Chinese company announces new GPU that can reach 37 TFLOPS

ixfd64 · 2021-04-04, 20:40

https://wccftech.com/china-readies-i...les-amd-nvidia

The bad news is that there is no mention of it supporting OpenCL. So it probably can't be used for GIMPS work out of the box.

ATH · 2021-04-04, 21:34

It is also 37 TFLOPS at FP32 which is for trial factoring, there is no mention of FP64 speed.

Trial factoring is ok too I guess, but there are just so many good FP32 cards compared to FP64 that it is not very exciting.

kriesel · 2021-04-04, 22:17

More on "Big Island" https://coreteks.tech/articles/index...ture-overview/
Memory bandwidth 20% higher than Radeon VII, memory capacity double which would be handy for P-1 on large exponents; no mention of DP/FP64 performance in any coverage I found, or any hint of pricing. No mention of APIs such as OpenCL or OpenGL
"The company did not provide double-precision performance figures." So presumably nothing to brag about there, if it even has it. Although with the "quality" of reporting, one wonders. (More than one site referred to INT32, INT16, and INT8 as floating point choices. There seems to have been some substantial copy/paste going on.)
Some of the pages discussing the announcement are dated mid January.
https://videocardz.com/newz/tianshu-...32-performance
Packaging design indicates passive cooling lengthwise, like a lot of Teslas, dependent on forced air through a server chassis. And a mezzanine form factor is offered too.

Neutron3529 · 2021-04-10, 14:06

Quote:

Originally Posted by kriesel

More on "Big Island"

For Chinese users or who knows Chinese, the name of the company is:

上海天数智芯

with website (in Chinese):www.iluvatar.com.cn

People who want to search the original messages(with or without google translation) may using this word to begin the search.

(I cannot believe that I know this Chinese news in a foreign forum.)

The information in Chinese is as confused as it was in English.
Some people says, Big Island(BI) is compatible to a cuda program, but some people says that is definitely impossible.

According to messages in the official website, BI is compatible to Tensorflow and Pytorch (I don't find their version, maybe that's something like ROCm)

ATH · 2021-04-10, 14:33

Quote:

Originally Posted by Neutron3529

with website (in Chinese):www.iluvatar.com.cn

Ilúvatar "Allfather" also called Eru "The One".

2021-04-04, 20:40	#1
ixfd64 Bemusing Prompter "Danny" Dec 2002 California 2³×313 Posts	Chinese company announces new GPU that can reach 37 TFLOPS https://wccftech.com/china-readies-i...les-amd-nvidia The bad news is that there is no mention of it supporting OpenCL. So it probably can't be used for GIMPS work out of the box.

2021-04-04, 21:34	#2
ATH Einyen Dec 2003 Denmark 110101111100₂ Posts	It is also 37 TFLOPS at FP32 which is for trial factoring, there is no mention of FP64 speed. Trial factoring is ok too I guess, but there are just so many good FP32 cards compared to FP64 that it is not very exciting. Last fiddled with by ATH on 2021-04-04 at 21:35

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
How long will it be until MM61 is within practical reach of PRP/primality testing?	Stargate38	Operazione Doppi Mersennes	14	2020-01-29 20:35
Asking permission of the company	sixblueboxes	Information & Answers	32	2015-12-10 19:36
Does the photon ever reach the spaceship?	retina	Puzzles	46	2011-02-05 11:26
Amazon is a greedy bastard of a company.	jasong	jasong	14	2007-12-13 21:02
math puzzles by Dell (puzzle company)....	ixfd64	Puzzles	0	2004-03-16 00:25

2021-04-04, 22:17	#3
kriesel "TF79LL86GIMPS96gpu17" Mar 2017 US midwest 2⁴·3·163 Posts	More on "Big Island" https://coreteks.tech/articles/index...ture-overview/ Memory bandwidth 20% higher than Radeon VII, memory capacity double which would be handy for P-1 on large exponents; no mention of DP/FP64 performance in any coverage I found, or any hint of pricing. No mention of APIs such as OpenCL or OpenGL "The company did not provide double-precision performance figures." So presumably nothing to brag about there, if it even has it. Although with the "quality" of reporting, one wonders. (More than one site referred to INT32, INT16, and INT8 as floating point choices. There seems to have been some substantial copy/paste going on.) Some of the pages discussing the announcement are dated mid January. https://videocardz.com/newz/tianshu-...32-performance Packaging design indicates passive cooling lengthwise, like a lot of Teslas, dependent on forced air through a server chassis. And a mezzanine form factor is offered too.