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]

RTX2070 Super never got a competition from AMD, RX5700XT was competing against the RTX2060 Super and they really did ok with 40% vs 60% of share in Steam Hardware Survey.

The 5700XT was always more of a 2070 SUPER competitor than it was a 2060 SUPER competitor even though it had the same MSRP as the 2060 SUPER. Hardware Unboxed did a recent test involving the three cards with the newest drivers and their new test system.

In the 14 games they benchmarked the 2070 SUPER and the 5700XT were basically identical for 1080p and 1440p. The 5700XT beat the 2060 SUPER by 20%. Even before the driver improvements, the 5700XT was always much closer to the 2070 SUPER in performance than it was to the 2060 SUPER.

People love to ignore steam when it's inconvenient to their narrative, but since most games are sold on Steam, a random selection of their users is going to be proportional to cards sold to gamers.

I suppose in theory it could under-represent some compute users, but that doesn't seem to be a significant concern. Not perfect, but It's the best data we have, and a reasonable representation.

Steam isn't a random sample. Users have to choose to participate.

If you wanted to have an real degree of strength to an argument that relies on a data source like that, you'd want several others to back it up. None might individually be perfect (or even great) but even poor sources in aggregate can lend credence to an argument when all of those sources point in a similar direction.

 

[Grooveriding]

I'm expecting we see a respin of some sort like we saw with Fermi and 480/70 to 580/70. Better thermals and power consumption, full chip etc. These things are using a ton of power, they have good cooling though.

The FE looks like a nice card for the 3080, but a total waste for the 3090. You want a beefy AIB if you're getting a 3090, even a 3080 really. Higher power limit, better power delivery and a binned chip.

I have my eye on the EVGA ftw 3080, or a Strix 3080.

 

[Heartbreaker]

Steam isn't a random sample. Users have to choose to participate.

What's the reasonable argument for significantly more AMD users, than NVidia users, choosing not to participate?

Plus it isn't like Steam Contradicts other sources:

Nvidia graphics cards now account for 80% of the PC market

Nvidia reportedly has 80% of the discrete PC GPU market, Jon Peddie Research claims.

www.pcgamer.com

 

[AtenRa]

The 5700XT was always more of a 2070 SUPER competitor than it was a 2060 SUPER competitor even though it had the same MSRP as the 2060 SUPER. Hardware Unboxed did a recent test involving the three cards with the newest drivers and their new test system.

In the 14 games they benchmarked the 2070 SUPER and the 5700XT were basically identical for 1080p and 1440p. The 5700XT beat the 2060 SUPER by 20%. Even before the driver improvements, the 5700XT was always much closer to the 2070 SUPER in performance than it was to the 2060 SUPER.

Simple no, from the start NAVI 10 was a direct competitor to the TU106 and RX5700XT was a direct competitor to the RTX2070 that was replaced with the RTX2060 Super.
AMD never released a RTX2070 Super (TU104) competing graphics card, just because RX5700XT can reach close to same Raster performance in the latest games, it doesnt mean that those two cards are at the same segment.

 

[AtenRa]

Why? You think that if they allow re


Sorry, thought it was you that downvoted my answer instead of addressing. Removed.

People love to ignore steam when it's inconvenient to their narrative, but since most games are sold on Steam, a random selection of their users is going to be proportional to cards sold to gamers.

I suppose in theory it could under-represent some compute users, but that doesn't seem to be a significant concern. Not perfect, but It's the best data we have, and a reasonable representation.

You can clearly understand that Steam Hardware Survey doesnt have much to do with Actual sales when you see the RTX2070 that is EOL over a year now to increase its share the last 4-5 months in the Steam Hardware Survey

NVIDIA GeForce RTX 2070
April 1.79%
May 1.81%
June 1.85%
July 1.89%
August 1.97%

For a dead card it sure makes lots of sales for NVIDIA

 

[Heartbreaker]

For a dead card it sure makes lots of sales for NVIDIA

It remained in production, there was a story about this soon after Super cards launched, and you could always go to newegg and find many in stock. Including today, where I see several in stock and shipping from newegg itself at a little over $400.

Similar price, performance and stock to the 2060 Super.

 

[AtenRa]

Similar price, performance and stock to the 2060 Super.

Well, im sure they had lots of them in stock not moving when they released the RTX2060 Super at $100 less.
Also to point out it has lower share per month in the Steam Hardware Survey for the last 4-5 months vs RX5700XT which is the direct competitor along the RTX2060 Super.

 

[Heartbreaker]

Well, im sure they had lots of them in stock not moving when they released the RTX2060 Super at $100 less.
Also to point out it has lower share per month in the Steam Hardware Survey for the last 4-5 months vs RX5700XT which is the direct competitor along the RTX2060 Super.

For monthly survey data, there is going to be some error rate, and if you understand math/statistics, you would recognize that looking at the much smaller monthly fluctuations, you are keeping the same amount of error, but placing it on a the smaller value of the fluctuation, multiplying the magnitude of the error on the fluctuation value.

Basically trying to read tea leaves of the fluctuations is completely useless.

 

[Kenmitch]

People love to ignore steam when it's inconvenient to their narrative, but since most games are sold on Steam, a random selection of their users is going to be proportional to cards sold to gamers.

I've been polled a couple times when using a Nvidia product, but ever remember being polled when using AMD's.

If you want to go buy the steam survey as the end all then nobody games at 4k nor uses the upper end gpu's.

 

[Bouowmx]

Can we please stop talking about old crap? Stick to Ampere and next-gen GeForce please!

 

[Kenmitch]

Can we please stop talking about old crap? Stick to Ampere and next-gen GeForce please!

Next-gen GeForce? Seeing how hard it is to get your hands on Ampere I'd guess it's way to early to discuss the next-gen.

 

[Midwayman]

Yup I'm curious now to see how the 3080 20gb or super/whatever they name them, will fit into the lineup. Why would they even release a 20gb 3080 unless that was the super card?

If it's just a card with more Vram, they could release it in late October following Navi announcement. If it is an actual Super type card, I don't see it releasing anytime soon. More like 6 months from now.

It wouldn't be the first time the only difference in cards was amount of vram. I feel like nvidia is pulling an apple here. The low end sku has just enough memory for people to worry day to day and then the 'real' usage version has a stupid upcharge for memory. Nvidia could have easily made the 3080 12gb, but they wouldn't be able to advertise a low base price and still capture a lot of profit on people upgrading memory. I kinda hope it blows up in their face with a 16gb rdna2 that matches it.

 

[Heartbreaker]

Why would they even release a 20gb 3080 unless that was the super card?

To satisfy these people:

Question - Is 10GB of Vram enough for 4K gaming for the next 3 years?

The question is simple. Will 10GB be enough moving forward for the next 3 years? We all know what happens when Vram gets breached: skips, stutters and chugging.

forums.anandtech.com

 

[Glo.]

Why would they even release a 20gb 3080 unless that was the super card?

To compete with AMD's offering.

 

[sze5003]

To compete with AMD's offering.

I want them to too because that means I can get my 20gb 3080 and bug off for another 2-3 years. Hopefully some concrete news comes after end of October.

 

[Bouowmx]

TechPowerUp review says undervolting is not possible, but undervolting still works the same way as Turing: editing the voltage-frequency curve.

Stock: 345 W
Undervolt to 0.806 V, 1800 MHz: 293 W
Performance in The Division 2 (frame rate): Identical

A more popular tech YouTuber redid the same undervolt (0.806 V, 1800 MHz) on an ASUS TUF GAMING card.
In Horizon Zero Dawn, frame rate in 4K Ultimate decreases from 73 to 72 fps, and power decreases from 320-330 W to 220-230 W. (Better silicon quality?) Other games showed negligible performance difference (within 2%).

Edit: stupid me forgot to put the link.

 

[tviceman]

I'm expecting we see a respin of some sort like we saw with Fermi and 480/70 to 580/70. Better thermals and power consumption, full chip etc. These things are using a ton of power, they have good cooling though.

The FE looks like a nice card for the 3080, but a total waste for the 3090. You want a beefy AIB if you're getting a 3090, even a 3080 really. Higher power limit, better power delivery and a binned chip.

I have my eye on the EVGA ftw 3080, or a Strix 3080.

I hope. The performance per watt improvement between Ampere and Pascal (let alone Ampere and Turing) is less than any generation since OG Fermi. Even the GTX 580 had a better perf/w improvement over the 480 - 98% the same chip on the same node. The reasons it's being more forgiven this time around is because the cooler is great, the price is better than expected, and the outright performance was in line with what we get from Nvidia (not including Turing's lackluster showing).

 

[antihelten]

A more popular tech YouTuber redid the same undervolt (0.806 V, 1800 MHz) on an ASUS TUF GAMING card.
In Horizon Zero Dawn, frame rate in 4K Ultimate decreases from 73 to 72 fps, and power decreases from 320-330 W to 220-230 W. (Better silicon quality?) Other games showed negligible performance difference (within 2%).

Edit: stupid me forgot to put the link.

Apparently the same holds true for the 3090:

https://twitter.com/x/status/1308492945495150599

If this holds true, I think it's quite safe to say that the rumored low performance improvement from the 3090 over the 3080 is not due to the 3090 being power limited.

 

[Qwertilot]

Well, they're definitely not going to have randomly added in 100w of power draw for ~2% performance!

Best guess has presumably to be something like variable quality silicon in the initial batches, with I presume the better quality stuff being harvested & held back to eventually go into notebooks and other such areas.

Fair chances at the mid life refresh slicing into the basic TDPs quite a bit at least.

 

[randomhero]

Have they tested compute workloads?
Games are different beast.

 

[Stuka87]

I am sure there are workloads that are impacted by it. Could be the card is fast enough to be more CPU bound, so no difference shows up. Shaving 100W by undervolting is completely unheard of. Polaris was a great undervolter, but it never got anywhere close to that much.

Even if we look at it as percentage based. Most Polaris cards could save about 15-18%. With these Ampere claims, they are saying 30-35%. So either nVidia set the voltage sky high to get better yields, or the test being run aren't adequately loading down the GPU.

 

[Midwayman]

Well, they're definitely not going to have randomly added in 100w of power draw for ~2% performance!

They would if the total uplift is only 10% with the 100w extra.

 

[Heartbreaker]

Things like undervolting are not guaranteed. Production cards need to have some engineering safety margin. Some users might be able to undervolt significantly, others may not. It's generally not something that can be counted on.

Real reviews are tomorrow.

 

[maddie]

I am sure there are workloads that are impacted by it. Could be the card is fast enough to be more CPU bound, so no difference shows up. Shaving 100W by undervolting is completely unheard of. Polaris was a great undervolter, but it never got anywhere close to that much.

Even if we look at it as percentage based. Most Polaris cards could save about 15-18%. With these Ampere claims, they are saying 30-35%. So either nVidia set the voltage sky high to get better yields, or the test being run aren't adequately loading down the GPU.

1080P relative performance shows the CPU bound situation very clearly. The 3080 performs almost identically at 1080P & 1440P at stock.

We need to look at the undervolt power savings as a %, not in absolute terms.

 

[Glo.]

NVIDIA confirms GeForce RTX 3090 is 10 to 15% faster than RTX 3080 in 4K gaming - VideoCardz.com

NVIDIA published a new blog post outlining the performance of the upcoming flagship ‘BFGPU’. NVIDIA: GeForce RTX 3090 is about 10-15% faster than RTX 3080 in 4K gaming, expect a limited supply The RTX 3090 graphics card featuring 10496 CUDA cores and 24GB of GDDR6X memory is launching tomorrow...

videocardz.com

RTX 3090 is going to be 10-15% faster in 4K than RTX 3080, according to Nvidia.

Yeah...