Next Generation Hardware Speculation with a Technical Spin [post E3 2019, pre GDC 2020] [XBSX, PS5]

[DavidGraham]

Your own math suggests there is a mere 5% error in linear scaling from 22 to 40 CUs. I can live with that.

Could be worse the further you scale the CU count up, in fact, the fact that AMD never released a 64 CU part first heavily implies that scaling is worse at the higher end of the spectrum.

In fact this is often the case, going from Vega 56 to Vega 64 yields a 40% increase in power consumption despite only a 14% increase in CU count and a mere 100MHz clock uplift.

You also make a ton of assumptions about AMD’s internal decisions

Even with bad yields, releasing this part would have made a lot more sense than releasing the short lived Radeon Vii. If you are postulating that a 64 CU die will have bad yields, then you can bet it's not going to be used in a console, as it will massively increase than BoM.

The worse assumption here is the 64 CU part accepting the same voltages as a lower CU part at iso clocks.

He said 260W, I would say more conservatively 280W.

My previous estimates puts it at 280w or more for the GPU alone.

His estimates puts it at 260w for the entire SoC + memory + everything else.

 

[MrFox]

My previous estimates puts it at 280w or more for the GPU alone.

His estimates puts it at 260w for the entire SoC + memory + everything else.

What is your estimate for the entire console from the wall?

 

[anexanhume]

Could be worse the further you scale the CU count up, in fact, the fact that AMD never released a 64 CU part first heavily implies that scaling is worse at the higher end of the spectrum.

No. It implies they hadn’t designed a 64 CU part. Anything else is supposition.

In fact this is often the case, going from Vega 56 to Vega 64 yields a 40% increase in power consumption despite only a 14% increase in CU count and a mere 100MHz clock uplift.

A different architecture on a different node with unknown voltages to compare with. Vega often had to be pushed to extreme voltages to hit speed targets. The core and the memory.

My previous estimates puts it at 280w or more for the GPU alone.

His estimates puts it at 260w for the entire SoC + memory + everything else.

A Xbox One X draws less at the wall than a 580 does despite having more TF.

 

[Proelite]

Hmm seems like a lot of chatter going around for Xsx on 7nm+...

Don't know what to make it of it. Still doubt.

 

[snc]

No. It implies they hadn’t designed a 64 CU part. Anything else is supposition.


A different architecture on a different node with unknown voltages to compare with. Vega often had to be pushed to extreme voltages to hit speed targets. The core and the memory.



A Xbox One X draws less at the wall than a 580 does despite having more TF.

xox has les tf (6 vs 6.175 )

 

[anexanhume]

xox has les tf (6 vs 6.175 )

At boost, an unguaranteed clock.

 

[MrFox]

Hmm seems like a lot of chatter going around for Xsx on 7nm+...

Don't know what to make it of it. Still doubt.

They are about a year after the start of TSMC volume production, which puts it somewhere a little earlier than ps4 pro in regards to 16nm (16ff was in mass production in august, so 15 months?). Maybe possible?

I would definitely see both of them use the intermediate N7P though. Which is still DUV not EUV.

 

[snc]

At boost, an unguaranteed clock.

indeed but most rx580 on market on avarage are close to boost so talking that xox has higher tf is little misleading

 

[RDGoodla]

They are about a year after the start of TSMC volume production, which puts it somewhere a little earlier than ps4 pro in regards to 16nm (16ff was in mass production in august, so 15 months?). Maybe possible?

I would definitely see both of them use the intermediate N7P though. Which is still DUV not EUV.

PS4 slim was released in sep. 2016, also used 16nm.

So the timing is ok for 7nm+ in 2020 holiday.


In fact all the info points to 7nm+: 64CUs, 350mm2, next-generation RDNA.

 

[milk]

In fact all the info points to 7nm+: 64CUs, 350mm2, next-generation RDNA.

What is all the info that points to those specific numbers?

 

[chris1515]

not on the right place sorry I think it was official

 

[vjPiedPiper]

Credit to DemonCleaner at NeoGAF for his table.

A theoretical 64 CU GPU at 1500 MHz with 10 GDDR6 chips and 3.5GHz Zen 2 APU might draw:

140W GPU
35W/2 memory
50W CPU
TOTAL: 208W
TOTAL AT THE WALL:~260W (18W memory power, 10W SSD, OD, misc, 90% efficient supply)

Well this looks like the result of some good work. I will have to see if i can chase up the original source and if they have documented their process and assumptions.
but to me, imho, the numbers look within the realm of possibility, maybe the GPU draw is a bit on the low side, but it's amazing how the usage goes down once you drop down the Speed.
Either way i think this is a somewhat realistic look at what we might get in the XBSX.
The two biggest things for me are a 64CU GPU, would be too big - imho, and 3.5Ghz for the CPU seems to high...
But i view it as a somewhat realistic attempt to model the properties of a future console, similar to the ryzen7 + 3700x system stuff we have discussed in the past.

 

[Tkumpathenurple]

Could the 64CU rumour be true, but not in the way it's been presented so far?

Given the PS4 sacrificed 2 of its 20 CU's, and both the Pro and X1X sacrificed 4, could the XSeX sacrifice more?

If so, maybe aggressive binning could be a means of maximising usable chips in a wafer across the entire Xbox brand?

For example:

• XSeX = 64CU's with 8 disabled = 56 total CU's. Clocked high @ 1800MHz.
• Lockhart = 64CU's with 16 disabled = 48 CU's. Clocked mid @ 1500MHz.
• X1X = The shit chips which are only just over 60% functional = 40 CU's. Clocked low @ 1200MHz.

At launch, just produce a single SoC (or GPU+I/O die) and use it as above. Two or three years in - once the process has matured, and the market made its demands clear - produce two chips: one for Anaconda, one for Lockhart, and bin the less performant ones of each for the tier below.

 

[turkey]

Even with bad yields, releasing this part would have made a lot more sense than releasing the short lived Radeon Vii. If you are postulating that a 64 CU die will have bad yields, then you can bet it's not going to be used in a console, as it will massively increase than BoM.

.

We come back to Lockhart and I would think if true also the next Xcloud blades.


Was there talk of cross team design?

Azure compute small / medium / large. Could even sub bin the rejected chips?

 

[Shifty Geezer]

My previous estimates puts it at 280w or more for the GPU alone.

His estimates puts it at 260w for the entire SoC + memory + everything else.

MrFox makes a very good point. We know XBSX is 12 TFs*. How many watts is MS working with here? You reckon 380-400W? To date, the most electron-guzzling console has been the PS3 at 200W tops.

* Edit : I suppose we don't know that, thanks to Naviflopgate. Possibly 10TFs.

 

[Pinstripe]

I'm pretty sure all console manufacturers won't be pushing their consoles beyond 200W, to stay in accordance with EU energy efficiency law for game consoles which they all have agreed on. And how would that make Sony and MS look like if they can't deliver a Green product?

Forget your 64 CUs pipe dream. Unless RDNA2 pulls a rabbit out of the hat and will be as Ultra efficient as Nvidia.

 

[Mskx]

I'm pretty sure all console manufacturers won't be pushing their consoles beyond 200W, to stay in accordance with EU energy efficiency law for game consoles which they all have agreed on. .

Do you have a source for that agreement?

 

[Tkumpathenurple]

And how would that make Sony and MS look like if they can't deliver a Green product?

 

[Pinstripe]

Do you have a source for that agreement?

http://efficientgaming.eu/

Neither PS4 Pro or X1X have been pushing past ~165-175W in gaming. I doubt they'll start doing that in next-gen. Greta wouldn't be happy.

 

[Shifty Geezer]

http://efficientgaming.eu/

I can't find a max limit of 200 W.