AMD employee confirms new GPU with HBM and 300W

[Kowar]

AMD employee Linglan Zhang (System Architect Manager) claims on his LinkedIn profile:

 

Developed the world’s first 300W 2.5D discrete GPU SOC using stacked die High Bandwidth Memory and silicon interposer

 

 

Ilana Shternshain (Physical Design Engineer, team leader) claims on her profile:

 

AMD R9 380X GPUs (largest in “King of the hill” line of products)

 

R9 380X (King of the hill) 300W HBM incoming?

 

https://www.linkedin.com/pub/linglan-zhang/5/75b/4b0

https://www.linkedin.com/in/ilanashternshain

 

edit:

 

I love when OPs dont explain what "HBM" even stands for

 

 

"High Bandwidth Memory (HBM) is proposed stacked DRAM memory standard from AMD and Hynix, started in 2010 and adopted by JEDEC as JESD235 in October 2013.[1] The HBM is similar but incompatible with Hybrid Memory Cube stacked memory from Micron (2011).[2] HBM is targeted for next-generation high-performance graphics accelerators (like Nvidia's Pascal[3] and AMD's Fiji) and network devices.[4]"

 

Wiki page on HBM

 

"In 2.5D stacking, two dies are flipped over and placed on top of an interposer. All of the wiring is on the interposer, making the approach less costly than 3D stacking but requiring more area. Heat dissipation is not much of a concern, since cooling mechanisms can be placed on top of the two dies. This approach is also lower cost and more flexible than 3D stacking because incorrect connections can be reworked."

 

WCCF write up

Edited January 14, 2015 by Kowar

[Faceman]

All I saw is 300W.

[Kobz360]

All I saw is 300W.

Full Chip Timing Leader, MTS Asic/Layout design engineer AMD

 

April 2010

 – Present (4 years 10 months)AMD

• Member of Technical Staff, leading projects with corporate VP-level visibility.

• Leading teams of up to 20 members in Canada, US, China and India.

• Responsible for full-chip timing methodology and closure for AMD’s most exciting chips:

o AMD APU (incorporated into Sony PS4)

o AMD R9 290X GPUs

o AMD R9 380X GPUs (largest in “King of the hill” line of products)

 

 

So looks like AMD plans on becoming GPU kings soon (hopefully not of just their own products lol). 

 

Anyway, hopefully rumours are right and the 380X is the 980 (it'll have to be at this point) competitor and the 390X will compete with whatever NVIDIA spits back at them

 

EDIT - I have a feeling he said all he saw is 300W because of the shock of 300W please excuse me if I look like a cooked dog now

[JAKEBAB]

300w card? are u kidding me?! not even going to consider that thing even if its fast lol

[PerfectTemplar]

AMD Solution: Let's make our cards more power efficient and output less heat after all the jokes people make about us. THEN CLOCK IT UP AND PUMP AS MUCH POWAH AS POSSIBLE INTO IT TO MAKE IT TAKE OFF!

[Bogica]

With a 300w TDP it better perform the same as 2x 980. Also, they should fix their current line up because a TON of people still have driver and BIOS issues (tearing and white/gray screens, browser hw acceleration) on their high end 280/280x and some on 270/270x. Last time I recommend a high end AMD product to anyone...

[Notional]

With a 300w TDP it better perform the same as 2x 980. Also, they should fix their current line up because a TON of people still have driver and BIOS issues (tearing and white/gray screens, browser hw acceleration) on their high end 280/280x and some on 270/270x. Last time I recommend a high end AMD product to anyone...

 

Who said anything about 300w TDP?

 

https://www.youtube.com/watch?v=yDWO177BjZY

[PerfectTemplar]

Who said anything about 300w TDP?

 

https://www.youtube.com/watch?v=yDWO177BjZY

300W power is still nearly double a gtx 980s.

[qwertywarrior]

2.5D discrete GPU !!!

 

i thought intel would have been the first with a 2.5D product

this is awesome

[Khvarrioiren]

I have a feeling these don't link together, since there's been HBM prototype GPU's before the 3xx if I recall correctly. Atleast I really hope so. Last year was riddled with disappointments more or less throughout, and I'm pretty sure AMD wouldn't want to actually release a product with such high TDP.

[patrickjp93]

Given that's the last graphics architecture that was designed with Rory Read and those old execs still giving their (terrible) input, I'm not ready to pick up the pitchfork.

[patrickjp93]

2.5D discrete GPU !!!

 

i thought intel would have been the first with a 2.5D product

this is awesome

What does 2.5D mean?

 

Edit NVM I did some reading. Also, heck no. Intel will skip 2.5D. It wouldn't be caught dead increasing chip sizes.

[HippY]

What does 2.5D mean?

it is like DD but larger

/badjoke

[Notional]

What does 2.5D mean?

 

http://www.eetimes.com/document.asp?doc_id=1279514

[Bogica]

Who said anything about 300w TDP?

 

https://www.youtube.com/watch?v=yDWO177BjZY

Teach me more senpai! /sarcasm.

[patrickjp93]

I have a feeling these don't link together, since there's been HBM prototype GPU's before the 3xx if I recall correctly. Atleast I really hope so. Last year was riddled with disappointments more or less throughout, and I'm pretty sure AMD wouldn't want to actually release a product with such high TDP.

And we questioned the reference liquid cooler?

[Khvarrioiren]

http://www.eetimes.com/document.asp?doc_id=1279514

 

Quite a worrisome line in that article: "Today’s 3D stacks would melt away in 100W applications,” said Reiter. “It’s much better if the memory and processors are isolated from each other in a 2.5D configuration. Processors can typically run at 120C junction temperatures, but DRAMs start to encounter problems if you go beyond 85C junction temperatures."

[Khvarrioiren]

And we questioned the reference liquid cooler?

 

I never questioned it personally. I just have a hard time believing AMD would continue to shoot themselves in the foot over and over again. But alas, it's quite possible.

[patrickjp93]

I never questioned it personally. I just have a hard time believing AMD would continue to shoot themselves in the foot over and over again. But alas, it's quite possible.

Assuming AMD means to make the 380X the bang for buck card, I'm guessing this is the power usage or TDP (source not clear) of the 390X.

[qwertywarrior]

What does 2.5D mean?

 

 

http://www.eetimes.com/document.asp?doc_id=1279540

 

explains it very well

[Me1z]

300W really isn't that bad. Especially when you consider that the GTX 970 is a 250W GPU and the GTX 980 is a 275W GPU. You guys are way over reacting, that or you actually believe Nvidia's bullshit efficiency claims.

 

 

Source: http://www.tomshardware.com/reviews/nvidia-geforce-gtx-980-970-maxwell,3941-12.html

[Notional]

Teach me more senpai! /sarcasm.

 

What did you expect? What you are saying is not necessarily correct. Noone mentioned TDP, so that could be anything from 300w TDP and under.

 

Quite a worrisome line in that article: "Today’s 3D stacks would melt away in 100W applications,” said Reiter. “It’s much better if the memory and processors are isolated from each other in a 2.5D configuration. Processors can typically run at 120C junction temperatures, but DRAMs start to encounter problems if you go beyond 85C junction temperatures."

 

Not really, I don't see anyone putting dram dies on top of any processing die. Either it would be put in silicon like cache or using 2,5D (side by side), or completely separate.

 

 

What does 2.5D mean?

 

Edit NVM I did some reading. Also, heck no. Intel will skip 2.5D. It wouldn't be caught dead increasing chip sizes.

 

Why not? The Iris 5200 you talk about is a separate large die, next to the CPU die. Using 2,5D could increase the interconnect, making similar tech like HSA easier and simpler to make.

 

[patrickjp93]

What did you expect? What you are saying is not necessarily correct. Noone mentioned TDP, so that could be anything from 300w TDP and under.

 

Not really, I don't see anyone putting dram dies on top of any processing die. Either it would be put in silicon like cache or using 2,5D (side by side), or completely separate.

 

Why not? The Iris 5200 you talk about is a separate large die, next to the CPU die. Using 2,5D could increase the interconnect, making similar tech like HSA easier and simpler to make.

But that's not chip stacking in the traditional sense. 2.5D and 3D chips are about linking copies of the same chips together for stronger performance at the same tasks. I suppose you are correct: you could increase interconnect for Intel's iGPU, but I get the sneaking suspicion you have to make the whole chip at once already joined together, not fuse 2 proven good ones later. That's my worry about this tech. To me it's a defect density death trap if that assumption is correct.

[Notional]

But that's not chip stacking in the traditional sense. 2.5D and 3D chips are about linking copies of the same chips together for stronger performance at the same tasks. I suppose you are correct: you could increase interconnect for Intel's iGPU, but I get the sneaking suspicion you have to make the whole chip at once already joined together, not fuse 2 proven good ones later. That's my worry about this tech. To me it's a defect density death trap if that assumption is correct.

 

Not chip stacking no, that is why it is called 2,5D, and not 3D.

 

Normal 2D:

 

The main difference between a traditional 2D IC/SiP as shown above and a 2.5D IC/SiP as shown below is that, in the case of the 2.5D version, a silicon interposer is placed between the SiP substrate and the dice, where this silicon interposer has through-silicon vias (TSVs) connecting the metallization layers on its upper and lower surfaces.

http://www.eetimes.com/document.asp?doc_id=1279540 Thanks qwertywarrior for link.

 

2,5D:

 

From what I gather, you should indeed be able to take two proven chips, and fuse them to the silicon interposer. I wonder if this is how the 380/390 series will be connected to the real 3D stacked HBM. If so, then we might see AMD cards, become quite small, for high end versions, which would be amazing. Think if a 390x could fit in a standard ITX system (if cooler can handle TDP). 

[patrickjp93]

Not chip stacking no, that is why it is called 2,5D, and not 3D.

 

Normal 2D:

 

 

http://www.eetimes.com/document.asp?doc_id=1279540 Thanks qwertywarrior for link.

 

2,5D:

 

 

From what I gather, you should indeed be able to take two proven chips, and fuse them to the silicon interposer. I wonder if this is how the 380/390 series will be connected to the real 3D stacked HBM. If so, then we might see AMD cards, become quite small, for high end versions, which would be amazing. Think if a 390x could fit in a standard ITX system (if cooler can handle TDP). 

Sorry, miscommunication. 2.5D is meant to be the halfway point between planar chip design and chip stacking, but attaching a GPU to a CPU this way violates what chip stacking is meant to do as a matter of Orthodoxy.

 

yeah, I get the feeling itx versions will either be slowed way down or be forced to use liquid cooling as well.