AMD Vega 10, Vega 11, Vega 12 and Vega 20 Rumors and Discussion

ImSpartacus said:
I'm a little suspect of that.

If you follow the sources through, you go to:

WCCFTech 5/15/17 http://wccftech.com/amd-radeon-rx-v...lineup-includes-3-skus-priced-at-599-499-399/

TPU 5/15/17 https://www.techpowerup.com/233325/entire-amd-vega-lineup-reportedly-leaked-available-on-june-5th

"Digiworthy" 5/14/17 http://digiworthy.com/2017/05/14/amd-rx-vega-lineup-fastest-vega-nova/

And that's curiously close to the baseless post on reddit documented here: https://forum.beyond3d.com/posts/1980465/

So my guess is that this "Digiworthy" site just copied the reddit rumor.
Click to expand...

Yes it seems it is based on a false rumor, sorry
 
SK Hynix updated its databook with the latest information about HBM2 and GDDR6 memories.
HBM2 was first expected in the third quarter of 2016, only to be pushed back to Q1 2017 six months later. The latest databook suggests that HBM2 will be available this later quarter. It’s important to note that it does not say ‘Available: Now’, which means HBM2 was not available at the time of making this databook (May 5th to be exact). This might be the main reason why there is no Vega yet.

Assuming that Radeon RX Vega was to feature HBM2 from SK Hynix, then we should expect a minimum bandwidth of 410 GB/s. Of course, this is just an illustration, Vega could easily have 4 stacks which would double the bandwidth and capacity.
...
Now let’s move on to products that are still far far away. SK Hynix currently expects GDDR6 memory to be available in the fourth quarter. The datasheet confirms that only two variants are in the making, both 8Gb (1GB). The difference is hidden in the frequency, the faster variant is to feature 14 Gbps interface, while the other is estimated at 12 Gbps. This means that at best we should get 672 GB/s bandwidth with 384-bit bus or 448 GB/s with the 256-bit bus. That’s a significant increase over GDDR5-based solutions.
Click to expand...
https://videocardz.com/69504/sk-hynixs-updated-memory-catalog-features-hbm2-and-gddr6
 
Something about Vega being teased for today by RTG head Raja Koduri and Sr. Marketing Director Chris Hook, as per some previous rumors:

https://twitter.com/GFXChipTweeter/status/864347035033522178
Raja Koduri said:
Ok..complete the sentence..we stand today on the edge of a ...... (first 10 people who get this right will get something special) @radeon
Click to expand...

https://twitter.com/GChip/status/864334949263986688
Chris Hook said:
@GFXChipTweeter Riddle me this: When I'm wild, Davy Crockett is king, and while I'm also an airline, some say I'm really a Vega #sweet16
Click to expand...




CSI PC said:
Right,
so it is a different reason when it is nothing to do with AMD.....
Click to expand...
Anyone reading this topic has become well aware of how you choose to interpret or form an opinion about the oncoming information (inclination is always the same).
But it's the jabs like this one I'm replying to that seem to really bring out the true nature of the whole thing.


So all those leaks showing 1200MHz core clock for Vega in early benchmarks are valid and you were convinced Vega couldn't clock over 1.3GHz. However, this last Compubench public score showing 1.6Hz core, together AMD showing slides claiming Vega is optimized for significantly higher clocks and their own Mi25 card being announced with 12.5 TGLOPs FP32 are all probably bogus.

Beginning of Micron's GDDR5X mass production was kept a secret until the formal announcement of the GP100, but 8-Hi stacks HBM2 being made by SK Hynix can't possibly be true even though there's a Compubench score of a Vega card showing 2 stacks of 8GB for a total of 16GB VRAM.

Apple made AMD hide the full Polaris 11 @ 35W from their portfolio so only the cut-down and lower-binned chips (destkop RX460) could be mentioned until the formal announcement of the Macbook Pros some 3 months later, but for you it's impossible for SK Hynix to hide 2Gbps HBM2 for a couple of quarters to give AMD the chance to formally present it to the public - even though it was already in their portfolio.
 
This is interesting, hope all of it is true. 1600MHz core clock and 16 gigs of HBM2. Although 16 gigs of vram is a bit of an overkill for games it might be very useful for other applications :)
 
Clukos said:
This is interesting, hope all of it is true. 1600MHz core clock and 16 gigs of HBM2. Although 16 gigs of vram is a bit of an overkill for games it might be very useful for other applications :)
Click to expand...

But I guess, with HBM2, it either that, or "only" 8 ? They can't go 11 or 12 like nVidia, right ?
 
Clukos said:
This is interesting, hope all of it is true. 1600MHz core clock and 16 gigs of HBM2. Although 16 gigs of vram is a bit of an overkill for games it might be very useful for other applications :)
Click to expand...

Hence why i will have first think that thoses benchmark are linked to the MI25 or a similar card.. 16GB should be the minimum for this type of card.
Now is there a possibility that the MI25 have a 4Hi x4Gb stack ( so different interposer ) that the Vega 10 "gaming part" ? ( or basically this could explain too why vega 20)
 
Last edited:
ToTTenTranz said:
....
Anyone reading this topic has become well aware of how you choose to interpret or form an opinion about the oncoming information (inclination is always the same).
But it's the jabs like this one I'm replying to that seem to really bring out the true nature of the whole thing.


So all those leaks showing 1200MHz core clock for Vega in early benchmarks are valid and you were convinced Vega couldn't clock over 1.3GHz. However, this last Compubench public score showing 1.6Hz core, together AMD showing slides claiming Vega is optimized for significantly higher clocks and their own Mi25 card being announced with 12.5 TGLOPs FP32 are all probably bogus.

Beginning of Micron's GDDR5X mass production was kept a secret until the formal announcement of the GP100, but 8-Hi stacks HBM2 being made by SK Hynix can't possibly be true even though there's a Compubench score of a Vega card showing 2 stacks of 8GB for a total of 16GB VRAM.

Apple made AMD hide the full Polaris 11 @ 35W from their portfolio so only the cut-down and lower-binned chips (destkop RX460) could be mentioned until the formal announcement of the Macbook Pros some 3 months later, but for you it's impossible for SK Hynix to hide 2Gbps HBM2 for a couple of quarters to give AMD the chance to formally present it to the public - even though it was already in their portfolio.
Click to expand...

Yeah right so secret....
Here is what I said 2 days before the Micron announcement 10th May announcement (which was 4 days before the PCGamer 'news about secret production' 12th May) and my info was referencing information that went back months before that with it having production status and contact for order:

May 8th 2016:
https://forum.beyond3d.com/threads/nvidia-pascal-announcement.57763/page-33#post-1912442
CSI PC said:
Well I checked and it seems the custom AIB for the 980 were just under 4 weeks later than the reference design.
If it goes over that then maybe supply issues, but I do not think that is an issue because I checked the status of the 10Gb/s GDDR5X awhile back and it was in production status.
Even the 12Gb/s went into sampling status back in early March, so not long until that can start to be ramped up for production (that will have supply issues but fingers crossed it will be used on the custom AIBs 1080 when it does become available)
Cheers
Click to expand...
Production status means available to calendar order.
12Gbps never made it further than Sampling status for a long while (unfortunately only hit production status around March 2017) and so my expectation on that was wrong, but shows reality and challenges of ETA/estimated spec into official catalogue or the parts reference and then importantly to Sampling status, and then critically to Production status.

And I said it looked more likely the clocks would be 1300-1350MHz for the reasons I gave, not as you infer that it would not even hit 1300MHz.
And this is sustained 1300-1350MHz rather than as I pointed out (either that post or earlier) what happens with Nano relative to its official TFLOPs/Boost that is not sustained or only very briefly hit unless one overclocks/tweaks it.
This is ignoring the massive power demand it takes Nvidia to push the 1080ti even towards 1900MHz (post you were referring to also was considering relative clock changes) and this is a smaller core/die than Fiji in an area Nvidia is well known for efficiency and operating at high clock speeds across all GPU sizes.

Anyway back to current subject, Micron GDDR5X production status or contact Micron for order was never kept secret to tie in with Nvidia product.
Nor would I expect SK Hynix to keep it secret when it comes to their catalogue/parts reference that goes to end of year for schedule and ordering, both would do announcements that aligns with products using their component but that is different to the parts book.

But happy to drop this until we get the facts from AMD rather than 'leaks'.
 
Last edited:
Will the price of HBM2 ever experience economies of scale if Nvidia decides to stay on the GDDR path for consumer cards? With Samsung and Micron ramping up GDDR6 production the likelihood of widespread standards acceptance/rollout and lower prices for HBM2 seems less each day.
 
SpaceBeer said:
Unless there are other products in which AMD, Samsung, Intel... could implement HBM (HPC, servers, workstations, ultra portable laptops, SFF PCs, mobile devices...).
Click to expand...

As mentioned a couple of pages ago, with HBM's interposer requirement, AMD isn't selling only the GPU in a substrate to OEMs.
AMD is selling the GPU, the HBM stacks and the Interposer all in one package.

AFAIK, there are only two companies with HBM in their roadmaps: AMD and nvidia. Nvidia for GPUs, AMD for GPUs and APUs.

So if an OEM wants to build a Vega graphics card, they have to buy the GPU+HBM from AMD. If a HPC builder wants to use high-performance APUs with HBM, they'll have to buy APU+HBM from AMD. Same thing for a laptop maker if they make a Raven Ridge successor with HBM.

Intel's architectures don't support HBM and aren't likely to support it in the future either (unless their own HMC really crashes and burns).
There are talks about Intel planning a MCM carrying a Core ix and a Radeon GPU, but if said GPU uses HBM then it needs to come in an interposer so again AMD would be selling the HBM together with the GPU in the interposer.


HBM isn't tailored for handhelds either, as its JEDEC ultra-low power counterpart is Wide I/O.

The only possible candidate for HBM is company trying to build an ARM-powered HPC/server chip using that memory, but ARM chips haven't been awfully popular in that area.
Perhaps if/when apple decides to ditch x86 and replace them Intel cores with their ARM windy stuff, they'll be looking to use the best-of-the-best in for their high-margins laptops and AiOs. But that's still a long time away IMO.





So the short answer to your question is: no, there doesn't seem to be anyone else other than AMD wanting to buy HBM from SK Hynix at the moment.
 
ToTTenTranz said:
Intel's architectures don't support HBM and aren't likely to support it in the future either (unless their own HMC really crashes and burns).
There are talks about Intel planning a MCM carrying a Core ix and a Radeon GPU, but if said GPU uses HBM then it needs to come in an interposer so again AMD would be selling the HBM together with the GPU in the interposer.


HBM isn't tailored for handhelds either, as its JEDEC ultra-low power counterpart is Wide I/O.

The only possible candidate for HBM is company trying to build an ARM-powered HPC/server chip using that memory, but ARM chips haven't been awfully popular in that area.
Perhaps if/when apple decides to ditch x86 and replace them Intel cores with their ARM windy stuff, they'll be looking to use the best-of-the-best in for their high-margins laptops and AiOs. But that's still a long time away IMO.


So the short answer to your question is: no, there doesn't seem to be anyone else other than AMD wanting to buy HBM from SK Hynix at the moment.
Click to expand...

Samsung announced a network processor with HBM2 recently, and that type of high-speed signalling chip is a target for HBM2.
As far as Intel goes--or technically its Altera acquistion, the Stratix 10 was announced with HBM2 and at the time it was supposedly from Hynix.
Intel's Lake Crest uses HBM2, but that is also from another acquisition.
HPC and networking were targets for HBM2, and at least some work has gone into that.

The original HBM appears to be even more limited to AMD's last-gen hardware.
 
Kaotik said:
More likely is that they'll announce when it will be unveiled ;)
Click to expand...

They've just said on stream they will be showing demos, hopefully it's not DOOM again (don't have anything against it but we've seen that already :p).
 
You must log in or register to reply here.
Back
Top