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

[MrTeal]

Ebay just isn't worth the hassle anymore. I'd glady pay $50+ more for the same thing here or on Amazon. Plus, all the obvious scam listing drown out the few legitimate ones. Every time I see a good price I get excited and jump in only to see a user with 0 feedback.

From both ends. I've had a couple bad experiences selling cards on eBay, to the point I'd rather bin it than go through the hassle of selling it there.

 

[n0x1ous]

yep, FB marketplace is so much easier and you don't have to worry about getting scammed by a buyer. Sold my other 2080 ti FE a month ago for $900 cash

 

[sze5003]

Ebay just isn't worth the hassle anymore. I'd glady pay $50+ more for the same thing here or on Amazon. Plus, all the obvious scam listing drown out the few legitimate ones.

Looking for a 2080S on ebay sucks - every time I see a good price I get excited and jump in only to see a user with 0 feedback, or a misleading title.

Not only cards but high value electronics in general. Had to do several re listings because scammers would win them and then want them shipped elsewhere. There's a way for them to scam you even with PayPal too. One guy won laptop I was selling and looking up the address it was some area in an outdoor shopping mall.

Could be their business, I don't know but seemed weird, they had feedback too. Then they refuse to pay because they wanted the configuration of the ram changed.

More headache and stress than it's worth dealing with eBay. Better to try and sell them here in the for sale section or possibly reddit too.

 

[raghu78]

Nvidia decided to play into AMD's strategy of pushing their cards well past the perf/w curve. There is simply no way AMD will match the RTX 3080 - it's more than twice as fast as the 5700 XT.

The best big Navi card will slot in between the RTX 3080 and RTX 3070. The question is if AMD will be able to play ball on prices, but it is going to be tough.

Based on digitalfoundry's preview RTX 3080 is around 30% faster than RTX 2080 Ti. RTX 2080 Ti is 53% faster than RX 5700XT at 4K.

Grafikkarten-Rangliste 2021: GPUs im Vergleich

Welche Grafikkarte kaufen? Für Juli 2021 gibt es neue Empfehlungen für Nvidia GeForce RTX und AMD Radeon RX von Einstieg bis High-End.

www.computerbase.de

That makes the RTX 3080 Ti two times as fast as RX 5700XT. Now just consider that Navi 2x stack will have 3 dies (with atleast 6 - 7 SKUs) all performing well above RX 5700XT

Navi 21 - 80 CU
Navi 22 - 60 CU
Navi 23 - 40 CU

We know RDNA2 is very efficient and more so than Ampere given the RTX 3070 and RTX 3080 TDP. Matching or even beating the RTX 3080 is not as difficult as you are making it out to be. The question is can Navi 21 80 CU match the RTX 3090 ?

 

[Konan]

Based on digitalfoundry's preview RTX 3080 is around 30% faster than RTX 2080 Ti. RTX 2080 Ti is 53% faster than RX 5700XT at 4K.

Grafikkarten-Rangliste 2021: GPUs im Vergleich

Welche Grafikkarte kaufen? Für Juli 2021 gibt es neue Empfehlungen für Nvidia GeForce RTX und AMD Radeon RX von Einstieg bis High-End.

www.computerbase.de

That makes the RTX 3080 Ti two times as fast as RX 5700XT. Now just consider that Navi 2x stack will have 3 dies (with atleast 6 - 7 SKUs) all performing well above RX 5700XT

Navi 21 - 80 CU
Navi 22 - 60 CU
Navi 23 - 40 CU

We know RDNA2 is very efficient and more so than Ampere given the RTX 3070 and RTX 3080 TDP. Matching or even beating the RTX 3080 is not as difficult as you are making it out to be. The question is can Navi 21 80 CU match the RTX 3090 ?

No.
1. The DF preview was much higher than 30% with RTX OFF and DLSS OFF - best to wait for proper reviews.
2. There is no 3080Ti

2 X 5700 XT = about 30% better than a 2080Ti

Edit: you have no idea how RDNA2 will scale to a large chip. Looks great on smaller ones. A 80CU could be a nuclear reactor for all we know and the mhz could drop off significantly. Say 5120SP @ less than 2k memory

 

[Glo.]

https://twitter.com/i/web/status/1301246897202573319

Your posts must contain your own personal written content, and cannot consists of content from others, be it a quote, an image, a tweet, etc.

AT Moderator ElFenix

 

[raghu78]

No.
1. The DF preview was much higher than 30% with RTX OFF and DLSS OFF - best to wait for proper reviews.
2. There is no 3080Ti

2 X 5700 XT = about 30% better than a 2080Ti

Edit: you have no idea how RDNA2 will scale to a large chip. Looks great on smaller ones. A 80CU could be a nuclear reactor for all we know and the mhz could drop off significantly. Say 5120SP @ less than 2k memory

The bolded part is the only thing I agree. 4K with TAA could be picked to show the RTX 3080 with 760 GB/s in BEST LIGHT against the RTX 2080 with 448 GB/s we will have to wait for press reviews. Anyway RDNA2 is not far behind. I think we will see it launch in Oct. So better to wait and see how these products actually perform.

 

[Hitman928]

https://twitter.com/i/web/status/1301246897202573319

So confirms the advancement of 1 fp + 1 int inside the Turing architecture to 1 fp + 1 fp or 1 int inside of Ampere.

 

[Saylick]

Radeon VII was $700 with 16 GB of HBM. I bet both TSMC 7nm and HBM are more affordable now than they were in early 2019.

Radeon VII was built on TSMC N7, if I'm not mistaken, at a die size of 331 mm2. Big Navi will be on TSMC's N7P process, which will have better yields but also be at a 50% larger die size (assuming 500mm2 die). I think yields and die cost will be comparable between the Radeon VII at 331mm2 and Big Navi at 500mm2.

 

[Konan]

Radeon VII was built on TSMC N7, if I'm not mistaken, at a die size of 331 mm2. Big Navi will be on TSMC's N7P process, which will have better yields but also be at a 50% larger die size (assuming 500mm2 die). I think yields and die cost will be comparable between the Radeon VII at 331mm2 and Big Navi at 500mm2.

RDNA1 Navi is on N7P. We're unsure of what RDNA2 will be on if not the same. I agree re: costs and yields comparable. ''Maybe" RNDA2 will have HBM2 and G6 memory according too _rogame on twitter. I think that Big Ampere and Big Navi will have node density parity (or be pretty close).

 

[GodisanAtheist]

Ebay just isn't worth the hassle anymore. I'd glady pay $50+ more for the same thing here or on Amazon. Plus, all the obvious scam listing drown out the few legitimate ones.

Looking for a 2080S on ebay sucks - every time I see a good price I get excited and jump in only to see a user with 0 feedback, or a misleading title.

- Odd, I've personally never had a bad experience buying anything from Ebay. I just make sure the seller is in the same country and its not one of those "too good to be true" type of deals (generally in line with other items of the same price) and I haven't been steered wrong yet.

yep, FB marketplace is so much easier and you don't have to worry about getting scammed by a buyer. Sold my other 2080 ti FE a month ago for $900 cash

Not only cards but high value electronics in general. Had to do several re listings because scammers would win them and then want them shipped elsewhere. There's a way for them to scam you even with PayPal too. One guy won laptop I was selling and looking up the address it was some area in an outdoor shopping mall.

Could be their business, I don't know but seemed weird, they had feedback too. Then they refuse to pay because they wanted the configuration of the ram changed.

More headache and stress than it's worth dealing with eBay. Better to try and sell them here in the for sale section or possibly reddit too.

- I'll check out some other marketplaces as well though and see what I can see there. Never bought anything through the Anandtech For Sale section (looks like a bit of a hassle with whatever heatware is) but I'd rather give my money to someone else on these forums than some rando on ebay.

 

[JoeRambo]

So confirms the advancement of 1 fp + 1 int inside the Turing architecture to 1 fp + 1 fp or 1 int inside of Ampere.

And if i understood correctly, they have also doubled ROPs per GPC. Incredible for NV to go for such a brutal GPU.

These ROPs have nasty implications for AMD who are stuck on 64 ROPs for nearly a decade now? Really worried about AMD.

 

[AtenRa]

Radeon VII was built on TSMC N7, if I'm not mistaken, at a die size of 331 mm2. Big Navi will be on TSMC's N7P process, which will have better yields but also be at a 50% larger die size (assuming 500mm2 die). I think yields and die cost will be comparable between the Radeon VII at 331mm2 and Big Navi at 500mm2.

Perhaps, but dont forget that GA102 is even larger at 627mm2 that makes yields even lower, increasing the cost further.
So even if SS 8N wafers are cheaper, yields on the 627mm2 die will be lower.

NVIDIA now will be forced to sell a bigger die (627mm2) to compete with a smaller die (500mm2).

 

[Konan]

NVIDIA now will be forced to sell a bigger die (627mm2) to compete with a smaller die (500mm2).

That's nothing new for them though, Turing, skip back a couple more gens and it was rinse and repeat.

 

[Saylick]

Perhaps, but dont forget that GA102 is even larger at 627mm2 that makes yields even lower, increasing the cost further.
So even if SS 8N wafers are cheaper, yields on the 627mm2 die will be lower.

NVIDIA now will be forced to sell a bigger die (627mm2) to compete with a smaller die (500mm2).

I wouldn't be surprised if the die costs end up being comparable between Big Navi at a low defect, high yielding but more expensive TSMC N7P and GA102 at 20%+ larger die size on a higher defect, lower yielding node but less expensive SS 8N. Both companies are getting comparable density on their respective nodes and if we assume that Big Navi at 500mm2 is roughly as performant at a 20% cut-down 627mm2 GA102, or roughly 500mm2 of effective silicon, then that tells us something.

 

[Tup3x]

I wouldn't be surprised if the die costs end up being comparable between Big Navi at a low defect, high yielding but more expensive TSMC N7P and GA102 at 20%+ larger die size on a higher defect, lower yielding node but less expensive SS 8N. Both companies are getting comparable density on their respective nodes and if we assume that Big Navi at 500mm2 is roughly as performant at a 20% cut-down 627mm2 GA102, or roughly 500mm2 of effective silicon, then that tells us something.

I guess that's why 3080 uses GA102 this time. The defective ones go there.

 

[sze5003]

- Odd, I've personally never had a bad experience buying anything from Ebay. I just make sure the seller is in the same country and its not one of those "too good to be true" type of deals (generally in line with other items of the same price) and I haven't been steered wrong yet.






- I'll check out some other marketplaces as well though and see what I can see there. Never bought anything through the Anandtech For Sale section (looks like a bit of a hassle with whatever heatware is) but I'd rather give my money to someone else on these forums than some rando on ebay.

Heatware is just a site that has feedback people can leave you. You can link your Anandtech or overclockers(and whatever other forum they support ), user id there and people will leave feedback for the sales you do in these forums. Takes 5 min to set up. I've bought and sold to people here and it was 100% super easy and stress free.

 

[Kenmitch]

- Odd, I've personally never had a bad experience buying anything from Ebay. I just make sure the seller is in the same country and its not one of those "too good to be true" type of deals (generally in line with other items of the same price) and I haven't been steered wrong yet.

Ebay is a buyers market which heavily favors the buyer/scammer. I buy from Ebay reputable sellers, but won't even consider selling there anymore.

 

[AtenRa]

I wouldn't be surprised if the die costs end up being comparable between Big Navi at a low defect, high yielding but more expensive TSMC N7P and GA102 at 20%+ larger die size on a higher defect, lower yielding node but less expensive SS 8N. Both companies are getting comparable density on their respective nodes and if we assume that Big Navi at 500mm2 is roughly as performant at a 20% cut-down 627mm2 GA102, or roughly 500mm2 of effective silicon, then that tells us something.

I dont know about the density, TSMC 7nm can reach up to 63M transistors per mm2 (Ryzen SOC and NVIDIA A100) when Ampere GA102 at Samsung 8nm reaching close to 45M transistors per mm2
But lets assume they have almost the same density, the AMD smaller die can have the same performance as the NVIDIA bigger die simple because AMD doesn't use Tensor Cores. Also if AMD will use HBM they will also use less space for its memory controllers vs NVIDIA with its GDDR-6X.

So i will say that a Big Navi could compete with 3080 even in price.

 

[DDH]

No.
1. The DF preview was much higher than 30% with RTX OFF and DLSS OFF - best to wait for proper reviews.
2. There is no 3080Ti

2 X 5700 XT = about 30% better than a 2080Ti

Edit: you have no idea how RDNA2 will scale to a large chip. Looks great on smaller ones. A 80CU could be a nuclear reactor for all we know and the mhz could drop off significantly. Say 5120SP @ less than 2k memory

Out of the 4 titles shown on the DF vid, the average of the uplift vs the 2080ti is 35%

All the titles chosen were to definitely hamstring the 8gb 448gbs 2080 memory, nevertheless the performance improvement is impressive.

The difference between the 3080 and the 2080 was 35% for control with RT and DLSS on and off ie the RT performance between the 3080 and 2080 was 1:1, which seems to indicate perhaps a limitation or bottleneck somewhere in that bench maybe. Hardware unboxed noticed the same thing.


We really have no idea where AMD is at with regards to their development and performance of RDNA2. So we can't rule one way or the other.

We have seen a openvr bench of an AMD GPU 30% faster than a 2080ti, so there is that

Anyway, 3080 at 699 is a pretty good deal tbh. Hope AMD gives us even more

 

[maddie]

We should all realize that SS8, at least for big die must have terrible parametric yield.

 

[tviceman]

Big words as always. We'll see.

"As always" haha I've barely posted in the past 4 years. My words are more like farts in the wind.

I see it this way: AMD needs perfect 2x scaling in performance AND a 25% TGB perf/w improvement just to match RTX 3080. Unfortunately for reality, perfect scaling is about as common as a unicorn. As far as efficiency is concerned, there is definitely an opportunity to improve it, but with AMD I won't hold my breath.

I look forward to being wrong, but I don't see it. I think Big Navi will be ~75% faster than 5700 XT in an absolute best case scenario.

 

[DDH]

"As always" haha I've barely posted in the past 4 years. My words are more like farts in the wind.

I see it this way: AMD needs perfect 2x scaling in performance AND a 25% TGB perf/w improvement just to match RTX 3080. Unfortunately for reality, perfect scaling is about as common as a unicorn. As far as efficiency is concerned, there is definitely an opportunity to improve it, but with AMD I won't hold my breath.

I look forward to being wrong, but I don't see it. I think Big Navi will be ~75% faster than 5700 XT in an absolute best case scenario.

Big Navi is not just a big Navi10, rdna2 is a change of architecture. So no linear scaling isnt necessarily relevant

 

[raghu78]

And if i understood correctly, they have also doubled ROPs per GPC. Incredible for NV to go for such a brutal GPU.

These ROPs have nasty implications for AMD who are stuck on 64 ROPs for nearly a decade now? Really worried about AMD.

We know Xbox Series X is a 2SE, 4SA ( 4 x 7WGP/14CU = 56CU) , 64 ROP design with 116 Gpixels/s . The 116 Gpix/s ( 64 x 1.825 Ghz = 116.8) was revealed at Hotchips 2020.

We also know Navi 21 has 4 RBE per SE and has 4 SE as per rogame's findings. So 4 x 4 = 16 RBE.



Navi 10 had 8 RBE per SE ( 4 per SA) with 4 ROP per RBE. Navi 10 had 2 SE for a total of 2 SE x 2SA x 4RBE x 4 = 64 ROP .

Given both Series X and Navi 21 are RDNA2 GPUs I believe the RBE per SE is same in both these GPUs. This would mean AMD has doubled the ROPs per RBE from 4 to 8. Incidentally Ampere has 2 RBE per GPC with 8 ROPs per RBE.

Series X = 2SE x 2SA x 2 RBE x 8 = 64 ROPs
Navi 21 = 4SE x 2SA x 2 RBE x 8 = 128 ROPs
Navi 22 = 3SE x 2SA x 2 RBE x 8 = 96 ROPs
Navi 23 = 2SE x 2SA x 2 RBE x 8 = 64 ROPs

 

[Stuka87]

And if i understood correctly, they have also doubled ROPs per GPC. Incredible for NV to go for such a brutal GPU.

These ROPs have nasty implications for AMD who are stuck on 64 ROPs for nearly a decade now? Really worried about AMD.

To be clear, nVidia ROPs and AMD ROPs are not equivalent. They ultimately do the same job, but aren't 1:1 the same.

And NOBODY had 64 ROPs 10 years ago. The first GPU to ever feature 64 ROPs was in 2014 with GM204 (GTX-980).