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

[Mopetar]

The 3090 is a dumpster fire.

I know it's a big card, but you probably don't need a dumpster for a case just to fit it.

 

[Kenmitch]

It might be power limited for FE, but it won't for AIB cards that aren't as power limited. With 20% more functional units, ~20% gains will happen.

I'm sure this fugly looking one won't be the only one with increased power limit....If you got the power to burn and are willing to deal with the heat.

Custom 410W GeForce RTX 3090 Brings The Bling And The Heat

No need for a space heater this winter.

www.tomshardware.com

 

[antihelten]

It might be power limited for FE, but it won't for AIB cards that aren't as power limited. With 20% more functional units, ~20% gains will happen.

AIB aren't really relevant here, since the discussion isn't how AIB are positioning their cards but rather how Nvidia is positioning them.

Also it is perfectly possible that there are other bottlenecks in the design that is limiting the 3090, in fact the experiments that some sites have done with power limiting the 3080 would suggest that this is likely the case (i.e. Computerbase showed that decreasing the power limit of the 3080 by over 15% only gave a 4% performance hit, or inversely that increasing the power limit by almost 20% only gives a 4% performance boost).

Even if they didn't, 3080 Ti, does NOT have to be slower than 3090.

Go back and check Kepler 780, Titan and 780 Ti. 780 Ti outperformed the Titan.

Also note that was the last time that x80, Titan and x80 Ti all shared the same die.

'
Honestly if you have to go all the way back to the Kepler generation to try and make your point, then it doesn't really get much weaker than that.

I mean the Nvidia basically invented an entire generation (700) based of a single die (almost everything else was just rebranded stuff). Obviously that will lead to some significant wonkiness with SKUs.

 

[Grooveriding]

If this rumoured 3090 performance is accurate, seems pretty crappy all round. Hardly any more performance than 3080 - but twice the cost, uses tons of power and has to be cooled for it.

These cards seem ripe for what we saw with 480/470. Some sort of respin with better power consumption and thermals and the full chip for the 3090 refresh.

 

[swilli89]

I was thinking about it today - NVIDIA is already fabbing the Ampere architecture on TSMC 7nm with the A100. I could see them porting A102 and possibly even A104 to TSMC for an immediate 25% boost if they wanted to next year. After all, NVIDIA DID run both Samsung and TSMC for Turing cards as well so they might whip out a tweaked 7nm Ampere for the 4000 series. Given what we know about density between the two processes a A102 (3080) would drop from 628mm2 to about 450mm2 and consume a good bit less power to boot.

 

[guidryp]

I was thinking about it today - NVIDIA is already fabbing the Ampere architecture on TSMC 7nm with the A100. I could see them porting A102 and possibly even A104 to TSMC for an immediate 25% boost if they wanted to next year. After all, NVIDIA DID run both Samsung and TSMC for Turing cards as well so they might whip out a tweaked 7nm Ampere for the 4000 series. Given what we know about density between the two processes a A102 (3080) would drop from 628mm2 to about 450mm2 and consume a good bit less power to boot.

People said similar things about Turing. Soon see the 7nm version in 6 to 12 months max.

They won't. It's tens of millions of dollars taping out and getting masking layers made for a chip. They want to get a good 2 year run out of the chip to pay back those big up front costs.

 

[Saylick]

People said similar things about Turing. Soon see the 7nm version in 6 to 12 months max.

They won't. It's tens of millions of dollars taping out and getting masking layers made for a chip. They want to get a good 2 year run out of the chip to pay back those big up front costs.

Good luck to Nvidia on getting 2 years out of SS 8N when AMD is poised to drop RDNA3 in 16 months on an Advanced Node, probably TSMC N5. I think Nvidia have to move to a chiplet based design and/or new node, either TSMC N5 or SS 7LPE, or else they'd face steep competition next generation. I also don't expect a monolithic design to be able to compete financially against a chiplet design, which is what RDNA3 will likely be.

 

[guidryp]

Good luck to Nvidia on getting 2 years out of SS 8N when AMD is poised to drop RDNA3 in 16 months on an Advanced Node, probably TSMC N5. I think Nvidia have to move to a chiplet based design and/or new node, either TSMC N5 or SS 7LPE, or else they'd face steep competition next generation. I also don't expect a monolithic design to be able to compete financially against a chiplet design, which is what RDNA3 will likely be.

People probably said good luck on getting 2 years out of 12nm FFN, with AMD poised to drop 7nm next gen...

Poised to Drop RDNA3? Is it time to start the the "wait for RDNA3" posts already?

You can't count your chickens before they're hatched. You should probably wait to to see what AMD delivers before pronouncing Ampere on 8N dead.

 

[moonbogg]

I'd like Nvidia and their partners to immediately release RTX 3080 20GB models for not much more money so all the scalpers will be stuck with the crappy 10GB models and anyone who bought one over MSRP would get equally screwed. That would be pretty epic.

 

[uzzi38]

I'd like Nvidia and their partners to immediately release RTX 3080 20GB models for not much more money so all the scalpers will be stuck with the crappy 10GB models and anyone who bought one over MSRP would get equally screwed. That would be pretty epic.

Hahahahahahahahahahahahaha

Nah. That'll be 1K+ buckaroos please.

https://twitter.com/i/web/status/1307917693312528384

 

[coercitiv]

Is it time to start the the "wait for RDNA3" posts already?

You can't count your chickens before they're hatched.

Funny you should say that, considering all the "wait for 3080Ti" I've been reading on this thread.

 

[guidryp]

Funny you should say that, considering all the "wait for 3080Ti" I've been reading on this thread.

Certainly not from me. I expect that is months away, but a lot closer than RDNA3 I would bet.

 

[StinkyPinky]

Well this all turned out to be pretty crappy. 20GB 3080 seems like the only reasonable 4K card.....assuming that exists.

Big Navi better be good. I'm definitely waiting for that now, especially since that is what the Xbox will be based around.

 

[sze5003]

I'm really curious of the 3090 reviews that will come this week. I too will now be waiting for a higher Vram version of the 3080. I feel bad for a friend of mine who sold his 1080ti for $450 and now he's without a card.

Luckily I still have mine and won't be buying anything until I have a better view of the market in a month or two. Not like you can get anything anytime soon right now anyways.

 

[n0x1ous]

do we know NDA/review date for 3090?

 

[sze5003]

do we know NDA/review date for 3090?

If it's anything like the 3080, they should be out on the 23rd.

 

[Mopetar]

I doubt NVidia does a TSMC 7nm refresh. Wasn't there a rumor from a while back about them booking some TSMC 5nm wafers?

If they can get enough of those it would be far better just to press ahead and getting to 5nm gives them a huge leg up over Ampere even if they don't make such aggressive parts. Going to 5nm is a bit closer to a node and a half improvement instead of just a full node step since the Samsung process is a lot closer to 10/12nm than it is to 7nm.

Even if Ampere is a bit rough around the edges in terms of the overall product, there's a good reason to be excited about future products. Even a die shrink itself will be a massive improvement, but there's no reason to think NVidia won't make further optimizations and improvements to their architecture.

 

[DooKey]

I doubt NVidia does a TSMC 7nm refresh. Wasn't there a rumor from a while back about them booking some TSMC 5nm wafers?

If they can get enough of those it would be far better just to press ahead and getting to 5nm gives them a huge leg up over Ampere even if they don't make such aggressive parts. Going to 5nm is a bit closer to a node and a half improvement instead of just a full node step since the Samsung process is a lot closer to 10/12nm than it is to 7nm.

Even if Ampere is a bit rough around the edges in terms of the overall product, there's a good reason to be excited about future products. Even a die shrink itself will be a massive improvement, but there's no reason to think NVidia won't make further optimizations and improvements to their architecture.

It was a news release. They have booked 5nm wafers.

 

[guidryp]

It was a news release. They have booked 5nm wafers.

I bet it was another rumor. This stuff tends to be confidential.

 

[Midwayman]

I'm really curious of the 3090 reviews that will come this week.

Leaks I've been hearing seem underwhelming. The 3090 is supposedly heavily power limited. Like 10% faster in most circumstances.

 

[sze5003]

Leaks I've been hearing seem underwhelming. The 3090 is supposedly heavily power limited. Like 10% faster in most circumstances.

Yea if that is the case it's an absurd waste of money to pick it over waiting for a 20gb 3080 (even if the performance of this higher ram card is no different than the 10gb 3080).

 

[guidryp]

Leaks I've been hearing seem underwhelming. The 3090 is supposedly heavily power limited. Like 10% faster in most circumstances.

I have seen 3 leaks. One with a bunch of gaming benches: ~10%. One with compute: ~20%, and one with Synthetic gaming ~20% (same synthetic was showing 10% in other leak).

You would only expect 20% in the best case anyway, so it was NEVER going to be any kind of reasonable gaming value.

It looks like FE card is limited to 350W, which is 10% more power. But AIB cards are 400W+, so go for an AIB card if you are inclined to get one of these.

But really they only makes sense if you are making money with the card, then time is money, and they will pay for themselves shortly.

 

[Mopetar]

The gaming benchmarks aren't surprising. I think NVidia is running into the same issues as AMD was when they made Fury. Sure you have an insane amount of hardware resources and the theoretical performance is jaw dropping, but most games can't utilize all of it and it gets to be a bigger problem as the core count grows increasingly higher.

I'm really interested in how NVidia is going to refine their design going forward. Ampere feels like a design built around making the most of a lackluster Samsung process. Getting back to TSMC will help a lot.

 

[Midwayman]

Yah, if you are going AI compute, or some other professional use where you know for sure that memory is a huge deal it might be worth it. Everyone else it'll be a very expensive e-peen.

 

[Ajay]

It was a news release. They have booked 5nm wafers.

Wonder if they will dual source again on TSMC & SS 5nm.

I bet it was another rumor. This stuff tends to be confidential.

I couldn't find an official statement, so , yes, rumor for now.
If NV is going with TSMC for Hopper, I'm sure they will talk it up early to assure large investors. Then it will leak.