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

Thought on the memory speeds, what if the fabric was running synchronously like Ryzen? 1600-2000MHz memory lines up with base/boost clocks.

This tech could be really nice on average to low end gpus, but this dont change the compute power, it only reduce the impact on memory.
Or for VMs ND separate apps where each could use up to the full memory capacity. Saves the driver from having to know what is active in an oversubscribed pool.

Remember, this is the same company that has the peace of mind to tape up their power connectors before opening up the case to Linus for his CES video.
They were likely covering the tachometer showing it running 30% of capacity. The initial benches could have only been using ~100W.

There are also the powered display connectors hitting in Vega's lifetime. I wouldn't be surprised to see a third connector show up for prosumer parts driving multiple displays approaching 75W each. Admittedly I haven't seen how any vendor is handling the passthrough.

I'm no EE, but I bet that there's additional power equipment necessary to deal with the current that those beefier/extra connectors can try to draw. Therefore, the cost to allow 2 8-pin connectors on a 150W card is more than just the physical connectors.
Many PSUs have a single power plane for 12V so the connectors don't matter much. The only real difference will be any diodes to prevent feedback and how they limit when overloaded. The system designer would pick a large enough supply to accommodate whatever was going in the system. Professional systems are often fixed and consumers won't adhere to specs in aftermarket anyways.
 
Well they are comparing Vega FE (presumably the watercooled one) to Tesla P100 in a presumably custom workload we don't know much about. Tesla P100 is rated for 9.3tflops (the pcie version, which they undoubtedly used) and is a 300W double precision card that is passively cooled.

I would not count it passively cooled when it relies on certain type of chassis and having certain setup on fans..
 
I would not count it passively cooled when it relies on certain type of chassis and having certain setup on fans..
Most of us are going by the slide showed back in December for the accelerators, Mi6 and Mi25 passive in a way similar to Tesla cards I assume as shown by Ieldra.

AMD-Radeon-Instinct_Final-for-Distribution-page-017-2_575px.jpg


Probably fair to say the Frontier cards are not designed for large scale project implementation specifically with nodes/clusters, which is what it seems to me the Mi25 is for in those setups as a passive.

Cheers
 
Last edited:
Most of us are going by the slide showed back in December for the accelerators, Mi6 and Mi25 passive in a way similar to Tesla cards I assume as shown by Ieldra.

AMD-Radeon-Instinct_Final-for-Distribution-page-017-2_575px.jpg


Probably fair to say the Frontier cards are not designed for large scale project implementation specifically with nodes/clusters, which is what it seems to me the Mi25 is for in those setups.

Cheers

Yeah. I would not count those passive either. Frontier cards are most likely very very limited availability and thus nothing to sell on volumes for oem. I believe that there has really been some issues with hbm2 memory availability, hopefully those are behind now and amd can get volume up for second half.
 
Yeah. I would not count those passive either. Frontier cards are most likely very very limited availability and thus nothing to sell on volumes for oem. I believe that there has really been some issues with hbm2 memory availability, hopefully those are behind now and amd can get volume up for second half.
I think Ieldra like myself are saying only the MI25 will be passive chassis based while the other Vega are air or water cooled.
Mentioning just in case we have our wires crossed on this, and yeah if AMD has some kind supply issues it may be a little while until AMD can build a supply of these *shrugs*.
Cheers
 
Fair point - but where was it mentioned before? The 480 GB/s I mean? Must have missed that.
In the posts above mine..

OT: Your constant bickering about how you were possibly right and other were wrong is mildly annoying. Remember, until yesterday, we were in a rumors/speculation thread.

Oh please do get a grip... It was one sentence inside one post in which I put a whole bunch of other information.

The Internet Rules clearly state I'm allowed of writing that one sentence. It's right there in article 14 b).
;)


Has there ever been a case where a professional card was clocked higher than its consumer equivalent?
Point well taken.
So we should expect the gaming cards to be clocked a bit higher, and perhaps that's where the 1600MHz Compubench result came from.


Why is the Vega card so much larger physically than the Fiji?
Maybe not the card itself but only the heatsink that is longer. It's 175W vs. 300W.
Had AMD chosen to use the Fury version of Fiji instead of the Nano one, the lengths would be similar.
 
Last edited by a moderator:
The memory could be tied to the fabric speed and limited by the cores. Or they simply didn't need that much bandwidth and dialed it down a bit. Allowing for a refresh in 6 months or so with full speeds and a corresponding clock boost. One thing we haven't seen with these cards is actual power usage. A water cooled Vega could be an overclocker's dream. And I'll show myself out now.
 
Link to full text of Raja's Q&A.
We’ll be showing Radeon RX Vega off at Computex, but it won't be on store shelves that week. We know how eager you are to get your hands on Radeon RX Vega, and we’re working extremely hard to bring you a graphics card that you’ll be incredibly proud to own. Developing products with billions of transistors and forward-thinking architecture is extremely difficult -- but extremely rewarding -- work. And some of Vega’s features, like our High Bandwidth Cache Controller, HBM2, Rapid-Packed Math, or the new geometry pipeline, have the potential to really break new ground and fundamentally improve game development. These aren’t things that can be mastered overnight. It takes time for developers to adapt and adopt new techniques that make your gaming experience better than ever. We believe those experiences are worth waiting for and shouldn’t be rushed out the door. We’re working as hard as we can to bring you Radeon RX Vega.

On HBM2, we’re effectively putting a technology that’s been limited to super expensive, out-of-reach GPUs into a consumer product. Right now only insanely priced graphics cards from our competitors that aren’t within reach of any gamer or consumer make use of it. We want to bring all of that goodness to you. And that’s not easy! It’s not like you can run down to the corner store to get HBM2. The good news is that unlike HBM1, HBM2 is offered from multiple memory vendors – including Samsung and Hynix – and production is ramping to meet the level of demand that we believe Radeon Vega products will see in the market.
https://www.reddit.com/r/Amd/comments/6bklro/we_are_radeon_technologies_group_at_amd_and_were/
 
View attachment 1982
So it's confirmed 2 stacks.

I wonder how AMD got access to things that can't possibly exist?

Yeah cynic me is wrong but something is up at SK Hynix because they then should had launched separately back 7 months ago the 1.6Gbps 4GB, it would not need to be held back with AMD using 8-Hi or higher density 8GB component.
With this situation SK Hynix are missing out on critical areas such as integrating this for production now into network processors as Samsung recently has or clients who want HBM2 such as Xilinx and those with solutions comparable to Stratix 10.

Anyway AMD I feel missed an opportunity to get one up on Nvidia, it was expected Volta would be more than 16GB HBM2 awhile back.
With a 3 or 4-stack Frontier would had been better designed specifically for HPC/data scientists/elite visualisation as it would have a more ideal BW (one reason Nvidia has increased theirs in this segment and 480GB/s is tight) and importantly 24GB+ that scientists and some others could really use, which would also had been a coup and great marketing (while also having uses) over V100.
Considering the market they are targeting the Frontier for, it could be comfortably priced in and still work out much cheaper than what Nvidia is charging.
 
Last edited:
With a 3 or 4-stack Frontier would had been better designed specifically for HPC/data scientists/elite visualisation as it would have a more ideal BW (one reason Nvidia has increased theirs in this segment and 480Gb/s is tight) and importantly 24GB+, whicht would had been a coup and great marketing (while also having uses) over V100.
It's also possible there is a dual GPU part providing that. Lower clockspeeds for ideal perf/watt might have made more sense. V100 double exposed an interposer to get up around an 800mm2 die. A dual Vega would be North of 1000mm2 with a similar 4 stacks. Raja did say the dual ASIC was possible with Infinity, but not something they have "mentioned". For a high end competitor that probably makes more sense.

As for the 8-Hi stacks, what if that was part of some HBM exclusivity clause. The assumption was always HBM in general, but it could have been more specific. Allow competitors to use it, but reserve the big stuff for a competitive advantage.
 
It's also possible there is a dual GPU part providing that. Lower clockspeeds for ideal perf/watt might have made more sense. V100 double exposed an interposer to get up around an 800mm2 die. A dual Vega would be North of 1000mm2 with a similar 4 stacks. Raja did say the dual ASIC was possible with Infinity, but not something they have "mentioned". For a high end competitor that probably makes more sense.

As for the 8-Hi stacks, what if that was part of some HBM exclusivity clause. The assumption was always HBM in general, but it could have been more specific. Allow competitors to use it, but reserve the big stuff for a competitive advantage.
It would not need to be a dual GPU though, Frontier performance is perfect for it and they would had a window advantage over Nvidia HBM2 Volta memory limitation.
Makes sense what you say about probably a dual GPU will be available down the line but when; I would say MI25 is more important than that as a chassis node-cluster solution and larger contracts-projects.
Regarding exclusivity; I guess it could had been a special order contract (like the version Nvidia has I think for Volta and 12n), but I cannot imagine how much AMD would need to pay SK Hynix for them to hold back also 1.6Gbps 4GB as they will be losing out on a lot of potential and playing catchup in lucrative areas such as network processors/backplanes/other clients/etc that Samsung just recently seems to be able to take a step into.
Cheers
 
Last edited:
Back
Top