Dec.17th; Dr.Lisa Su's keynote at CES 2019, where it is expected to get news about AMD's 7nm consumer CPU and GPU line-ups, will be taking place January 9th, in a one hour presentation:
So the previous thread was locked and I decided to create a new one.
What's new? Microsoft made the first public move by actually announcing that they're working on their next console:
Amazing post by anexanhume in the previous thread a few weeks before E3 2018. Thanks to Socky by pointing it out:
So the previous thread was locked and I decided to create a new one.
What's new? Microsoft made the first public move by actually announcing that they're working on their next console:
Amazing post by anexanhume in the previous thread a few weeks before E3 2018. Thanks to Socky by pointing it out:
I will try to construct a summary of the past few months. Others have hit on a few points already, but I'll try to make it all self-contained.
Backwards compatibility
1. Sony has made comments that indicate that BC could be present in next gen.
2. There are roughly 5/6 patents with Cerny's name attached that go into methods of altering how a processor works to enable backward compatibility with previous architectures.
CPU
1. AMD CEO Lisa Su has made comments as recent as last week that confirm Zen 2 is design complete and will sample this year, with high volume next year. This is both server and desktop versions.
2. We do not know if it will be manufactured in TSMC or Global Foundries' 7nm.
3. We also don't have much architecture insight into Zen 2. It is believed that CCX will go up to 6 cores in Zen 2, which means multiple CCX would be required for an 8 core console APU.
4. Someone posted Ryzen power consumptions vs. frequency graphs a few pages back that show that the power scales down quite nicely around the 3 GHz mark, supporting a 25-35W CPU TDP in a likely console APU budget.
5. LLVM (compiler) improvements have been spotted for Zen version 1 from a verified Sony SN Systems engineer. Zen 2 hasn't been spotted yet, so it's possible the design similarities or NDAs could explain this if Zen 2 is still the target.
6. I've spotted AVX512 commits from the same engineer. AVX512 is not currently supported on any AMD processor, but Zen supports AVX2 (256) by 2-cycling it's 128 bit units. Jaguar had AVX support (128 bit wide). AVX really accelerates vector workloads and can be optimized to (and is more general purpose with more register space despite how general purpose GPUs have become), but it can be a big hit on die space and power consumption.
7. Zen 3 design is underway and is targeted 2020 in 7nm+ per current roadmaps. Adoption of this design would be very aggressive, even with a 2020 launch date. Of course, customizations are always possible, as seen with GPUs.
GPU
1. Vega is not expected to make it to consumers in 7nm at all. Strictly for the AI/compute/research crowd.
2. Navi is shown on AMD's roadmap for 2019, though we have few details other than the focus is on scalability and it will be the last GCN iteration.
3. A patent for a "Super SIMD" design was recently discovered, and the concept sounds very similar to Nvidia's latest designs. It's unclear if this would be Navi or post-Navi.
4. Navi's successor is 2020 per AMD roadmap, but would be very aggressive to include. Customizations are always possible, of course.
5. AMD has stated that they've brought some Zen designers over to the GPU team in the interest of improving efficiency, since Zen was a big step forward compared to excavator and piledriver. It's unclear when we can expect to see the fruits of that effort.
APU
1. Known die scaling factors for 16nm to 7nm suggest that an 8 core Zen, 64 CU (Vega chips) could fit in an area around 350mm^2. This is where original Xbox One and PS4 APUs were at. Depending on clocks, this is a 10-14 TF GPU.
2. The idea of MCMs with GPU chiplets has been tossed around due to smaller die costing less, but the cost structures and Infinity Fabric penalties likely don't make sense in a console budget.
Memory
1. GDDR6 will hit mass production this year and is being supplied by all major memory suppliers.
2. GDDR6 spec goes up to 18Gb/s per pin, and has been demonstrated up to 20Gb/s by Micron. This is on the order of 550GB/s for a 256-bit bus ( a la PS4 and Pro), or 770GB/s on a 384-bit bus (Xbox One X).
3. 384-bit bus is problematic from a cost standpoint because it necessitates at least 12 chips, and is a larger PCB. It also can keep the die size larger, which means it can't be shrunk as much.
4. All three major manufacturers have been accused of price fixing, and pricing has relaxed a little lately. The crypto craze is also lessening demand as it tapers off. Lisa Su commented she believes the consumer aspect of this will dwindle.
5. HBM remains a possibility, but there are issues with cost, yield, testability, and access granularity according to industry trends and Microsoft's insight from designing the Xbox One X. Cost could be near $100 higher than GDDR6 for a similar capacity solution according to an article I posted before. TSV and interposer are the cost drivers.
6. Micron recently became a HBM supplier, so that would help cost and availability some.
7. Current HBM timelines show HBM3 could be a possibility, but HBM2 would be more likely.
8. Capacity likely between 16-32 GB. GDDR6 or HBM could support these capacities. The solution could be anywhere from 8-16 chips in GDDR6 or 2-4 stacks in HBM.
Non-volatile storage
1. AMD's HBCC has been getting some attention for its potential to act like a cache manager between main and non-volatile memory. This could allow for a modestly-sized NAND drive in the realm of 64-128GB where the OS and relevant game data can sit to reduce load times and enhance system responsiveness.
2. AMD claimed up to a 50% reduction in VRAM footprint using HBCC, but that's yet to be realized in real-game benchmarks.
3. SSDs are finally reaching a $/GB level where they could be in the discussion for console usage. Perhaps multiple SKUs or replaceable drives, as seen in PS4.
Foundry Technology
1. TSMC and Global Foundries' 7nm both would be ready for a 2019 launch.
2. 28nm had a 1+ year gap from HVM to console APUs, which is necessary from a cost, yield and throughout perspective given the larger die (mobile devices around ~100mm^2 tend to be first).
3. Current 7nm+ (EUV) timelines put TSMC ready for HVM within 2019. This could support a 2020 console, but would be very aggressive. The gains over 7nm are 10-15% performance/heat/shrink, but the processing is also easier due to fewer mask steps.
Timing
1. Sony has explicitly stated no hardware announcements at E3 this year.
2. Sony recently shared FY guidance that showed a huge profit hit in March 2021 (2020's results). They purposefully did not disclose March 2020, so they obscured the exact launch date, but guaranteed 2019 or 2020.
3. Microsoft has been fairly quiet, but have made statements that show their investment in console gaming is still strong.
4. Sony has stated that PS4 has entered into the end of its life cycle, which matches with the above implications of the FY guidance indicating a launch within the next two years.
VR
1. There have been a ton of patents in VR with and without Mark Cerny's name attached. Many relate to foveated rendering and other optimizations.
2. Sony has said that VR growth has fallen below industry expectations across the board, including them. They identified it as a growth area to improve upon, so they still seem fully invested and I expect PSVR2, as do most I think.
Operating system
1. The aforementioned Sony engineer's name has turned up with regards to FreeBSD commits as well (recent versions). This could mean that their OS of choice for PS3 and PS4 could stick around as the basis for PS5.
Financial considerations
1. Sony showed an up to 400 million profit hit in their March 2021 results, which would be a 30% drop. They only dropped 1% with the launch of PS4. Given that the hardware and packaged software sales only make up 45% of their current revenue model, and that PS+ and digital comprise the rest and have been growing, some are taking this to mean Sony could be heavily subsidizing PS5 costs.
2. Launch price is assumed $400-$500. There are compelling cases for the entire range.
Last edited:
