Jingjia Micro developing GTX1080 perform-alike GPU

[HEXUS]

Chinese 28nm JM9271 GPU design is paired with 16GB HBM, has PCIe 4.0, 200W TDP.

Read more.

[Yoyoyo69]

28nm? Should be cheap. Imagine what they could do with latest fans.

[Kato-2]

Developing a GPU is one thing, developing the software stack to support it in the consumer market is another and a big one at that. Hence why these are military only pieces of hardware.

Of course a military entity like JM receives heavy state subsidies from the Chinese government to break Chinese dependency on western technology, will be interesting to watch what happens but this hardware definitely wont be sold beyond China for a long time if at all.

[globalwarning]

They have been hacking US companies long enough so... time to counter hack? Maybe find out what are they are doing that allows them to get GTX 1080 (16nm) like power on 28nm?

[Spreadie]

Developing a GPU is one thing, developing the software stack to support it in the consumer market is another and a big one at that.

That's what APIs like direct x are for, surely? Of course, if it is direct to metal, it could explain the claimed comparison (albeit skewed) with a 1080 which would be running via the HAL -the same way relatively low end GPUs perform better in consoles than in PCs. I'm thinking it's probably the only way they can claim that kind of performance on 28nm and only 200W TDP.

Either way, something smells off.

[Spud1]

Could be interesting - even if it's several years off, we could do wtih a new player in the market. Nvidia could do with some competition really. One to keep an eye on

[ohmaheid]

#globalwarning I think you'd need to have a word with Nvidia's CEO. Since they signed a deal with China's Baidu to develop AI.

[b0redom]

Well, It's time to change a computer. noah-pm.en.alibaba.com . Get your weapons .

Wow, I think this is the worst targetted spam I've *ever* seen.

[peterb]

Wow, I think this is the worst targetted spam I've *ever* seen.

It is no more!

[scaryjim]

That's what APIs like direct x are for, surely? Of course, if it is direct to metal, it could explain the claimed comparison (albeit skewed) with a 1080 which would be running via the HAL -the same way relatively low end GPUs perform better in consoles than in PCs. I'm thinking it's probably the only way they can claim that kind of performance on 28nm and only 200W TDP.

Either way, something smells off.

Well, firstly the comparison appears to be based on theoretical throughput, and it's actually more than 10% behind GTX 1080 in that regard - 8000 GFLOPs v 8900 for the 1080. 2015's Fury X hit > 8600 GFLOPS on paper at 275W TDP, so I see no reaosn to think that a 28nm chip with 4 years of process maturation couldn't hit 8000 GFLOPs at 200W.

In fact, the Fury X is quite a nice comparison, because it also had an HBM interface providing 512GB/s bandwidth. So you could equally say this chip gives Fury X-like performance. Remind me again how the Fury X stands up to the GTX 1080 in real world performance...? That's why the software stack is so critical.

Secondly - and perhaps more significantly - the JM chip only lists OpenGL/OpenCL compatibility. At the very least that means they'd need to develop a DX driver if they wanted to push them into the consumer space, and it could even mean that the shader design is heavily tuned for OpenGL/CL performance and they'd be very poor DX performers anyway.

So nothing here makes me think the chip isn't as described - it's just being described to try to show it in the best possible light, rather than a realistic one...

[cheesemp]

They have been hacking US companies long enough so... time to counter hack? Maybe find out what are they are doing that allows them to get GTX 1080 (16nm) like power on 28nm?

A bigger chip?

[globalwarning]

They have been hacking US companies long enough so... time to counter hack? Maybe find out what are they are doing that allows them to get GTX 1080 (16nm) like power on 28nm?

A bigger chip?

That's one variable in a large, complex equation. They haven't indicated how big the chip is, but I suppose on could calculate how much bigger it would need to be, approximately, to have a similar number of transistors on a 16 vs 28nm process.

[Tabbykatze]

That's one variable in a large, complex equation. They haven't indicated how big the chip is, but I suppose on could calculate how much bigger it would need to be, approximately, to have a similar number of transistors on a 16 vs 28nm process.

Condoning cyber warfare because you don't understand the relation between the number of transistors, chip volume and resultant performance for something as mediocre as a GPU fascinates me.