Rumor: XBox dual SKU next-gen launch

[mrcorbo]

This got me thinking.
In the past kept selling the previous gen due to cost.
But why would they need to sell any XO's next gen. Could do a top to bottom Scarlett launch.

Maybe could keep selling 1S for markets where streaming isn't viable and they want a very cheap console.
Otherwise the lowest end Scarlett arcade/s model should suffice.

I have my doubts about being able to get the 1X down in price enough to compete with Scarlett arcade, even in 2 years. But question is why keep selling it?

I don't see them continuing to sell the One X for very long into the next generation. The One S and disc-less SKU could continue to exist as budget/GamePass consoles.

Scarlet X - $499
Scarlet base - $299
One S - $199
One mini (discless SKU) - $149

Shifting to after some time:

Scarlet X - $399
Scarlet Base - $199
One mini or Streaming only box - $99

Eventually to:

Scarlet Mid-gen refresh - $499
Scarlet X Slim - $299
Scarlet Base Slim - $149
Streaming only box - $99

 

[Shifty Geezer]

15% better performance?
Scarlett X could be around the 12tf mark if not more, that's double the performance right there.

How are you getting double the performance just by enabling a few CUs and having 10% higher clockspeed? I mean, you state Scarlett S as 4 disabled CUs, 10% lower clocks. Scarlett X is 10% higher clocks and +4 CUs over Scarlett S in your model. What are the performance and price differentials of these two devices? Surely the X isn't double the performance to the S in this case.

 

[BRiT]

Uber-CUs!

 

[Silent_Buddha]

I considered that, but smartphones (at least the ones people pay attention to) can command prices close to $1000 and actually have real profit margins. I see one console SKU carrying the other in a dual SKU launch.

It also happens with phones in the 200-400 USD price range from both the bigger and smaller smartphone manufacturers.

Regards,
SB

 

[mrcorbo]

What is the upper bound for number of CUs expected in the SoC for the next-gen consoles using following as references:

The MI60 and has 64 CUs at a die size of 331mm2 @ 7nm

The OG Xbox One Soc had a die size of 363 mm2 (It had the biggest die of the current gen. Even the Scorpio Engine is smaller)

The Zen 2 chiplets are being estimated as being between 72mm2

Then we can start estimating clockspeeds with the reference that the aforementioned MI60 is @ 1800 MHz.

Once a full fat configuration can be determined and the TFs calculated, we can start trying to see what combination of CU cut-downs and clock speed drops get us down to ~6TF and whether that makes sense.

Edit: Corrected MI60 clockspeed.

 

[Jay]

How are you getting double the performance just by enabling a few CUs and having 10% higher clockspeed? I mean, you state Scarlett S as 4 disabled CUs, 10% lower clocks. Scarlett X is 10% higher clocks and +4 CUs over Scarlett S in your model. What are the performance and price differentials of these two devices? Surely the X isn't double the performance to the S in this case.

Super Navi+ NCU's.

I did say that the figures themselves weren't correct.
if you want you could up the ghz much more as i expect the X to use vapor chamber again,
the arcade to be clocked a lot lower with disabled cu's, to use cheap cooling and power delivery.

Going by what we currently know (yes based on leaks) we expect an arcade, an x model, and Scarlett also in cloud. Just spitballing how could do it, I should've just said x and y for figures to be fair, was more concept of the idea.

It would be a waste of silicon if was disabling lot of good cu's just so have enough for arcade, but the fact got the x, dev box (minor amount), cloud (possibly big supply) to also produce from the wafer it could ballance the cost a lot better.
Someone with better idea of yield and wafer usage/cost would need to say how reasonable it is to disable 40% of cu's and be clocked lower and the affect to overall cost, when factoring in the other apu's being produced.

 

[mrcorbo]

What is the upper bound for number of CUs expected in the SoC for the next-gen consoles using following as references:

The MI60 and has 64 CUs at a die size of 331mm2 @ 7nm

The OG Xbox One Soc had a die size of 363 mm2 (It had the biggest die of the current gen. Even the Scorpio Engine is smaller)

The Zen 2 chiplets are being estimated as being between 72mm2

Then we can start estimating clockspeeds with the reference that the aforementioned MI60 is @ 1800 MHz.

Once a full fat configuration can be determined and the TFs calculated, we can start trying to see what combination of CU cut-downs and clock speed drops get us down to ~6TF and whether that makes sense.

Edit: Corrected MI60 clockspeed.

I was hoping someone would pick up that gauntlet and do a better job than I am about to, but here's a very rough target with rounding and shortcuts taken all over.

363mm2 maximum size for die minus the 72mm2 for the CPU portion leaves 291mm2 of die available for the GPU. If 331mm2 can accommodate 64 CUs then 291mm2 can accommodate (64 * 291 / 331) 56 CUs.

The TDP of MI60 @ 1800 MHz is 300W. Yeah, not so much for a console part. 7nm allowed a 300Mhz boost over 14nm with the same TDP when comparing MI60 to MI25. The Radeon 590 on 14nm has a base clock of 1469 MHz for 36 CUs @ 175W TDP. So, pulling number from nether regions, 1500 MHz for 56 CUs @ 7nm fits within a reasonable power budget for a console?

56 * 1500 * 128 = 10.75 TF for high-end SKU. For reference, 1600 MHz gets you 11.46TF, 1700 gets you 12.18TF, and 1800 gets you 12.9TF.

52 * 1100 * 128 = 7.3 TF for low-end SKU

There. Feel free to pick at this, but if you do I ask you to show your work. None of this next-gen needs to be x% faster or x multiple of the performance of last gen crap with zero consideration of what can actually be manufactured.

 

[DuckThor Evil]

There. Feel free to pick at this, but if you do I ask you to show your work. None of this next-gen needs to be x% faster or x multiple of the performance of last gen crap with zero consideration of what can actually be manufactured.

Good post, I do have to remind that using the Radeon Vega 20 chip is not that great of a base for calculations, even if we don't have much to go on at this point. It's very reasonable to argue that 64CUs could be put into a smaller die. Vega 20 has a ton of features that the console chip will not have. I believe TSMC's 7nm should offer far better logic density improvement compared to their 16nm process than what the shrink of Vega chip would imply. I am not expecting less than 64Cus and doubling Scorpio should imo be doable... Not sure about clocks though.

 

[Jay]

56 * 1500 * 128 = 10.75 TF for high-end SKU. For reference, 1600 MHz gets you 11.46TF, 1700 gets you 12.18TF, and 1800 gets you 12.9TF.

52 * 1100 * 128 = 7.3 TF for low-end SKU

Thanks for putting some figures to the sort of thing I was thinking.

I think the X will be clocked like crazy with vapor chamber, and better power delivery.
52 * 1700 * 128 = 12.18TF X model

The arcade, cheap cooling and power delivery to get costs down as far as possible.
52 * 925 * 128 = 6.1TF for arcade.

Don't know how reasonable it is regarding TDP etc.
The X has shown that high frequency is possible in console though.

 

[DuckThor Evil]

Just to extend a bit to my previous post. On paper TSMCs 7nm process should offer larger density improvement compared to their 16nm process vs what the improvement was with their 28nm vs 16nm processes. According to TSMC the Logic size improvement from 28nm to 16nm was around 100%, meaning double the transistors per mm2 of silicon. Actually they pretty much achieved that with their 20nm process already and 16nm "only" brought performance and power saving improvements, but not much area improvements.

Now TSMC is stating that same 100% improvement with 16nm vs 10nm and further 60% improvement to 7nm, so the logic density jump should be bigger than what we saw before. Then consider that the CU counts for the mid gen consoles went up from 12/14 to 40/44 on Xbox (on chip/enabled, over 3x increase) and from 18/20 to 36/40 (2x increase) on the PS4 side, while the die sizes went slightly down. (Of course on the Xbox side getting rid of the ESRAM helped a ton there.) The size for a singular CU was measured to be between 3-3.5mm2 on the 16nm process.

Now we are getting a bigger density improvement or if being conservative a similar density increase and we should expect only 1.4x (56/40) increase in CU count as being possible?
Like I said I expect 2X to be possible with the area given, I'm just not sure if they want to keep similar die size or if there are other architectural limits with the CU counts. I remember 64CUs being some type of hard limit for the current AMD tech, so that might be a limiting factor, however if that ends up being the case, the die size should be smaller, unless they fill it up with something else like on chip memory again.

Teraflop count is hard to come by with, as we have limited information on Navis performance on 7nm, but for the X replacement, which I expect to be the strongest next gen console, I think it's going to be around 12-13TF, whether that is with 64CUs and higher clocks or more CUs and lower clocks, I am not sure.

 

[ultragpu]

Since the other hardware release strategy thread was closed I'll just put this here. Which one do you guys prefer?
Plan A
$499 single SKU, 8c/16t 3.0 ghz cpu, 14tf Navi, 24gb gddr6 + 8gb ddr4, 1TB hdd

Plan B
Dual SKU launch
(Weak sauce SKU) $399 8c/16t 2.6 ghz cpu, 10tf Navi, 16gb gddr6 + 8gb ddr4, 500GB hdd
(Beast SKU) $599 8c/16t 3.4 ghz cpu, 16tf Navi, 32gb gddr6, 1TB hdd

Plan A to me is the far better choice, simple, very powerful baseline and not over the top price point.

 

[BRiT]

But what if Mfg 1 does plan A while Mfg 2 does plan B?

 

[ultragpu]

Pardon my ignorance but what Mfg do you speak of?

 

[BRiT]

Pardon my ignorance but what Mfg do you speak of?

Imaginary, since we're talking about imaginary console builds anyways. So one manufacturer does Plan A while a different one does Plan B. Which one is the better plan now?

 

[ultragpu]

Plan A still, it's just much more streamlined and offers better looking baseline exclusives. Plan B isn't too far off at this point tho since the baseline would still be powered by a 10tf machine so in either case we get a noticeable generation leap.

 

[AlNom]

If your criteria hinges upon preference for exclusives, nothing about the hardware matters.

Just get the system you want to play the games. End-of-story. Move on.

 

[mrcorbo]

Since the other hardware release strategy thread was closed I'll just put this here. Which one do you guys prefer?
Plan A
$499 single SKU, 8c/16t 3.0 ghz cpu, 14tf Navi, 24gb gddr6 + 8gb ddr4, 1TB hdd

Plan B
Dual SKU launch
(Weak sauce SKU) $399 8c/16t 2.6 ghz cpu, 10tf Navi, 16gb gddr6 + 8gb ddr4, 500GB hdd
(Beast SKU) $599 8c/16t 3.4 ghz cpu, 16tf Navi, 32gb gddr6, 1TB hdd

Plan A to me is the far better choice, simple, very powerful baseline and not over the top price point.

From what perspective am I evaluating this? As a consumer considering buying one of the SKUs, as a wannabe product planner trying to predict which will sell better, or ???

As a consumer, I will almost always prefer more (within reason) choices.

 

[Proelite]

What's up with the ridiculous amounts of DDR4. 8GB can enable a full desktop Windows 10 PC experience.
If they have separate OS ram I don't expect more than 4GB in the highest scenario.

My plausible SKUS:

Base:
$299, 8c/16t 2.4 ghz CPU, 8tf Navi, 16GB GDDR6 on 256 bit bus, ~450 GB/s, 0-2GB DDR4, 500GB HDD

Premium:
$499, 8c/16t 3.0 ghz CPU, ~15tf Navi, 24GB GDDR6 on 384 bit bus, ~760 GB/s, 0-2GB DDR4, 2TB HDD

Ultra:
Some as premium but with 2TB SDD

 

[ultragpu]

From what perspective am I evaluating this? As a consumer considering buying one of the SKUs, as a wannabe product planner trying to predict which will sell better, or ???

As a consumer, I will almost always prefer more (within reason) choices.

From a consumer perspective of course. Sometimes more choices don't mean the best choices tho.

If your criteria hinges upon preference for exclusives, nothing about the hardware matters.

Just get the system you want to play the games. End-of-story. Move on.

This does matter to a great extent admittedly but a balanced console is equally as important.

 

[ultragpu]

What's up with the ridiculous amounts of DDR4. 8GB can enable a full desktop Windows 10 PC experience.
If they have separate OS ram I don't expect more than 4GB in the highest scenario.

My plausible SKUS:

Base:
$299, 8c/16t 2.4 ghz CPU, 8tf Navi, 16GB GDDR6 on 256 bit bus, ~450 GB/s, 0-2GB DDR4, 500GB HDD

Premium:
$499, 8c/16t 3.0 ghz CPU, ~15tf Navi, 24GB GDDR6 on 384 bit bus, ~760 GB/s, 0-2GB DDR4, 2TB HDD

Ultra:
Some as premium but with 2TB SDD

Just in case they bloat up the next gen OS to a ridiculous level but I guess I'll agree to the 4GB amount.

Spoiler

When can we expect some new leaks?