Next Generation Hardware Speculation with a Technical Spin [pre E3 2019]

Status
Not open for further replies.
What you will get in consoles is at most 14gbps ram. 16 is laready stretching, 18 is a no-go (pw consumption).
Burning a couple of more watts is not going to kill any kind of system budget. The biggest impediment would be yield or cost for the higher binned modules.
 
Samsung said their 20gbps test was bypassing the actual dram, because current process cannot yield a fast enough dram array. Not even for a test chip. So they only tested the interface.

Maybe 18gbps needs to wait for process improvements to get a reasonable percentage of chips passing that speed.
 
Do you guys think the bare min will be 16GB for next-gen lower end models or will it match X's 12GB? Obviously the speed/bandwidth will be faster than the X.
 
Samsung announced mass production for up to 18Gbps over a year ago.

https://www.anandtech.com/show/12338/samsung-starts-mass-production-of-gddr6-memory

Only 16Gbps is visible on their site, though.
Ok, this is tricky. They actually didn't announce the mass production of 18Gbps GDDR6 in January 2018. They only announced mass producing 16-Gigabit GDDR6 and only talked about 18Gbps speeds available eventually.

That's their actual announcement:
Samsung Electronics Starts Producing Industry’s First 16-Gigabit GDDR6 for Advanced Graphics Systems
Beginning with this early production of the industry’s first 16Gb GDDR6, we will offer a comprehensive graphics DRAM line-up, with the highest performance and densities, in a very timely manner

And then in the article they talk about the max available speed [eventually available].

https://news.samsung.com/global/sam...6-gigabit-gddr6-for-advanced-graphics-systems

The anandtech author (and everybody else) got tricked by the dubious wording. But I think they actually started mass producing those speeds relatively recently, I read an article about it, but I can't find the source anymore.
 
Last edited:
Ok, this is tricky. They actually didn't announce the mass production of 18Gbps GDDR6 in January 2018. They only announced mass producing 16-Gigabit GDDR6 and only talked about 18Gbps speeds available eventually.

That's their actual announcement:



And then in the article they talk about the max available speed [eventually available].

https://news.samsung.com/global/sam...6-gigabit-gddr6-for-advanced-graphics-systems

The anandtech author (and everybody else) got tricked by the dubious wording. But I think they actually started mass producing those speeds relatively recently, I read an article about it, but I can't find the source anymore.

Thanks for clearing that up!
 
If Sony or MS ship a 8c/16t Zen2, I'd bet money on them using the exact same chiplet that's in the CPUs. Sure, it costs some extra power for the link between the CPU and the IO/GPU chiplets, but using the same chiplet for the CPU means they get to bin them, from the set of all the Zen2 chips AMD will make. This buys them more performance and less power and most importantly, much lower cost than shipping a huge monolithic chip that's bust if there is a single flaw in a cpu core.
It may be more complicated than optimum power/performance since using the same chiplet also gets them a major bidding competitor: AMD's own much more lucrative products.
A chiplet that performs well at low power could go into semi-custom products that might be in the teens for margins, or go into a server SKU that sells for an order (orders?) of magnitude more.

So maybe they can buy the most mediocre-clocked ones that are just okay for power?

I think the GPU could be smaller than the VII, because of no need for 64-bit support, but broadly in the same class. Assuming that this launches for the holiday sales in 2020, 330mm^2 of 7nm would not be terrible.
It looks like half of Vega20's area is everything but graphics, so a good chunk of xGMI, infinity fabric, and high-end controllers could be removed.
If the GPU's are backwards-compatible at a hardware level to the Sea Islands architecture--and the customized ones in the consoles in particular, it might have an area cost that Vega 20 didn't have.

I don't believe the supposed leak, but wouldn't it be possible for the first iteration to use GDDR6 on a 512-bit bus, then transition to a couple of stacks of HBM3?
That looks like an expensive lateral move. The amount of bandwidth cannot significantly change for consistency reasons, but the standards likely don't behave identically in ways that may cause games to diverge in performance between consoles. At least for mainline consoles, the core elements are something the vendor has to commit to.
 
... so you're confirming 16 GB on a 256-bit bus, clocked at 14 gb/pin, so they can match clocks with dev kits running with more memory in clamshell mode?

Right, I'm off to the bookies!
Yeah I'm having too much fun "exploring" fringe possibilities, but before the crazy leak talk, most predictions have been narrowing it down to almost a concensus somewhere around 10TF and 16GB/256bit.

It seems to be at a nice "knee" of the diminishing returns from raising the BOM. Anything more gets expensive really fast for not much more performance. And anything less means next gen isn't exactly worth spending hundreds of millions developping a new platform.

But... We also thought 8GB for the ps4 would be too expensive, so I don't know if such a surprise is possible again.

(the reason something is a surprise is because the chances were slim)
 
Last edited:
Let's remember that current gen came in at low-mid PC specs of the time, with the exception of a relatively large pool of memory enabled by low pricing which may not be as favorable again (and MS still dropped the ball). This time we're getting lucky with the CPU chiplet but I wouldn't expect an expensive GPU/RAM combo.
 
Let's remember that current gen came in at low-mid PC specs of the time, with the exception of a relatively large pool of memory enabled by low pricing which may not be as favorable again (and MS still dropped the ball). This time we're getting lucky with the CPU chiplet but I wouldn't expect an expensive GPU/RAM combo.
PS4 GPU was between a 7850 and 7870. That’s not low-mid. Only enthusiast GPUs were above it (7950, 7970).

The rumored forthcoming Navi 10 and Navi 12 are in that tier, with Navi 20 to follow later.
 
290x (and 780 Ti) came out in the same month. Plus the massive CPU deficit.
And a lag of some features is to be expected given when the design is finalized. I don’t think anyone would dispute the CPU claim, which is why I didn’t address it. I also don’t think anyone would dispute that Zen 2 should be a huge leap forward, and actually relevant to current desktop CPU performance.
 
Let's assume $400-450 is the price target. Let's also assume that 350mm of 7nm die in 2020 likely costs more than 350mm of 28nm in 2013. You may be able to balance that by decoupling and binning the CPU chiplet, which also improves yield on the GPU die. Vega 20 is 330mm without the chiplet IF link a console version would need. Without density improvements on Navi, at best you're looking at 75% of Vega 20 die size, and 75% of the frequency (< 150W console TDP). That could be < 8 TFLOPS, and definitely < 10.

I'm not even factoring in the larger memory costs, or the need for NAND.
 
Let's assume $400-450 is the price target. Let's also assume that 350mm of 7nm die in 2020 likely costs more than 350mm of 28nm in 2013. You may be able to balance that by decoupling and binning the CPU chiplet, which also improves yield on the GPU die. Vega 20 is 330mm without the chiplet IF link a console version would need. Without density improvements on Navi, at best you're looking at 75% of Vega 20 die size, and 75% of the frequency (< 150W console TDP). That could be < 8 TFLOPS, and definitely < 10.

I'm not even factoring in the larger memory costs, or the need for NAND.

The memory costs won’t be larger if they only go with 16GB.
 
Let's assume $400-450 is the price target. Let's also assume that 350mm of 7nm die in 2020 likely costs more than 350mm of 28nm in 2013. You may be able to balance that by decoupling and binning the CPU chiplet, which also improves yield on the GPU die. Vega 20 is 330mm without the chiplet IF link a console version would need. Without density improvements on Navi, at best you're looking at 75% of Vega 20 die size, and 75% of the frequency (< 150W console TDP). That could be < 8 TFLOPS, and definitely < 10.

I'm not even factoring in the larger memory costs, or the need for NAND.

I follow your line of reasoning, but, sincerely, if Sony releases a lower than 8tf next-gen machine, next year.... then better not releasing anything imo.
 
Looking at the latest consoles from Sony and MS (PS4>PS4 Pro> ONE X) you get 30-35 % increase in TFLOPS each year. I just cant see going from ONE X 6 TFLOPS to 8 TFLOPS in 2-2,5 years. I think we will see 10+.
 
Status
Not open for further replies.
Back
Top