Question 'Ampere'/Next-gen gaming uarch speculation thread

[Ottonomous]

How much is the Samsung 7nm EUV process expected to provide in terms of gains?
How will the RTX components be scaled/developed?
Any major architectural enhancements expected?
Will VRAM be bumped to 16/12/12 for the top three?
Will there be further fragmentation in the lineup? (Keeping turing at cheaper prices, while offering 'beefed up RTX' options at the top?)
Will the top card be capable of >4K60, at least 90?
Would Nvidia ever consider an HBM implementation in the gaming lineup?
Will Nvidia introduce new proprietary technologies again?

Sorry if imprudent/uncalled for, just interested in the forum member's thoughts.

 

[Mopetar]

Were those speeds for a particular product? If it's the base clock for the top end part with a massive number of shaders then 1700 MHz is really good.

The last Titan only had a base clock of 1350 MHz but could boost to over 1700 MHz. If we have a similar situation here then a 2100 MHz boost seems reasonable.

 

[moonbogg]

Who thinks this Samsung 8nm GPU is going to heat up faster than a nuclear-powered toaster?

 

[jpiniero]

Who thinks this Samsung 8nm GPU is going to heat up faster than a nuclear-powered toaster?

That's why nVidia made the new cooler.

 

[moonbogg]

That's why nVidia made the new cooler.

It could actually be a good thing. A waterblock could potentially offer a huge performance advantage. Everyone else is like "ugh, 90c at 80% load" and I'd be like "yawn, 50c at 120% load"

 

[jpiniero]

It could actually be a good thing. A waterblock could potentially offer a huge performance advantage. Everyone else is like "ugh, 90c at 80% load" and I'd be like "yawn, 50c at 120% load"

I think it's just that the FE is going to be like an overclocked model, like the Lightning Z. There may not be much room beyond what the stock FE offers.

 

[Stuka87]

It could actually be a good thing. A waterblock could potentially offer a huge performance advantage. Everyone else is like "ugh, 90c at 80% load" and I'd be like "yawn, 50c at 120% load"

They could go the AMD Fury route and offer a factory AIO cooler.

But a 20% OC would mean more like 40% more power dispersed. That would turn a 250W GPU into a 350W GPU. Or a 300W GPU into a 420W GPU. Put that 600W connector to work

 

[BenSkywalker]

So elementary English it is-

"AMD never publicly stated that"

Means something entirely different than

"You made that up"

In no way are they remotely close. Considering you all have so much faith in the Tweets of some kid who keeps contradicting himself, it is odd that you take such exception to what the mainstream press has been using as a term for the last eight months.

It also speaks volumes about how much confidence you actually have in AMD.

 

[Campy]

If the rumours about the price of nvidia's new cooler are true, they really should just slap an AIO on it instead.

 

[moonbogg]

If the rumours about the price of nvidia's new cooler are true, they really should just slap an AIO on it instead.

Don't worry man, it's going to be so expensive neither of us will care one way or another about whatever cooler they go with. I can't wait for a good laugh while I contemplate buying a GPU so expensive I'd have to trade in the family mini-van and take a second out on the house + an extra $50/mo to Edison.

 

[FaaR]

10? WoW, Minecraft and Cyberpunk are all pretty high profile too.

WoW uses it for shadowmap generation only. As the game already has a solid shadow rendering implementation it's basically just a checkbox feature for funsies, nothing more.

Also, again, when you respond to someone's post directly as you did in this case, use Proper Fricken Quoting. Everybody else can do it! So can you. I don't care about your stupid excuses why you think you shouldn't have to. None of them make the least bit of sense. Just do it. Stop obstructing!

 

[BenSkywalker]

You explicitly stated AMD made the "nVidia killer" claim here

Is English your third or fourth language? Explicitly stated and implied are two very different things.

Furthermore, I'm not the one who implied that in the first place, it has been implied hundreds of times by mainstream news outlets everywhere.

Meanwhile, you all are treating Twitter kiddie as a messiah even when he comes out and says the rumors are lies.... Rumors he created in the first place. He has now contradicted every single one of his own claims- yet people still cling to it, while the term widely used by most of the major tech sites for the entire year... that's what you take issue with?

Do you people have *any* standards at all?

Personal insults/attacks are not allowed (bolded by me above).

AT Mod Usandthem

 

[DooKey]

Don't worry man, it's going to be so expensive neither of us will care one way or another about whatever cooler they go with. I can't wait for a good laugh while I contemplate buying a GPU so expensive I'd have to trade in the family mini-van and take a second out on the house + an extra $50/mo to Edison.

Learn to code. /s

 

[DXDiag]

Ampere to have 2X FP32 units per SM compared to Turing, this accelerates RT performance greatly in combination with RT cores improvements.

https://twitter.com/x/status/1294120100400832513

https://twitter.com/x/status/1293528946571542529

 

[BFG10K]

Alleged 3090 PCB pictures showing 3x8-pin: https://www.techpowerup.com/270986/nvidia-geforce-rtx-3090-ampere-alleged-pcb-picture-surfaces

That configuration allows a theoretical maximum of 525 watts.

 

[Glo.]

Micron confirms NVIDIA GeForce RTX 3090 gets 21Gbps GDDR6X memory - VideoCardz.com

Micron has confirmed that it has been developing GDDR6X memory in collaboration with NVIDIA for its next-generation Ampere graphics cards. NVIDIA GeForce RTX 3090 gets 19-21Gbps GDDR6X memory In a technology brief, Micron has confirmed that GeForce RTX 3090 will sport memory based GDDR6X...

videocardz.com

Edit:

https://twitter.com/x/status/1294239371160358913

Here goes for the naysayers.

 

[BFG10K]

12GB in a high-end part is piss-poor given we've had 11GB in consumer space for over 3 years. Especially considering this thing will almost certainly be horrifically overpriced.

 

[Gideon]

GDDR6X does exist didn't expect that

 

[exquisitechar]

GDDR6X does exist didn't expect that

Me neither, but kopite7kimi delivers once again.

Now I want to see what's up with the double FP32 per SM rumor.

Edit: Speaking of that:

https://twitter.com/x/status/1294241335935819776

 

[Karnak]

Well, since even GDDR6X is now official I think all Igor said so far was true. Will probably be the same for his rumoured up to 350W TDP. Guessing that's due to Samsung's bad 8nm LPP in terms of higher clock speeds.

 

[Glo.]

High power draw is all due to Nvidia being required to clock those GPUs to hell.

 

[Bouowmx]

Alleged 3090 PCB pictures showing 3x8-pin: https://www.techpowerup.com/270986/nvidia-geforce-rtx-3090-ampere-alleged-pcb-picture-surfaces

That configuration allows a theoretical maximum of 525 watts.

That's the back side of the PCB. Looks like 22 GB on that example, double-sided memory coming to GeForce, and GDDR6X 19-21 GT/s only coming in 1 GB packages.

 

[Stuka87]

That's the back side of the PCB. Looks like 22 GB on that example, double-sided memory coming to GeForce, and GDDR6X 19-21 GT/s only coming in 1 GB packages.

They have front and back sides of the PCB. There is very clearly no places for memory on the backside of the board. There is 11 modules on the front.

 

[Bouowmx]

They have front and back sides of the PCB. There is very clearly no places for memory on the backside of the board. There is 11 modules on the front.

Because front-side of graphics card has the visible screws to the heatsink that is still attached, no visible VRM components, and just the 8-pin connector solder points, not the connector itself. It's the back side.

 

[Bouowmx]

Ampere to have 2X FP32 units per SM compared to Turing, this accelerates RT performance greatly in combination with RT cores improvements.

A weird rumor flying under the radar.
I thought doubled FP32 would mean:
7 GPC * 6 TPC * 2 SM * 128 cores = 10752 cores

kopite7kimi thinks it's a bit more nuanced:

https://twitter.com/x/status/1293857308552736769

SM would have 128 FP32 + 64 INT32 ALUs, compared to Turing 64 FP32 + 64 INT32.

 

[Stuka87]

View attachment 27970

Because front-side of graphics card has the visible screws to the heatsink that is still attached, no visible VRM components, and just the 8-pin connector solder points, not the connector itself. It's the back side.

I missed the heatsink behind the card on that image. I took it for a mirrored front side since it has a GPU sitting there. And the pads for the vregs and power connectors and such are there.

This image very clearly has no memory on the backside, so maybe we are seeing two different PCBs?

EDIT: Looking again, I am quite sure those first two images are of the front side, the heatsink is there behind it, but not attached. The inner screws attach to the backplate.