Predict: Next gen console tech (9th iteration and 10th iteration edition) [2014 - 2017]

Status
Not open for further replies.

Intel calls it Embedded Multi-Die Interconnect Bridge. Seems like it would allow discrete CPU and GPU while using less board space. So maybe that discrete gpu for PS5 rumour wasn't so far-fetched?
 

Intel calls it Embedded Multi-Die Interconnect Bridge. Seems like it would allow discrete CPU and GPU while using less board space. So maybe that discrete gpu for PS5 rumour wasn't so far-fetched?
You could do that without EMIB too, that's not the point.
 
It's probably good news for AMD, but I don't think we'll see this combination in the next generation of consoles. Maybe the generation thereafter, once the relationship has matured and everyone knows they're doing more than test the waters.

The main thing I took notice of is the usage of HBM2: it's suitable for a portable device, and AMD appear to be committed to it.

Something else to make me hope for a two tier PS5 launch: a portable, 1080p focused model, and a full fat, non-portable, 4K machine. HBM3 in both, micro SD card/SSDD as cache (portable/home console respectively.)
 
I really want to see how that Zen APU that launched a week or so back performs, but proper reviews seem none existent.
 
Yeah I guess simply switching from GDDR5 to HBM is really what saving the space. Either way care to elaborate?
MCM chips in general can be done without much special techniques, they've been used forever in embedded and mobile spaces, EMIBs biggest merit is replacing interposer for HBM/GPU-connection
 
I reckon you're underestimating, I bought an iMac in 2013 which a 4Gb Geforce 780M. I replaced it this year with a 5K iMac with 8Gb Radeon Pro 580. If Apple are shipping those cards, the PC world has been shipping them longer.
You're correct, the amd 290 also had 4GB. Still, that's half of the Ps4's.

Honestly I think the current ram prices are just massive price gouging and they'll be forced to come down within the next year or two.
 
Specs/benchmarks of Kaby Lake G Series with integrated discrete radeon gpu's:
https://wccftech.com/intel-kaby-lak...-radeon-gpus-first-benchmarks-specifications/

- 3 variants
- CPU is 4 core 3.1Ghz
- GPU has 24 compute units @~1011MHz-1190MHz depending on variant
- 4GB HBM

Kaby-Lake-G-Specs.jpg
 
32 GB HBM3 could be cheaper than 32 GB HBM2 and offer the same bandwidth (~1TB/s) and have the benefit of power savings compared to GDDR. HBM3 is likely going to end up in many more devices than HBM2.

Hynix-HBM3-900x505.png



Of course, HBM3 scales beyond HBM2, allowing upto 64 GB @ 2 TB/s bandwidth, which should come into play for AMD's large scale APUs / chiplets / MCMs, etc, for HPC and the ultra enthusiast graphics cards during 2019-2021+, encompassing both Navi GPU architecture and post-Navi "Next Gen".

Samsung-HBM3.jpg


iEue8i2.png
 
Last I read about it, the cost issue with HBM remains the inability to fully test the device, the yield depending in part on the number of tsv, and that yield causes losses of the complete assembly, not just the defective die. I think we should watch closely for the apearance of HBM on lower cost GPUs.

The companies who need it are currently paying whatever it cost because there are no alternatives at the high end. But that's temporary while production ramps up I guess.
 
Last edited:
I agreen
32 GB HBM3 could be cheaper than 32 GB HBM2 and offer the same bandwidth (~1TB/s) and have the benefit of power savings compared to GDDR. HBM3 is likely going to end up in many more devices than HBM2.

Hynix-HBM3-900x505.png



Of course, HBM3 scales beyond HBM2, allowing upto 64 GB @ 2 TB/s bandwidth, which should come into play for AMD's large scale APUs / chiplets / MCMs, etc, for HPC and the ultra enthusiast graphics cards during 2019-2021+, encompassing both Navi GPU architecture and post-Navi "Next Gen".

Samsung-HBM3.jpg


iEue8i2.png


Why would we need 1TB/s of bandwidth?

Assuming Sony make an APU with 4096sp in the gfx part... It won't need a Terabyte of bandwidth.
 
Last edited:
Stop thinking in current gen terms. I want fully dynamic voxel cone traced global illumination, multi-layer g-buffers, volumetric fluid sim, etc... Go see how much memory/bandwith that shit takes.
I just want good quality alpha and shadow effects again. And AF. And no/minimal pop ins. Basic shit. :p

All that high end stuff would get in the way of that xD
 
Status
Not open for further replies.
Back
Top