updatevideocardzAMD Polaris 10 engineering sample ‘pictured’

[msi2]

I would like to ask, does anyone of you know if Micron, Hynix, Samsung or anyone developed 8-Hi stacks of HBM1?

As you can see: Tech Shrink.

That picture is a fake.

 

[ShintaiDK]

I would like to ask, does anyone of you know if Micron, Hynix, Samsung or anyone developed 8-Hi stacks of HBM1?

Only Hynix makes HBM1 and they gave up on 8Hi stacks. HBM2 replaces HBM1 and HBM1 goes out. They are not to coexist as such.

Nobody have shown 8Hi stacks yet. 8Hi also gives some challenges because every DRAM part will only have half the bandwidth as with 4Hi.

 

[Silverforce11]

For those interested in whether Polaris is HBM or not, listen closely to what is being said:

https://youtu.be/hvD37UUcdIo?t=3m15s

Everytime HBM is talked about by AMD, the issue of cost always creeps in. Note that they specifically say it will only use HBM for high-end segments due to cost sensitivities.

 

[ShintaiDK]

Exactly. Also why Polaris naturally enough will be GDDR5(x).

HBM will still in 2016 and 2017 be for flagship parts and enterprise.

 

[Glo.]

I think that is because HBM2 is able to work in Legacy mode with HBM1 memory controller, and offer the option for 8 GB or more at 4096 bit memory bus.

 

[JDG1980]

Only Hynix makes HBM1 and they gave up on 8Hi stacks.

Do you have a source for this?

 

[JDG1980]

Of course specific uarch paths are needed with DX12. Same reason why the fabled AOTS wont run on an Intel IGP. Because there is none.

So I assume you finally admitted that DX12 requires specific optimization for every single uarch, even as low as SKUs. And its a perfect place to do please your sponsor via.

Just wait till you see new uarchs and no patches, that will be a fun thing for older games. GCN 1.1 owners can rejoice over any console ports. But GCN 1.0, 1.2, 1.3, Kepler, Maxwell, Pascal, Gen8, Gen9, Gen10 owners can only pray to the dev gods for every game and see if they get lucky.

I still haven't seen any evidence that GCN 1.2 requires different optimizations from GCN 1.1. Both Tonga and Fiji are unbalanced designs (front ends too narrow for the shader counts), so the fact that they perform relatively worse than Hawaii doesn't prove much. It is true that GCN 1.0 may perform worse in some cases (especially involving async compute) since it has fewer hardware resources available. The slides at GDC talk about "AMD" and "Nvidia" optimizations; for Nvidia, they presumably mean Maxwell, but AMD is still selling all three iterations of GCN, and nothing in the slides indicates that these should be treated differently.

Pascal is likely going to be a modestly updated and die-shrunk Maxwell, so it's likely that one set of Nvidia optimizations will cover both architectures. Kepler will be left out in the cold, but that's already happening with DX11 games anyway.

 

[crisium]

Tonga and Fiji had visual artifacts and worse performance than Hawaii and Tahiti in Gears day 1. This can invariably only mean there is a difference in optimizations.

 

[Silverforce11]

Tonga and Fiji had visual artifacts and worse performance than Hawaii and Tahiti in Gears day 1. This can invariably only mean there is a difference in optimizations.

Yes, and they fixed it in a week or so, the 380 spanks the 960 after the optimizations, likewise for the other SKUs.

This really shows uarch specific optimization can be done quite quick in DX12. Good news for gamers.

 

[ShintaiDK]

Do you have a source for this?

That only Hynix makes HBM1 or they gave up on 8Hi HBM1? Last part is gone from all their newer material from 2015.
https://www.skhynix.com/eng/product/dramHBM.jsp

Example:

This is also what burned Fiji, AMD had expected 8GB cards due to 8Hi stacks. Instead they now have to use a hotfix solution with 2 engineers.

 

[Glo.]

As I said, they abandoned 8-Hi stacks, because HBM2 is able to work with HBM1 memory controller in legacy mode. Thats why it will be possible to make 8 GB HBM1 GPUs using HBM2 stacks.

http://www.overclock.net/t/1595065/...performance-entries-spotted/210#post_25003734
The same person says that Polaris 10 is 10-15% faster than Fury X.

Time will tell.

Or not. Looks like same person posts here and there under different nicks. I suggest observation of situation.

 

[ShintaiDK]

Where does this legacy information come from? Got any links?

Also what would the point of it be?

 

[Glo.]

http://monitorinsider.com/HBM.html Here you have explanation.

The point of observation is that people can post things to hype unreleased hardware. I've learned my lesson with Fury X, and mobile Maxwell GPUs on notebookreview forum. Same person hyped mobile Maxwell GPUs there, and hyped AMD GPUs here. Hyping mobile GPUs was obvious, but hype for AMD GPUs? The only reason was to make people feel disappointed with it so it could make more track for Nvidia GPUs. Same person have had access to information that was not obtainable easily, like getting info on upcoming refresh of Maxwell GPUs, and info on desktop GTX980 coming to laptops in MXM form. After the releases he/she shut his mouth here and not spoken once, till recently. I am wondering why...

I think you know what Im saying here. Also, Im not exactly pointing at anyone.

 

[Aristotelian]

Also what would the point of it be?

That's a good question, even moreso when taken generally. I've been reading article after article, from clickbait to reliable, and forum to forum, to try to figure out a bit more about what is coming.

All I'm seeing points to the mid range next generation coming first, using GDDR5 for cost concerns (including new process yields) while at the same time aiming for smaller chips on the new process and touting the power efficiency and equivalent performance roughly of the previous generation. Meaning, the new mid range will roughly equal the current high end.

But we will have to wait for 2017 to see the real deal next generation chips (larger ones, with HBM2), e.g., the Ti version, the Titan version, the Fury X full replacement chip (the Vega, I think).

I'm getting a bit of burnout. At this rate. those who want to build a new pc around the new graphics generation are better off waiting for skylake-e it seems if broadwell-e is coming out this may-june period - why pair an extreme build with mid range cards?

Just my thoughts here.

 

[ShintaiDK]

Maybe its just me, but I haven't seen this legacy before. And I wish for something more than a newly created unknown site.

 

[Glo.]

http://www.anandtech.com/show/9969/jedec-publishes-hbm2-specification
Is it enough for you?

If, for some reason, an ASIC developer believes that Pseudo Channel mode is not optimal for their product, then HBM2 chips can also work in Legacy mode.

Also worth noting, here: http://www.stonearch.net/forums/showpost.php?p=255569&postcount=438

 

[ShintaiDK]

Thanks

 

[arandomguy]

Yes, and they fixed it in a week or so, the 380 spanks the 960 after the optimizations, likewise for the other SKUs.

This really shows uarch specific optimization can be done quite quick in DX12. Good news for gamers.

We have a rather optimal case with GPUs that are currently selling and a developer with a vested interest in showing DX12 on a game that just launched. Yet the issue still occurred and was tested and supported prior to launch?

What happens in the 2018 time frame when we will have another set of GPU architectures added into the mix combined with games possibly at 1yr+ post release? What is the real support situation going to be?

I'm surprised this potential issue with low level APIs is not getting more attention.

 

[poofyhairguy]

Instead they now have to use a hotfix solution with 2 engineers.

Source on the two engineers? You always say two so I assume that is from somewhere. Thank you in advance.

 

[96Firebird]

Source on the two engineers? You always say two so I assume that is from somewhere. Thank you in advance.

Which is why for Fiji, AMD tells us they have dedicated two engineers to the task of VRAM optimizations. To be clear here, there’s little AMD can to do reduce VRAM consumption, but what they can do is better manage what resources are placed in VRAM and what resources are paged out to system RAM. Even this optimization can’t completely resolve the 4GB issue, but it can help up to a point. So long as game isn’t actively trying to use all 4GB of resources at once, then intelligent paging can help ensure that only the resources that are actively in use reside in VRAM and therefore are immediately available to the GPU when requested.

Source

 

[poofyhairguy]

Source

Thank you.

 

[IEC]

Well I'd be interested in a Polaris 11/1070 gtx @ ~ 110 watts, 2x the performance of my overclocked gtx960 (overclocked gtx980 performance) @ ~ 300$.

I think Polaris 10 /1080gtx might be out of the price range I'm comfortable paying for a card.

Think its possible?

Otherwise I might as well grab a G1 gamer gtx970 overclock it (182watts) to 980 speeds for ~240$ and eat the extra 70 watts.

A 182 watt card is at toaster oven levels, not quite a space heater.:thumbsup: j/k

Heh, you should feel the heat from a tri-fire/tri-SLI setup. Nothing like 1000W+ usage to heat a room in winter

If Polaris 10 exceeds 390X performance while coming in at less than half the TDP, all for <$400 I'd buy at least two (for VR), or possibly 3 for Crossfire in a PLP Eyefinity setup (assuming it is supported, I'd likely only buy 2 initially before playing beta tester).

 

[Silverforce11]

We have a rather optimal case with GPUs that are currently selling and a developer with a vested interest in showing DX12 on a game that just launched. Yet the issue still occurred and was tested and supported prior to launch?

What happens in the 2018 time frame when we will have another set of GPU architectures added into the mix combined with games possibly at 1yr+ post release? What is the real support situation going to be?

I'm surprised this potential issue with low level APIs is not getting more attention.

It's already been mentioned various times. DX12 and Vulkan are not for every studio/game, only for those who are capable and seeking peak performance.

DX11 will be for the rest.

The issue with GOW is they had an NV sponsorship, part of the GameWorks program. That may explain their lack of testing on AMD hardware, as well as their lack of communication with AMD about their release intention.

What we've seen with Ashes is that it performs very well on all hardware, Async Off, NV GPUs get a performance boost with DX12 too. It's a win-win scenario.

 

[Ma_Deuce]

We have a rather optimal case with GPUs that are currently selling and a developer with a vested interest in showing DX12 on a game that just launched. Yet the issue still occurred and was tested and supported prior to launch?

What happens in the 2018 time frame when we will have another set of GPU architectures added into the mix combined with games possibly at 1yr+ post release? What is the real support situation going to be?

I'm surprised this potential issue with low level APIs is not getting more attention.

You can go the complete opposite direction of DX12 and look at titles released using gameworks. This should be a best case scenario in which the developers have to do minimal work. They have a tendency to run terribly though, across the board, especially at launch.

I don't see how the FUD of DX12 is really justified.

 

[poofyhairguy]

I am really excited about Polaris 10 if it really is a 390X that uses less power.

I might get one when they are released at the jacked up early price and then get another in a year or so to crossfire in preparation of the Oculus gen 2.