Reposting this because I believe the source rumor to be very credible.
That one post at reddit by a throwaway account that PS5 uses HBM + DDR4 after the Wired article is either true or made by a madman with shit ton of time on his hands.
As someone stated here earlier, Sony has patented a heatsink last year to work on a stacked die on top of an APU, which is how HBM chips would be at. The 1.6 HBM chips could very well be similar to the low cost HBM chips Samsung has been developing and revealed in 2016 that a GPU made for laptops could be used;
It's not like AMD doesn't have experience using HBM2 chips on their GPUs either.
Although, the heatsink could be from a PS5 prototype that won't see the light of day, as well as Cerny has been championing unified RAM so...
The reddit post actually went up shortly
before the Wired article.
Another interesting point to make regarding this rumor is that TSMC actually consulted with Sony (among others) during the development of InFO_MS, which would make Sony familiar with it and probably in the back of their mind when developing consoles. I'd also like to point out that Sony is not a stranger to being inventive with memory. The Vita had a
special packaging method in use to get very high bandwidth to the Vita's GPU.
Also, let's assume for a second there was a typo in the original rumor, because it doesn't quite make sense as it is. It says:
allowed them to go below ~50 GFLOPs per GB/sec. bandwidth but still keep above 40 GFLOPs per GB/sec.
Which doesn't make sense. Typically when we talk about GPUs, we talk about GB/s per GF/TF, not the other way around. However, 40GB/s per GF would be a stupidly low number, so let's now assume they meant TF. We need a second part of the comment for future information
InFO_MS allows them to drive their 1.6 Gbps chips @ 1.7 Gbps (435 GB/sec.) without having to increase the voltage above 1.2v
1.6Gbps per second would result in 409.6GB/s in a 2-stack config. 1.7Gbps bumps it to 435.2GB/s, which is competitive with 256-bit GDDR6 solutions. If we're now assuming the 40-50 GB/s per TF is valid, this gives us a range of
8.7TF to 10.875TF for PS5.
EDIT: if you add DDR4 bandwidth, it goes up further
anexanhume
If you tack on the DDR4 bandwidth to the total (yielding 537.6 total), doing the GB/s per TF calculation gets you a range of 10.75TF to 13.44TF.
I didn't notice before, but the reddit post indicates a quad-channel DDR4 interface @ 3200 MT/s. That's quite substantial, especially when you consider desktop Ryzen is paired with dual-channel DDR4. Maybe allotting extra bandwidth for overhead from moving data between the two pools and for the GPU to use directly when the GPU and CPU are working on the same data?
Finally, regarding my skepticism around HBM2 supply, some key things have happened since Samsung's comments about low capacity. SK Hynix and Micron have both entered the market in full force (after the latter abandoned HMC development). And the crypto market crashed. With DRAM
and NAND markets easing up, that capacity has to shift somewhere.
Regarding HBM pricing, it's hard to know much it has eased over the past few years (we do have some Vega VII rumored costs for reference), but I think it may be possible to get it down to less than 50% more than GDDR6 per GB, perhaps even just
35% to 40% higher. When you consider that they just need 8GB instead of 16GB of GDDR6, their solution is extremely cost competitive. At that point, it becomes a lot more attractive. HBM is also done on contract pricing (i.e. not floating with market costs), so a big order from Sony locks that factor in and sets up a mutually beneficial relationship with that partner to help them build up their own capacity.
The only rumor around this giving me pause is the digitimes rumor that stated
ASE will do the packaging. Other than that, a lot of this rumor makes sense the more I dig into it.
Also, I imagine if that PS4 rev mentioned is coming, it's definitely this Fall. Since it's a console rev, it may not get cracked open to confirm the 7nm EUV from Samsung part, but the timing makes so much sense with MS pushing costs down with the SAD model and the rumored E-revision of the device internals.
Finally, here's the
rumor in its entirety for posterity:
PS4 refresh
- sometime between september and november
- 199
- fabbed on samsung 7nm EUV
- best wafer pricing in the industry
- die size 110mm²
- no PRO refresh, financially not viable yet
- too close to PS5 as well
PS5 memory and storage systems
- 24 GB RAM in total (20 GB usable by games)
- 8 GB in form of 2 * 4-Hi stacks HBM2
- Sony got "amazing" deal for HBM
- in part due to them buying up bad chips from other customers which can't run higher then 1.6 Gbps while keeping 1.2v.
- HBM is expected to scale down in price a lot more than GDDR6 over the console lifetime
- Samsung, Micron and SK Hynix already shifting part of their capacity towards HBM due to falling NAND prices
- Sony will be one of the first high volume customers of TSMCs InFO_MS when mass production starts later this year (normal InFo already used by Apple in their iPhone)
- InFO_MS brings down the cost compared to traditional silicon interposers - has thermal and performance advantage as well
- InFO_MS allows them to drive their 1.6 Gbps chips @ 1.7 Gbps (435 GB/sec.) without having to increase the voltage above 1.2v
- HBM is more power efficient compared to GDDR6 - the savings were invested into more GPU power
- additional 16 GB in form of DDR4 @ 256 bit for 102.4 GB/sec.
- 4 GB reserved for OS, the remaining 12 GB usable by games
- memory automatically managed by HBCC and appears as 20 GB to the developers
- HBCC manages streaming of game data from storage as well
- developers can use the API to take control if they choose and manage the memory and storage streaming themselves
- memory solution alleviates problems found in PS4
- namely that CPU bandwidth reduces GPU bandwidth disproportionately
- 2 stacks of HBM have 512 banks (more banks = fewer conflicts and higher utilization)
- GDDR6 better than GDDR5 and GDDR5x in that regard but still less banks than HBM
- at the same time trying to keep CPU memory access to slower DDR4
- very satisfied with decision to use two kinds of memory for price to performance reasons
- allowed them to go below ~50 GFLOPs per GB/sec. bandwidth but still keep above 40 GFLOPs per GB/sec.
Also, this DF video came out today:
