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.

 

[Kenmitch]

When you have iPhone margins, you don't need to switch. Apple ships millions of units every year on TSMC's bleeding edge node and doesn't even blink.

So in a round about way your saying Apple drives the success of the foundries?

 

[shiznit]

So in a round about way your saying Apple drives the success of the foundries?

It drives TSMC's success for sure. Charging ahead with a bleeding edge node is a lot easier when you know that Apple will pay top dollar for every wafer.

 

[xpea]

Is this Samsung foundry's first major contract for a complex chip like this? If so they must've bend over and backward to please the first big customer like nvidia. Plus, nvidia has a history of being extremely aggressive in negotiations as well.

IBM Power10 and Z monster datacenter CPUs will be made by Samsung in 7nm EUV next year

IBM Expands Strategic Partnership with Samsung to Include 7nm Chip Manufacturing

IBM (NYSE: IBM) today announced an agreement with Samsung to manufacture 7-nanometer (nm) microprocessors for IBM Power Systems™, IBM Z™ and LinuxONE™, high-performance computing (HPC)...

newsroom.ibm.com

 

[traderjay]

So in a round about way your saying Apple drives the success of the foundries?

Its a symbiotic relationship, they are virtually inseparable and they feed on each other's success. Apple's huge volume (revenue for foundry) drives innovation and more R&D. Just look at Huawei's Kirin w/o fab, they are on life support now and won't last too much longer.

IBM Power10 and Z monster datacenter CPUs will be made by Samsung in 7nm EUV next year

IBM Expands Strategic Partnership with Samsung to Include 7nm Chip Manufacturing

IBM (NYSE: IBM) today announced an agreement with Samsung to manufacture 7-nanometer (nm) microprocessors for IBM Power Systems™, IBM Z™ and LinuxONE™, high-performance computing (HPC)...

newsroom.ibm.com

Interesting but I guess the IBM Power CPUs doesn't move anywhere near the same volume as nvidia, hence I think nvidia is the make or break moment for Samsung's foundry.

 

[maddie]

It will most likely help Samsung. When other vendors of large ASICs see the volume and margins Nvidia will be pushing with Ampere, they'll come knocking. Nobody wants to fight over TSMC's gaps if they can help it.

Also, Samsung doesn't need Apple's SoC business to fund innovation. Their enormous NAND and RAM business does that already.

Well they obviously not quite doing everything right. One only has to compare the 3080 & 3090 parameters to realize that something is very rotten with SS8.

Edit:
Just a thought experiment for argument sake and pushing the corner here. Let's assume that the fab cost of a die is free.

Do we have any idea of what the product using it will have to sell at to breakeven?

Assume now that the performance characteristics are worse than the competitor. Power, performance, etc.

What is the new selling price to compete in the marketplace?

In other words, at what price point is a free die equal or worse than one with superior parameters.

 

[Kenmitch]

In other words, at what price point is a free die equal or worse than one with superior parameters.

Too many variable to consider. You'd have to give example's of who's die's your talking about. Once that's done you know what the outcome will be already.

 

[shiznit]

Well they obviously not quite doing everything right. One only has to compare the 3080 & 3090 parameters to realize that something is very rotten with SS8.

Edit:
Just a thought experiment for argument sake and pushing the corner here. Let's assume that the fab cost of a die is free.

Do we have any idea of what the product using it will have to sell at to breakeven?

Assume now that the performance characteristics are worse than the competitor. Power, performance, etc.

What is the new selling price to compete in the marketplace?

In other words, at what price point is a free die equal or worse than one with superior parameters.

Many variables at play here so I'll oversimplify for the sake of argument. I think it's safe to say that TSMC 7nm is superior to Samsung 8nm, nobody is disputing that. It just boils down to capacity and margins. Apple and AMD are paying top dollar for 7nm wafers, their tiny 100mm2 and smaller chips probably get great yields. So if you're Nvidia, if you can't get the capacity you need from TSMC, you're stuck with Samsung. If you can't get the yields for good margins from TSMC (remember these are 600mm2 or greater chips), you're stuck with Samsung. So you choose high power consumption with Samsung and market the hell out of your performance per dollar, as we see them doing. If it succeeds, other players with large, low-medium margin ASICs not sensitive to power consumption should follow.

GA102 and GA104 are in no man's land for TSMC 7nm. At capacity-constrained prices, the die size is way too large for great yields and the margins are too small for poor yields.

*Edited for clarity

 

[Mopetar]

It drives TSMC's success for sure. Charging ahead with a bleeding edge node is a lot easier when you know that Apple will pay top dollar for every wafer.

Apple only has to pay what someone else would be willing to. TSMC can ask for a price, but it has to be one Apple is willing to pay. If Apple won't buy at that price TSMC would have to sell to someone who's only willing to spend less than Apple. Unsold wafers are a huge money pit.

Smaller chips always work better on a new process because even with a high defect density you still get a lot of working silicon. If you're making massive chips like NVidia is you need a more mature process or the ability to sell each of the far fewer functional chips for a lot of money.

TSMC is more successful because they have a wide variety of customers with different needs that allows them to support new processes in a financially feasible way. Intel has problems because a lot of the chips they want to sell aren't viable on a bleeding edge process node and they don't have any volume parts that are a good fit for a new node going through refinement that makes larger consumer chips feasible.

It's easy to point to 10nm being an unmitigated disaster, and there's truth to that, but the 14nm process had a rocky start as well. TSMC would have more problems if they didn't have companies manufacturing SoCs on their new nodes where it's still viable even if the defect density is exceptionally high. Companies like NVidia wouldn't be able to make large dues for consumer products, at least not without heavily impacting their own margins

 

[maddie]

Too many variable to consider. You'd have to give example's of who's die's your talking about. Once that's done you know what the outcome will be already.

Agreed on the many variables, but the takeaway is that for any given set of variables, there is a crossover point where a free die isn't worth it. A unique solution has to be derived for each set of values.

My argument is that most assume that cheaper or free is always better. I think the reality is, not always, it depends.

 

[Kenmitch]

My argument is that most assume that cheaper or free is always better. I think the reality is, not always, it depends.

Maybe I get it?....Kind of like turkey bacon vs real bacon? Might taste the same at 1st, but in the end it leaves a bad taste in your mouth.

 

[KompuKare]

If you can't get the yields for good margins from TSMC (remember these are 600mm2 or greater chips), you're stuck with Samsung.

Well, except that GA102 wouldn't be 600mm²+ if they were on TSMC's 7nm.
The GA100 on TSCM's 7nm gets around 65.6M / mm² according to TPU

NVIDIA GA100 GPU Specs

8192 Cores, 512 TMUs, 192 ROPs

www.techpowerup.com

While GA102 on Samsung's 8nm gets around 44.7M / mm², again according to TPU

NVIDIA GA102 GPU Specs

10752 Cores, 336 TMUs, 112 ROPs

www.techpowerup.com

So if GA102 was on TSMC, instead of being 627mm² it might be just 430mm² or so.

 

[Saylick]

Well, except that GA102 wouldn't be 600mm²+ if they were on TSMC's 7nm.
The GA100 on TSCM's 7nm gets around 65.6M / mm² according to TPU

NVIDIA GA100 GPU Specs

8192 Cores, 512 TMUs, 192 ROPs

www.techpowerup.com

While GA102 on Samsung's 8nm gets around 44.7M / mm², again according to TPU

NVIDIA GA102 GPU Specs

10752 Cores, 336 TMUs, 112 ROPs

www.techpowerup.com

So if GA102 was on TSMC, instead of being 627mm² it might be just 430mm² or so.

Great info and post.
I'd like to add the caveat that A100 has 40 MB of L2 cache, which likely inflated the transistor density. Navi 10 is on TSMC N7P and it only gets around 41 MT/mm2 (10.3B xtors, 251mm2) so for a typical gaming GPU, TSMC N7 and SS 8N are comparable. We'll see how much L2 cache GA102 had but I wouldn't be surprised if it's far less than the 40MB of A100.

 

[MrTeal]

Great info and post.
I'd like to add the caveat that A100 has 40 MB of L2 cache, which likely inflated the transistor density. Navi 10 is on TSMC N7P and it only gets around 41 MT/mm2 (10.3B xtors, 251mm2) so for a typical gaming GPU, TSMC N7 and SS 8N are comparable. We'll see how much L2 cache GA102 had but I wouldn't be surprised if it's far less than the 40MB of A100.

GA100 also doesn't need to drive offboard GDDR6X memory, with probably also helps a little with density.

 

[Asterox]

Hm, i expect Class Action Lawsuit RTX 2080ti owners vs Nvidia.

 

[Konan]

New arch details - Also Ryan has a write up coming out shortly on Anandtech to keep an eye out for.

NVIDIA details GeForce RTX 30 Ampere architecture - VideoCardz.com

A more detailed look at the Ampere GeForce architecture has been provided by NVIDIA. The company released block diagrams, technology descriptions, and detailed specifications of the next-generation GeForce graphics cards. NVIDIA details Ampere architecture NVIDIA has confirmed the expected...

videocardz.com

 

[Markfw]

Hm, i expect Class Action Lawsuit RTX 2080ti owners vs Nvidia.

Its in spanish, and the closed captioning in in spanish, so I have no idea what this is saying. Can you please give be a short synopsis of what its saying.

 

[Tup3x]

Its in spanish, and the closed captioning in in spanish, so I have no idea what this is saying. Can you please give be a short synopsis of what its saying.

It's hard subbed and no, the subs don't obviously have anything to do with what the guy is actually saying (it's a meme).

 

[MrTeal]

Read it on Techpowerup rather than the Videocardz tl;dr. The pictures are actually legible on TPU.

NVIDIA GeForce Ampere Architecture, Board Design, Gaming Tech & Software

NVIDIA's new Ampere architecture brings many interesting improvements, especially for raytracing and DLSS. In this special article we're going into all the technical details: how the shader counts are doubled and what makes GeForce 3000 RTX so much faster. We also take a closer look at the...

www.techpowerup.com

 

[n0x1ous]

112 ROPs on 3090
96 on 3080

3090 is not fully enabled ga102

 

[sze5003]

Its in spanish, and the closed captioning in in spanish, so I have no idea what this is saying. Can you please give be a short synopsis of what its saying.

That's odd for me the subs are in English. It's a very old Spanish show where he's is taking about some cooks. But people dub the video to make fun of recent events.

He's saying the 2080ti owners got screwed, they were made to pay 1200 for 2 extra Rays and beta test Nvidia products. And now they released the 3070 which is better than the 2080ti for $499 and now idiots are selling their 2080ti's on eBay for $500 lol.

 

[Det0x]

112 ROPs on 3090
96 on 3080

3090 is not fully enabled ga102

Very interesting
My wild guess for the future is:

Nvidia release RTX3070, RTX3080 and RTX3090.

AMD responds with Navi21 RX6800XT and RX6900XT
(RX 6900XT performs between RTX3080 and RTX3090)

Nvidia counters with 3080 TI/super which have the same 10496 CUDA Cores as RTX3090 but with a 320-bit memory bus and 20GB G6X gigabytes clocked @ 20-21 Gbps for 999$
They also finally release the Ampere RTX TITAN with the full GA102 die + 384-bit memory bus and 48GB G6X clocked @ 20-21 Gbps for 2500$
(G6X should be able to hit clockspeeds upwards to 21 Gbps according to Micron who produce the memory, and they are even talking about graphic cards with over 1TB/s bandwidth)

 

[Markfw]

That's odd for me the subs are in English. It's a very old Spanish show where he's is taking about some cooks. But people dub the video to make fun of recent events.

He's saying the 2080ti owners got screwed, they were made to pay 1200 for 2 extra Rays and beta test Nvidia products. And now they released the 3070 which is better than the 2080ti for $499 and now idiots are selling their 2080ti's on eBay for $500 lol.

I am one of these. I got 5 of them for an average of #1200. But Use them for F@H, and do do twice what my 1080TIs did, so I got my moneys worth.

Now 1 or 2 3090's are in my future.....

 

[jpiniero]

(G6X should be able to hit clockspeeds upwards to 21 Gbps Micron who produce the memory)

Almost wonder if the memory speed decision was based upon power consumption. Both GA102 FE's max power is 375, if the TPU article is correct.

I wonder if AIBs would be allowed to sell 21 gbps cards if they go 3x 8-pin.

 

[sze5003]

I am one of these. I got 5 of them for an average of #1200. But Use them for F@H, and do do twice what my 1080TIs did, so I got my moneys worth.

Now 1 or 2 3090's are in my future.....

Yea well then you used them to get money back so then you didn't completely lose out. As long as you use them for something worth your while, it's not a lost cause. I skipped 20 series all together after reviews were out and I'm glad I did because I didn't do much more than what I needed with my 1080ti. I'll probably bite on the 3090 now though too. Tired of doing the yearly upgrade.

 

[MrTeal]

112 ROPs on 3090
96 on 3080

3090 is not fully enabled ga102

It's not fully enabled, but it's pretty close. With only 2.4% of the shaders disabled and everything else on the chip being fully enabled there's really no room for a Titan on top of it, especially with it already having 24GB.
I expect we might see a fully enabled Quadro product at some point.