development__ is more about __FP32__ and nowadays also a lot about __FP16__ calculations, therefore the designers of __graphic cards__ (for the consumer market) are more into __optimizing this__ than the FP64 power. <br>The __outcome__ of this can be seen in the following table: <br> 
(GFlops means Giga Floating Point Operations per Second = 10^9 Flops, the more the better )

| Graphic card | Chipset | FP64 GFlops| FP32 GFlops | Year
---|---|---|---|---
| Radeon 6970| Cayman XT| 675| 2703|2010
| Radeon 7870 Ghz Edtion | Pitcairn XT | 160 | 2560| 2012
| Radeon 7870 XT| Tahiti LE| 749| 2995| 2012
| Radeon 7950| Tahiti PRO2/PRO|761|3046|2012
| Radeon 7970| Tahiti XT|947|3788|2012
| Radeon 7970 Ghz Edition| Tahiti XT2|1075|4300|2012
| Radeon 7990| New Zealand|1894|8200|2013
|Radeon R9 270x| Pitcairn XT| 160 | 2560| 2013
|Radeon R9 390| Hawaii Pro| 640 | 	5120| 2015
|Radeon R9 Fury X |Fiji XT| 537| 8601| 2015
| Radeon RX 480|Ellesmere XT|323|5161| 2016
| RX Vega 56| Vega 10|659|10544|2017
|RX Vega 64|Vega 10| 792| 12665| 2017
|Nvidia GTX1050ti|GP107-400-A1| 66| 2123| 2016
| Nvidia GTX1060| GP106-400-A1|137|4372|2016
|Nvidia GTX1070|GP104-200-A1| 202| 6463| 2016
|Nvidia GTX1080ti|GP102-350-K1-A1| 354| 11340| 2017



There are some __very remarkable things__ to be seen here. First of __old__ Radeon Cards like the __7970__ or even the __7870 XT__ are __superior__ to even the __GTX 1080ti__ in terms of __FP64__. Even the __6970__ with the Cayman XT Chip is almost __doubling the FP64__ performance of a __1080ti__. What also can be seen that in terms of performance increase over the years there was a lot movement, just not in FP64. The AMD cards are no better, the brand new __Vega 64__ is inferior to an old __7970__. <br>
__What does that mean ? Are those cards really faster ?__<br> Well there is no short answer to that, but in this __specific field__ the answer is __clearly yes__. In a lot of other topics like memory bandwith, FP32 calculations, in short in all relevant things __for gaming__ the answer is clearly __no__.<br>
__Ok but for what do we need this FP64 power ?__<br>
This is where the fun starts, there are __scientific projects__ that require this double precision calculations. __Milkyway@home__ for instance is such a project. This project uses the FP64 calculations to create a __highly accurate three dimensional__ model of the __Milky Way galaxy__.


## The shopping List ##

This all got me thinking and i came up with a shopping list for a 4 GPU crunching machine (should yield somewhere >150 mag easily). 
![](https://image.ibb.co/hmYxYa/FP64_pieces.jpg)


This is just an idea, for sure the hardware can differ if you want to build such a thing for yourself. I took the __Asus Board__ because i already have a few of them running and they are __solid as a rock__ :) and on the other hand brings enough PCIe ports to connect the GPUs to. Iam not yet fixed in what GPUs i want to have, 7970 would be fine but 7950 or 7870XT would also do. Only in the __PSU__ i will make __no compromise__, i want to have a stable new one, i wont trust in something used. <br>
I will try __Ebay__ and other sources of old hardware, lets see what i can get there. Maybe there is also a __price drop__ with the new __Vega Cards out__, altough i dont really believe in that, there are just so few shipped ...


## Shopping ##

I will start the shopping now, once it is concluded iam going to follow this up with the next chapter :)



----

If you want to know more about Gridcoin please take a look here: 

__Gridcoin:__ [Gridcoin Website](http://www.gridcoin.us/)

If you want to know more about FP64 crunching in specific, please look into the posts from @vortac. There are quite a few and they are brilliant.<br> 
__Example on Vortacs posts for GPUs:__ [GPU Mining 6](https://steemit.com/gridcoin/@vortac/gridcoin-gpu-mining-6-obtaining-the-maximum-performance-out-of-your-gpus)


----------

Content credit:
Gridcoin Header, @joshoeah
👍 , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , ,