[EuroGamer] AMD has developed a 20nm version of Xbox One APU

RussianSensation

Elite Member
Sep 5, 2003
19,458
765
126
"Chip designer AMD has developed a smaller, more cost-effective 20nm version of the main processor found in Xbox One, opening the door to a cheaper 'slim' version of the Microsoft console. The news comes from the LinkedIn bio of AMD's senior manager of SOC (system on chip) physical design, as discovered by Mosen from the Beyond3D forums. The entry in question says he "successfully planned and executed the first APU for Microsoft's Xbox One Game Console in 28nm technology and a cost-reduced derivative in 20nm technology."

Technology becomes smaller and more powerful according to the density of transistors crammed onto a piece of silicon, with production processes measured at the nanometre level. Currently, the Xbox One and PS4 processors are fabricated at 28nm, meaning each transistor measures 28 thousand-millionths of a metre. The new chip design shrinks that to 20nm. According to TSMC (Taiwan Semiconductor Manufacturing Company) - the factory that physically makes the Xbox One and PS4 processors - shrinking down the same chip provides 30 per cent higher speed or 25 per cent less power than the existing processor, at 1.9 times the density."


Source

^ This news is a positive ray of light that 20nm GPUs could be available in 2015. It is often said that there is hardly any benefit of shrinking from 28nm to 20nm for GPUs but it's certainly not true from a technological perspective. 1.9X the density and 30% higher speed or 25% lower power is worth it if 16nm is still 1.5-2 years away. Certainly if 20nm was not cost effective either than MS and Sony would not want AMD to shrink their APUs to 28nm and would have waited to lower nodes. Therefore, both of these myths that financially and technically the 20nm node is "essentially worthless" are bogus claims.
 

Gloomy

Golden Member
Oct 12, 2010
1,469
21
81
The Bone SoC may be a special case because of the huge chunk of ESRAM it has on-die. This probably takes much better gains from a shrink than a plain old chip would.

This is just a wild guess, though.
 

jpiniero

Lifer
Oct 1, 2010
14,510
5,159
136
The Bone SoC may be a special case because of the huge chunk of ESRAM it has on-die. This probably takes much better gains from a shrink than a plain old chip would.

This is just a wild guess, though.

That sounds about right. Even if they don't really gain much in terms of transistor/$, they could move to a cheaper cooler to save costs that way. I'm sure going 20 nm was part of the contract so AMD probably has to do it anyway regardless of the cost.
 

Techhog

Platinum Member
Sep 11, 2013
2,834
2
26
Before people make crazy assumptions, they won't make the chip faster. It'll take advantage of the power consumption benefits.
 

RussianSensation

Elite Member
Sep 5, 2003
19,458
765
126
The Bone SoC may be a special case because of the huge chunk of ESRAM it has on-die. This probably takes much better gains from a shrink than a plain old chip would.

This is just a wild guess, though.

I find it hard to believe that AMD would make a 20nm version of Xbox One APU but would not do the same for Sony. It's highly unlikely to me that AMD would allow MS to enjoy a 20nm advantage in terms of manufacturing costs and lower power usage, which allows a reduction in complexity of the console's motherboard, cooler, etc. but wouldn't propose the same to Sony. The information just hasn't leaked yet on PS4's shrink.

Also, the die sizes aren't much different for both consoles.

"The Xbox One die is 363 square millimeters, up from the PS4’s 348 sq mm. The 5% additional space, despite having the smaller GPU core, is mostly due to RAM."
http://www.extremetech.com/gaming/1...ered-reveals-sram-as-the-reason-for-small-gpu

That sounds about right. Even if they don't really gain much in terms of transistor/$, they could move to a cheaper cooler to save costs that way. I'm sure going 20 nm was part of the contract so AMD probably has to do it anyway regardless of the cost.

But the 20nm node's advantage apply to the entire node, not specific to Xbox One's APU. Why wouldn't AMD want to use 20nm for discrete GPUs if they already have experience shrinking Xbox One APU to that node at TSMC?
 
Last edited:
Aug 11, 2008
10,451
642
126
Sounds legit, but is there an official announcement? I don't consider somebody's linked-in profile a particularly reliable source.
 

monstercameron

Diamond Member
Feb 12, 2013
3,818
1
0
Before people make crazy assumptions, they won't make the chip faster. It'll take advantage of the power consumption benefits.

sony bumped up the psp ram and cpu speed with the slim. It isn't impossible and if they can get a few extra MHz then maybe they go for it.
 

tviceman

Diamond Member
Mar 25, 2008
6,734
514
126
www.facebook.com
The 30% higher speed or 25% lower power consumption and the 1.9x density improvement are all best case scenarios. I don't know what the best-case scenarios moving from 40nm to 28nm were, but 1.9x density sounds great while the 30% higher speed sounds meh, especially when reiterating those are best case scenarios. I think to take full advantage of node, the chip probably needs to be designed from the ground up for that node in mind.

I don't know though. TSMC has said 20nm FF (16nm) will be in volume product this year, so 20nm will likely be short lived. My guess is that AMD shrank the XB1 chip down to 20nm because Microsoft is trying to lower costs since their console generally costs more than the PS4. Anyways, I think AMD is going with 20nm in the first half of 2015 for their GPU's, while Nvidia will ride out 28nm Maxwell GPU's until 20nmFF is ready and skip 20nm altogether. If Nvidia goes with 20nm GPU's, I think they'll just be shrinks with slightly higher clocks.
 

Erenhardt

Diamond Member
Dec 1, 2012
3,251
105
101
Aren't the console apus designed to be easily portable to other fabs/nodes?

If they can bring the cost down, it will be amazing. New console may seem dirt cheap, but that's US standards, actually any developed country standard. In developing (poor) countries consoles are quite expensive.

Interesting fact: If going 20nm with console SOC is viable, then its quite contrary to what some posters say about transistor cost and how only intel can keep shrinking nodes and costs at the same time.

They could boost clocks a bit to help maintain steady FPS in the most demanding scenes.
 
Last edited:

Mondozei

Golden Member
Jul 7, 2013
1,043
41
86
This news is a positive ray of light that 20nm GPUs could be available in 2015.


Russian, why are you perpetuating the myth of the "no 20 nm GPU"? You should - and I know that you do - know better. Lisa Su has stated many times over that 20 nm GPUs are coming in 2015 and so has Nvidia.

For me it is just amazing how long a shelf life a myth has that has been debunked over and over again. Just amazing.
 
Sep 27, 2014
92
0
0
Russian, why are you perpetuating the myth of the "no 20 nm GPU"? You should - and I know that you do - know better. Lisa Su has stated many times over that 20 nm GPUs are coming in 2015 and so has Nvidia.

For me it is just amazing how long a shelf life a myth has that has been debunked over and over again. Just amazing.

Arent the GPUs Lisa Su refers to actually 16nm FF?
 

ShintaiDK

Lifer
Apr 22, 2012
20,378
145
106
Interesting fact: If going 20nm with console SOC is viable, then its quite contrary to what some posters say about transistor cost and how only intel can keep shrinking nodes and costs at the same time.

People seem to have read what they wish to and closed their eyes

A direct shrink without any additional transistor count only marginally increases the cost (5-10%). And in the case of MS/Sony. They can regain on 1 part what is lost on the other. (Smaller PSU, smaller cooler, smaller case etc.)

Now try to make a new chip with say 1.9x the transistor budget and then look at the price. Thats the issue. Unlike every precious node, the transistor cost is higher. So instead of lower cost, you have to do the economics against the electrical properties that have to outweight the higher cost.

However all of this is still based on a linkedin account.
 
Last edited:

OVerLoRDI

Diamond Member
Jan 22, 2006
5,494
4
81
They could boost clocks a bit to help maintain steady FPS in the most demanding scenes.

If the 360 is anything to go buy, they aren't likely to make any significant changes to the performance of the hardware, just make changes to save power and lower costs.
 

Techhog

Platinum Member
Sep 11, 2013
2,834
2
26
sony bumped up the psp ram and cpu speed with the slim. It isn't impossible and if they can get a few extra MHz then maybe they go for it.
That was to improve system performance, not game performance. Same goes for the New 3DS, other than one game which is causing a crapstorm.
 

Erenhardt

Diamond Member
Dec 1, 2012
3,251
105
101
Yes, but there is a sweat spot somewhere. 20% speed 1.5 times density and 15% power savings at the same time sounds nice.
 

AtenRa

Lifer
Feb 2, 2009
14,000
3,357
136
Uh... pretty sure the only operand is OR. There is no AND.

"TSMC's 20nm process technology can provide 30 percent higher speed, 1.9 times the density, or 25 percent less power than its 28nm technology."
http://www.tsmc.com/english/dedicatedFoundry/technology/20nm.htm

Meaning TSMC's 20nm process technology can provide:
30% higher speed OR 1.9 time density OR 25% less power.

Nope,

20nm has 1.9x density over 28nm.
30% higher speed than 28nm OR 25% less power at the same speed of 28nm.
 

jpiniero

Lifer
Oct 1, 2010
14,510
5,159
136
But the 20nm node's advantage apply to the entire node, not specific to Xbox One's APU. Why wouldn't AMD want to use 20nm for discrete GPUs if they already have experience shrinking Xbox One APU to that node at TSMC?

There is a still a significant cost to building the die even if the "hard part" is essentially being funded by MS/Sony. It may or may not be worth it to AMD.

I'm sure you will see a 20 nm PS4 shrink as well.
 

f1sherman

Platinum Member
Apr 5, 2011
2,243
1
0
Nope,

20nm has 1.9x density over 28nm.
30% higher speed than 28nm OR 25% less power at the same speed of 28nm.

Thats not what TSMC says.

You are saying they can double the transistor count at the same area(1.9x density) while providing 25% less power(same clocks).

Straight shrink giving you double the transistor count and 1/4 less power.
If that was the case we'd be running at passive coolers by now and there wouldn't be need for arch changes.
We need both shrink+arch to come up with something that's somewhat numerically similar to what you claim is a pure shrink.

Transistor density on the same fab process is not constant even across the same IHV/arch/gen.
 
Last edited:

RussianSensation

Elite Member
Sep 5, 2003
19,458
765
126
Russian, why are you perpetuating the myth of the "no 20 nm GPU"? You should - and I know that you do - know better. Lisa Su has stated many times over that 20 nm GPUs are coming in 2015 and so has Nvidia.

For me it is just amazing how long a shelf life a myth has that has been debunked over and over again. Just amazing.

I am not. I am stating the opposite that it's highly likely that GPU makers, at least AMD, will have 20nm GPUs in 2015. I can't say 100% because AMD constantly misses their deadlines. Therefore just because Lisa Su stated that 20nm GPUs will be out in 2015, it's not 100% until they actually launch.
 

DeathReborn

Platinum Member
Oct 11, 2005
2,743
734
136
Aren't the console apus designed to be easily portable to other fabs/nodes?

If they can bring the cost down, it will be amazing. New console may seem dirt cheap, but that's US standards, actually any developed country standard. In developing (poor) countries consoles are quite expensive.

Interesting fact: If going 20nm with console SOC is viable, then its quite contrary to what some posters say about transistor cost and how only intel can keep shrinking nodes and costs at the same time.

They could boost clocks a bit to help maintain steady FPS in the most demanding scenes.

If they do boost performance it would most likely be to have the Kinect "10%" back & not to create a 2 tier console ecosystem.
 

Enigmoid

Platinum Member
Sep 27, 2012
2,907
31
91
Thats not what TSMC says.

You are saying they can double the transistor count at the same area(1.9x density) while providing 25% less power(same clocks).

Straight shrink giving you double the transistor count and 1/4 less power.
If that was the case we'd be running at passive coolers by now and there wouldn't be need for arch changes.
We need both shrink+arch to come up with something that's somewhat numerically similar to what you claim is a pure shrink.

Transistor density on the same fab process is not constant even across the same IHV/arch/gen.

I suggest you double check the literature.

Density scaling is an effect of the smaller geometries involved. With smaller features power usage decreases. With that headroom you can now ramp up the clocks getting more performance at the same power (though always the performance gain is less than the equivalent power scaling due to the non-linear frequency vs. power graph).

Also note that the 28 nm -> 20nm isn't getting the full electrical benefits of a node shrink as TSMC doesn't have Finfets.

Its density gain + less power at the same clocks OR higher clocks at the same power.

Note that 1.9x density shrinking with only 25% power scaling means that heat produced per unit area is going up.
 

f1sherman

Platinum Member
Apr 5, 2011
2,243
1
0
I suggest you double check the literature.

Density scaling is an effect of the smaller geometries involved. With smaller features power usage decreases. With that headroom you can now ramp up the clocks getting more performance at the same power (though always the performance gain is less than the equivalent power scaling due to the non-linear frequency vs. power graph).

Also note that the 28 nm -> 20nm isn't getting the full electrical benefits of a node shrink as TSMC doesn't have Finfets.

Its density gain + less power at the same clocks OR higher clocks at the same power.

Note that 1.9x density shrinking with only 25% power scaling means that heat produced per unit area is going up.

REALLY?

If you cut the 20nm chip in half you'll end up with half the area of some 28nm chip that has roughly the same(1.9/2) the number of transistors.
And you'll need (100-25)% / 2 = 37.5% power for your 20nm cut in half chip.
So I suggest you double check the literature. Geometry and basic algebra mostly.

Further on, TSMC's own quote (notice "OR" between the density and the power?) is not totaly clear, but it never claims uniformal 1.9 times density on which you insist on.
And you insist on it by by using supremely bleeding obvious and tangential argument - things get smaller therefore they require less space.
You don't say! Like I claim otherwise.

But the resulting density is combination of fab process and the chip design, and it's not solely the result of fab process, so even ideally you won't end up with uniform 1.9x density

Doesn't the density increase come in all situations and the only trade off is between speed or power consumption?
SRAM can use larger cells than the process minimum for performance or reliability reasons.
Logic has tended to scale less than SRAM with each node, and more complex or higher speed logic has shrunk more slowly.
Less regular structures may need extra measures to manufacture, and wires in more complex logic can restrict density improvement.
When increasing for drive strength for FinFETs, extra fins and thus extra area can used to provide more current.

For control of leakage or manufacturability, parts of the chip can use physically larger transistors to reduce leakage and resist variation. Power gates are physically large relative to the rest.
Even with ideal scaling, the leakage and power efficiency issues will require allocating more of the larger transistor budget to more aggressive power control measures.
http://forum.beyond3d.com/showpost.php?p=1684671&postcount=108
 

raghu78

Diamond Member
Aug 23, 2012
4,093
1,475
136
REALLY?
If you cut the 20nm chip in half you'll end up with half the area of some 28nm chip that has roughly the same(1.9/2) the number of transistors.
And you'll need (100-25)% / 2 = 37.5% power for your 20nm cut in half chip. So I suggest you double check the literature. Geometry and basic algebra mostly.

completely wrong. you have not understood how a node shrink works.

Further on, TSMC's own quote (notice "OR" between the density and the power?) is not totaly clear, but it never claims uniformal 1.9 times density on which you insist on. And you insist on it by by using supremely bleeding obvious and tangential argument - things get smaller therefore they require less space. You don't say! Like I claim otherwise. But the resulting density is combination of fab process and the chip design, and it's not solely the result of fab process, so even ideally you won't end up with uniform 1.9x density
That statement on TSMC website is poorly worded. The density increase is not a OR choice. So you will get the density increase when you go for a node shrink. You can either reduce power at the same specs or improve perf at same power. see slide 19 . comparison of TSMC processes 16FF vs 20SOC vs 28HPM

http://www.eda.org/edps/edp2013/Papers/4-4 FINAL for Tom Quan.pdf

16FF provides 38% more perf at same power wrt 28HPM as 20% wrt 20SOC

perf at same power 28HPM perf is 1x and 16FF - 1.38x (38% speed gain at same power)

16FF is 20% faster than 20SOC - Y * 1.2 = 1.38x , Y = 1.38x / 1.2 = 1.15x

so 20 SOC provides 15% perf gain over 28HPM at same power

Now coming to power at same perf 28HPM - 1x 16FF - 0.46x (54% power reduction)

16FF provides 35% power reduction over 20SOC

Y * 0.65 = 0.46x , Y = 0.46x / 0.65 = 0.707x

so 20SOC provides roughly 30% power reduction over 28HPM at same perf

[FONT=&quot]Gate density 16FF provi[FONT=&quot]des 2x the density of 28H[FONT=&quot]PM

[/FONT][/FONT][/FONT][FONT=&quot][FONT=&quot][FONT=&quot][FONT=&quot]28HPM 1x , 16FF [/FONT][/FONT][/FONT]- 2x
[/FONT]
[FONT=&quot]16FF provides 1.1x density of 20SOC
[/FONT]
[FONT=&quot]Y * 1.1 = 2x , Y = 2x/1.1 = 1.8x[/FONT]
[FONT=&quot]
so 20SOC provides 80% density increase over 28HPM
[/FONT]
[FONT=&quot]Now this is simple algebra :thumbsup:
[/FONT]
 
Last edited: