Huawei now take on Nvidia with her own super AI chip (2x faster than Nv's best)

[52051]

On 10 Oct 2018, Huawei launched her own AI platform and AI chips called Shengteng (Ascend) series.

The flagship of this AI series, Shengteng 910, can deliever over 256Tflops performance and 512Tiops (int8) at 350Watt, 2X better than Nvidia's best offering in terms of performance and performance per watt.

Huawei also launched her own complete toolchain and framework, including complier, core maths libraries and machine learning libraries and AI framework/cloud etc.

This is very serious stuff, and judging by Huawei's track record, Nvidia, and its stockholders should be extremely worried, to the world of AI computing hardware, it is like a shark come to a pond.

Source (in Chinese):
http://news.mydrivers.com/1/598/598312.htm

 

[52051]

With the aid of trade war, I bet Huawei will dominate the AI chip making world very soon.

There are already several tech start-up in AI chip making in China, but now this is the real deal, far better than all others, and Huawei do it very quietly, this is the Chinese style, quiet mouth and hard working, unlike the opposite in the other side of pacific

 

[Offshore]

Huawei really exceptional company.. they always exceeds My expectation.. damn.
No wonder the US hated huawei lol

 

[moweike]

Next, the US will attack HUAWEI.

 

[LeGenD]

What kind of benchmark is that? Too ambiguous.

Most powerful GPU from NVIDIA is GeForce RTX 2080 Ti.

 

[delta2]

Better than NVidia? Fake news!!!

 

[52051]

What kind of benchmark is that?

Most powerful GPU from NVIDIA is GeForce RTX 2080 Ti.

By raw computing power.

And no Nv's best GPU is by no means 2080Ti, their best computing chip is Tesla V100 in terms of raw flops power, which deliver 125T flops and 50T iops(int8) at 300W.

And their best chip for int8 computing is Quadro RTX8000, deliever 220 Tiops (int8) at 250W.

While Huawei's Shengteng 910 deliever 256T flops and 512T iops (int8) at 350W.

And this is Huawei's first AI series, so you see the picture.

 

[LeGenD]

By raw computing power.

And no Nv's best GPU is by no means 2080Ti, their best computing chip is Tesla V100 in terms of raw flops power, which deliver 125T flops and 50T iops(int8) at 300W.

And their best chip for int8 computing is Quadro RTX8000, deliever 220 Tiops (int8) at 250W.

While Huawei's Shengteng 910 deliever 256T flops and 512T iops (int8) at 350W.

And this is Huawei's first AI series, so you see the picture.

Raw computing power does not tell us much.

See for yourself: https://lambdalabs.com/blog/best-gpu-tensorflow-2080-ti-vs-v100-vs-titan-v-vs-1080-ti-benchmark/

 

[52051]

Raw computing power does not tell us much.

See for yourself: https://lambdalabs.com/blog/best-gpu-tensorflow-2080-ti-vs-v100-vs-titan-v-vs-1080-ti-benchmark/

And in your link you proved my point: in terms of computing power, Nvidia's best GPU is Tesla V100 in terms flops.

Lets cut the crap, I am quite familiar with AI stuff, and actually did and doing alot of CUDA coding, so stop wasting my time on this one.

 

[LeGenD]

And in your link you proved my point: in terms of computing power, Nvidia's best GPU is Tesla V100 in terms flops.

Lets cut the crap, I am quite familiar with AI stuff, and actually did alot of CUDA codes, so stop wasting my time on this one.

Depend upon your usage and how you see things.

RTX 2080 Ti is much more efficient per $ in comparison, and will trounce V100 in complex gaming applications. Therefore, I stick with my comments.

 

[Offshore]

And in your link you proved my point: in terms of computing power, Nvidia's best GPU is Tesla V100 in terms flops.

Lets cut the crap, I am quite familiar with AI stuff, and actually did and doing alot of CUDA coding, so stop wasting my time on this one.

i'm not familiar with computer things, so what's the different between Gtx 2080 TI series and tesla v100?
different segment or what?

 

[Hack-Hook]

i'm not familiar with computer things, so what's the different between Gtx 2080 TI series and tesla v100?
different segment or what?

Tesla v100 is 8000$ and Gtx 2080 TI is 1000$
another difference is 32gb memory vs 11gb
or the fact Gtx 2080 TI is only have 80% of Tesla V100 power
or the fact that nVIDIA didn't cripple double precision calculation in Tesla and Quadro series

 

[Beast]

And in your link you proved my point: in terms of computing power, Nvidia's best GPU is Tesla V100 in terms flops.

Lets cut the crap, I am quite familiar with AI stuff, and actually did and doing alot of CUDA coding, so stop wasting my time on this one.

LeGenD is a US lackey, despite US drone violating Pakistan Airspace, killing pakistanis at will. He will try lick the boots for American. Blindly defend anything from US. He is good at selective post and facts to try pull off his points.

 

[oprih]

Lmao the western slave is already crying and defending his white master passionately. What a funny scene!

 

[52051]

i'm not familiar with computer things, so what's the different between Gtx 2080 TI series and tesla v100?
different segment or what?

Tesla V100 support much better floating computing performance: FP64/FP32/FP16, and its design is more towards computing (register/cache/programmable cache/memory bandwidth etc).

Whilst Turing like RTX2080Ti or Quadro RTX8000, simply design for game/rendering and some AI computation, but lack the high precious floating point power the Tesla V100 have.

For HPC, machine learning thats involve anything beyond simply GEMM, V100 is much better, of cause, if price is not a factor.