Big Navi is coming

My prediction is 15% behind the 3090...
That's assuming the 3090 is more than 15% faster than the 3080
 
Last edited:
Ampere is no where near tflops performance of Turing. So you need to take 2080ti flops which is really similar to RDNA per flop and calculate that. With your calculations RDNA 2 will be 1.6 times faster than 2080ti.

Youre not calculating the 128mb infinity cache

That is an absolute game changer.

THe 3080 has what? 6mb l3 cache?

This is the big massive hidden weapon AMD will have. The cache can allow significant uplift in the overall performance of the chip. Its not just about clock speed and CUs. Its also about the architecture feeding the Compute Units.

I wished they would release sooner
 
Absolutely. There os no comparing TFs across different architectures. Still, when a 5700xt matches a 2070 Super, it sure is promising.

I am just curious on what voodoo they are using for the bandwidth deficiency.



Paul at RedGamingTech said last week Biggest Navi has a 128MB cache on chip ,but I find that very hard to believe,even on TSMC 7nm+,as SRAM ,etc is damn expensive real estate.
 
Paul at RedGamingTech said last week Biggest Navi has a 128MB cache on chip ,but I find that very hard to believe,even on TSMC 7nm+,as SRAM ,etc is damn expensive real estate.
I was under the impression it was on some sort of infinity bus, so seperated silicon, but I may be mis-remembering or confusing different things at this point, lol. Either ways will be interesting to see how well it works. Here's to hoping.
 
I'm sure that nvidia would be happy to sell them, but I can't shake off the impression that the primary reason for the 3090's existence is in emphasizing how "reasonably" priced the 3080 is.

They'll be glad to sell to gamers, but it seems to me they expect to sell most of these to prosumers. The last gen Titan for prosumers was expensive as hell and this gen it's way more affordable. Remember that Jensen called the 3080 the flagship GPU of this generation. With that said it's obvious that Samsung put a big crimp in what NV was going for this round. It is what it is.
 
They'll be glad to sell to gamers, but it seems to me they expect to sell most of these to prosumers. The last gen Titan for prosumers was expensive as hell and this gen it's way more affordable. Remember that Jensen called the 3080 the flagship GPU of this generation. With that said it's obvious that Samsung put a big crimp in what NV was going for this round. It is what it is.

Yea but 3090 so far has no pro drivers. So it’s is not truly a titan even though Jensen would make it seem that way. You can call it titan level card for gamers without titan pro features. 🤷‍♂️
 
I'm sure that nvidia would be happy to sell them, but I can't shake off the impression that the primary reason for the 3090's existence is in emphasizing how "reasonably" priced the 3080 is.

From a pricing perspective pricing ramps fast as BOM increases (BOM goes through a lot of multipliers on a pricing sheet) and as quantity made goes down (non-reoccuring gets factored in fast). If you only plan on moving a 100,000 units and non reoccuring is $20,000,000 then that’s $200 extra a card just for non-reoccuring costs (engineering, swis, line setups, lrips, Ect.) Then say the higher denisty VRAM is another $200 (pulling that out of thin air to make a point) after pricing multipliers it’s likely more around $300-350 onto the price.

But the 3090 is practically the full size die and the 3080 has parts disabled. So it was going to exist regardless. From a die perspective the 3090 is equvilant to a 3080ti or Titan; depending on the historical launch you are comparing it to.

So my overall point is costs ramp faster than people think - $1499 does seem a little high though. The card would exist regardless though since the “good dies” have to be used someplace.
 
I'm sure that nvidia would be happy to sell them, but I can't shake off the impression that the primary reason for the 3090's existence is in emphasizing how "reasonably" priced the 3080 is.

As the 3080s climb in price with the extra memory, I'm guessing the "average" price is going to settle in around $900 for a 20GB 3080 (maybe your EVGA "Black" cards in the $800s range and the Strix/FTW cards in the $1000 range). I think the whole point of the 10GB cards was to try to gain some good will back after the Turing price debacle. People are going to "think" that the 3080 is reasonably priced when it will actually be comparable or slightly higher than the outgoing 2080 Super that it replaces.
 
3090 @ 5-10% faster for twice the price is the Greatest thing I've seen this year.

for reference:

bench : 3090 : 3080 : % gain (for twice the price XD)


3DMark Time Spy Extreme99489000+10.5%
3DMark Port Royal1282711981+7.1%
Metro Exodus RTX/DLSS OFF54.448.8+10.2%
Metro Exodus RTX/DLSS ON74.567.6+10.2%
Rainbow Six Siege275260+5.8%
Horizon Zero Dawn8476+10.5%
Forza Horizon156149+4.7%
Far Cry10799+8.1%
Assassins Creed Oddysey7165+9.2%
Shadow of the Tomb Raider RTX/DLSS Off9183+9.6%
Shadow of the Tomb Raider RTX/DLSS On111102+8.8%
Borderlands 367.661.3+10.3%
Death Stranding DLSS/RTX ON175164+6.7%
Death Stranding DLSS/RTX OFF116104+11.5%
Control DLSS/RTX ON7165+9.2%
Control DLSS/RTX OFF6257+8.8%
 
Last edited:
It definitely seems to have been marketed as more of a prosumer card, but the pro features are MIA?

Seems that Nvidia knows there is a market willing to pay $1500 plus to say they have the best, while the majority care more about value. Win win I guess.
 
It definitely seems to have been marketed as more of a prosumer card, but the pro features are MIA?

Seems that Nvidia knows there is a market willing to pay $1500 plus to say they have the best, while the majority care more about value. Win win I guess.
Well they know ppl are willing to pay 1k for a 3080 and $800 or so for used 2080tis lol.

Impulse control has been dead for quite some time.
 
Yea but 3090 so far has no pro drivers. So it’s is not truly a titan even though Jensen would make it seem that way. You can call it titan level card for gamers without titan pro features. 🤷‍♂️

I think they had to change the name after the backlash of the "The 2080Ti is the Titan equivalent in this generation" bullshit that the fanboys were claiming after the outrageous price increase over the 1080Ti. Then the RTX Titan dropped with a $2499 price tag.

It sounds like there will be a Ampere Titan at some point. 48GB of memory more cores, and extremely expensive.
 
I think they had to change the name after the backlash of the "The 2080Ti is the Titan equivalent in this generation" bullshit that the fanboys were claiming after the outrageous price increase over the 1080Ti. Then the RTX Titan dropped with a $2499 price tag.

It sounds like there will be a Ampere Titan at some point. 48GB of memory more cores, and extremely expensive.
I mean why not? They know people will pay.
 
Well dang... That not a very good value..
CPU Bottleneck? lol
If that is the case and they can't fix it, then there is a real possibility of Big Navi being faster than the 3090, at least without DLSS...

Doesn't seem like a CPU bottleneck as the 3090 actually lost some lead with RTX on. I think most games can only go so "wide" so a 3080ti that has similiar less tflops but higher boost clocks may actually do better. That, or maybe the 2 Gb modules have more latency or something.
 
Doesn't seem like a CPU bottleneck as the 3090 actually lost some lead with RTX on. I think most games can only go so "wide" so a 3080ti that has similiar less tflops but higher boost clocks may actually do better. That, or maybe the 2 Gb modules have more latency or something.

I'm pretty sure the 2GB models aren't shipping until 2021 as per the press release below. The 3090 PCB has 12 1GB models on the front and 12 on the back.

https://investors.micron.com/news-r...iscrete-graphics-memory-micron-powers-nvidias

GDDR6X is now available as part of Micron’s new Ultra-Bandwidth Solutions portfolio. Micron delivers GDDR6X memory in 8 gigabits (Gb) density, with speeds of 19 to 21 Gb/s. Starting in 2021, 16Gb density units will be added.
 
I have to admit, I am in full popcorn mode.

NV has a really powerful part. But - pretty power hungry, and at least currently, in what could at best be called "limited" supply. The next couple weeks will be informative.
I am interested in how AMD exploits some of these cracks in the armor. Or not.

I also want Su to make a kitchen picture with like > eleven thousand spatulas in the background with no explanation.
 
Ampere is no where near tflops performance of Turing. So you need to take 2080ti flops which is really similar to RDNA per flop and calculate that. With your calculations RDNA 2 will be 1.6 times faster than 2080ti.

??????

The only assumption that I made is that there are 64 shaders per CU (Compute Unit), which is consistent with all the AMD GPUs that I've seen here beginning with RX 400 series.
https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units#Radeon_R5/R7/R9_300_series

Here's the tables from my spreadsheet. I've already calculated just about everything.
gpus_20200919.png
 
Absolutely. There os no comparing TFs across different architectures. Still, when a 5700xt matches a 2070 Super, it sure is promising.

I am just curious on what voodoo they are using for the bandwidth deficiency.
So I read the patent paper. https://adwaitjog.github.io/docs/pdf/sharedl1-pact20.pdf (it's a good read)

Obviously take all this with a pinch of salt, but this new cache hierarchy yields quite a bit of better VRAM bandwidth efficiency for workloads with high data migration.

Basically one of the major challanges when designing a cache is premature eviction of data. Cache thrash I would imagine is a big issue with GPUs in particular since SIMD paradigm works on so much data making GPUs use so much data and bandwidth for rendering. So what AMD has done is made a new architecture which partitions L1 caches in order to improve data locality. The new method also allows L1 sharing across neighboring CUs. Which lowers data migrations significantly. These are two major improvements which can yield massive uplift from one patent.

It remains to be seen how well this pans out in the real world (we know AMD doesn't have the best record of new GPU features panning out, but we also know they are highly capable of pulling this off due to their successes with Ryzen).

We have also seen some confirmation of RDNA2 being really efficient due to what we know about the next gen consoles so far.

So I am cautously optimistic that AMD may give Ampere a run for its money in a big way. Everything points to a real strong offering from AMD, including the superior fab process TSMC is providing (compared to Ampere being on Samsung).
 
Since there is mention of "Pro" and the Titan series, what makes any of the Titan "Pro" exactly? I thought Quadro were the "Pro" cards? Or is there another "Pro" thing about it.

I ran my Titan Xp cards with geforce drivers, there were never any pro drivers / features that I knew of. Or did that start after the Pascal series?
 
So I read the patent paper. https://adwaitjog.github.io/docs/pdf/sharedl1-pact20.pdf (it's a good read)

Obviously take all this with a pinch of salt, but this new cache hierarchy yields quite a bit of better VRAM bandwidth efficiency for workloads with high data migration.

Basically one of the major challanges when designing a cache is premature eviction of data. Cache thrash I would imagine is a big issue with GPUs in particular since SIMD paradigm works on so much data making GPUs use so much data and bandwidth for rendering. So what AMD has done is made a new architecture which partitions L1 caches in order to improve data locality. The new method also allows L1 sharing across neighboring CUs. Which lowers data migrations significantly. These are two major improvements which can yield massive uplift from one patent.

It remains to be seen how well this pans out in the real world (we know AMD doesn't have the best record of new GPU features panning out, but we also know they are highly capable of pulling this off due to their successes with Ryzen).

We have also seen some confirmation of RDNA2 being really efficient due to what we know about the next gen consoles so far.

So I am cautously optimistic that AMD may give Ampere a run for its money in a big way. Everything points to a real strong offering from AMD, including the superior fab process TSMC is providing (compared to Ampere being on Samsung).

Whoa, great find!
 
  • Like
Reactions: muxr
like this
It definitely seems to have been marketed as more of a prosumer card, but the pro features are MIA?

Seems that Nvidia knows there is a market willing to pay $1500 plus to say they have the best, while the majority care more about value. Win win I guess.

I think you nailed it. They positioned the 3080 as the flagship so people can’t blame them for “gouging” like they did with the 2080 Ti. They also marketed the 3090 as a gaming card for the people who won’t blink at spending $1500 for the “best”.

Did they even have a choice? If yields on 3090 aren’t great then there was no other option but to sell it as a low volume / low demand product.
 
So I read the patent paper. https://adwaitjog.github.io/docs/pdf/sharedl1-pact20.pdf (it's a good read)

Obviously take all this with a pinch of salt, but this new cache hierarchy yields quite a bit of better VRAM bandwidth efficiency for workloads with high data migration.

Basically one of the major challanges when designing a cache is premature eviction of data. Cache thrash I would imagine is a big issue with GPUs in particular since SIMD paradigm works on so much data making GPUs use so much data and bandwidth for rendering. So what AMD has done is made a new architecture which partitions L1 caches in order to improve data locality. The new method also allows L1 sharing across neighboring CUs. Which lowers data migrations significantly. These are two major improvements which can yield massive uplift from one patent.

It remains to be seen how well this pans out in the real world (we know AMD doesn't have the best record of new GPU features panning out, but we also know they are highly capable of pulling this off due to their successes with Ryzen).

We have also seen some confirmation of RDNA2 being really efficient due to what we know about the next gen consoles so far.

So I am cautously optimistic that AMD may give Ampere a run for its money in a big way. Everything points to a real strong offering from AMD, including the superior fab process TSMC is providing (compared to Ampere being on Samsung).
Great Stuff! From paper:
  • We extensively evaluate our proposal across 28 GPGPU applications. Our dynamic scheme boosts performance by 22% (up to 52%) and energy efficiency by 49% for the applications that exhibit high data replication and cache sensitivity without degrading the performance of the other applications. This is achieved at a modest area overhead of 0.09 mm2 /core.
  • We make a case to employ our dynamic scheme for deeplearning applications to boost their performance by 2.3×.
 
  • Like
Reactions: muxr
like this
MLID stated that AMD expects to compete with the 3080 and that 2x 5700 XT performance is possible. That's where my expectations are, hopefully the price is good. I'm still leaning to getting a 3080 if I can actually find one.
 
Last edited:
Hmmm......
And AMD said 50% higher perf/w
That is just from the cache, process node, manual layout of chip, slower more power efficient memory can be used. So the rumored 60% perf/w improvement maybe not too far off. Will be interesting if certain game engines or games see more dramatic improvement over others. This is a major change that may also have some negatives, we just have to see how it plays out.
 
I'm dying to learn which is more power efficient, AMD or Nvidia this round. Will it be very different between them?
 
IF AMD can be more power efficient and come within 10% of the performance of the 3080 at a lower cost they will have a winner.
I’d consider the card. I’m still hesitant when it comes to AMD GPU drivers though. Had a not so great experience when I had my R9 390 a few years back. My co-worker has the 5700XT and says he still gets black screen issues in some games (though much improved over the past year). Nvidia drivers have been far less buggy from my personal experience over the past decade.
 
I've never had any serious driver issues with Radeons. I've probably 6-8 different models dating back to 2000. Eyefinity was great and Radeons always seemed to have excellent bang/buck value.
To be fair, I've also had quite a few Geforce models dating back to the TNT2 Ultra. Currently I'm on a Geforce 1070 (4 years now), haven't had any issues with it.

The shenanigans with scalpers and the shortage of 3080's makes me more anxious for AMD to deliver. I belive nVidia is complicit with this bullshit and the retailers are glad to help by doing nothing. Retailers could have implemented better captchas. They've known about high demand for products with many years of console system sales and should have prepared for it. Xbox Series X and PS5 pre-orders all sold out on the first day.

Even if nVidia decides to release abundant supply so prices go back to MSRP and the 3080 and 3070 become widely available from retailers, I'll likely still buy the new $500 Radeon. This assumes scalpers don't snap up all the supply of the Radeons and hold them hostage as well.
 
Last edited:
  • Like
Reactions: muxr
like this
Right, the retailers are acting like this is the first time ever a hot product came out with high demand.

And the scalping / mining crowd has been a known quantity for years. They probably just don't care, they still get their money.
 
Right, the retailers are acting like this is the first time ever a hot product came out with high demand.

And the scalping / mining crowd has been a known quantity for years. They probably just don't care, they still get their money.
Mining wise it looks decent, 100h/s at 250w with one sample for the 3080. Still the 5700XT can do 50-60 h/s, I do 51h/s, if my memory would OC better it would go up from there. Buying used 5700XTs would make more sense but I am sure we will see Mining rigs filled with 3080's, be a lot of hashrate power in one rig. Big Navi with double the CUs, faster clocks except ETH is memory extensive, it may do poorly overall for the cost (lol) yet work great for everything else :ROFLMAO:. Hard to say with the new Cache system, it may cream the hashrate or just stall out.
 
Mining wise it looks decent, 100h/s at 250w with one sample for the 3080. Still the 5700XT can do 50-60 h/s, I do 51h/s, if my memory would OC better it would go up from there. Buying used 5700XTs would make more sense but I am sure we will see Mining rigs filled with 3080's, be a lot of hashrate power in one rig. Big Navi with double the CUs, faster clocks except ETH is memory extensive, it may do poorly overall for the cost (lol) yet work great for everything else :ROFLMAO:. Hard to say with the new Cache system, it may cream the hashrate or just stall out.
Most serious miners don't care about this card, they want the most efficient hash for electricity. Even if they are running off of solar, they have and energy budget of how much energy they can use per day/hour/week/month. The more hashes they can fit into this budget, the more money they make. It's not about buying the biggest and baddest for any serious outfit (well, by and far anyways, obviously there will be some exceptions, I guess if it's a single rig vs 15 rigs the less hardware may be worth it, but if it's like 10% more perf for a rig and it takes 50% more power, probably not). They will see what the card can do, and buy appropriately, but so far it doesn't seem worth the money unless they can get some good under-volting to bring power down and keep hashes up. Anyways, like I said, you're probably pretty safe from the majority of miners. Now, the few that may be space constrained will be more worried about getting the most hashes per space, but this is a much smaller group (but i'm sure it exists and they'll be buying anything that they feel they can make money off of). Luckily right now, mining on GPU's is in a dip in profits, so hopefully not as many are looking at expanding.

Mining wise it looks decent, 100h/s at 250w with one sample for the 3080. Still the 5700XT can do 50-60 h/s, I do 51h/s, if my memory would OC better it would go up from there. Buying used 5700XTs would make more sense but I am sure we will see Mining rigs filled with 3080's, be a lot of hashrate power in one rig. Big Navi with double the CUs, faster clocks except ETH is memory extensive, it may do poorly overall for the cost (lol) yet work great for everything else :ROFLMAO:. Hard to say with the new Cache system, it may cream the hashrate or just stall out.
They said in workloads with data that doesn't change to much you can see around 50% increase, and they even said deeplearning type apps can get up to a 2.3x boost. Although, most crypto tries to constantly hit large dynamic parts of ram (the DAG), so the cache may not be that useful, but it should be useful while it's processing as a scratch pad of quick access memory. My guess is it'll be similar to the past, AMD cards have been better at mining in general (performance / electricity cost, and most of the times better hash rates for cheaper cards).

I'm more interested in the gaming aspect of the GPU's, but I'm not ignorant to mining because those are the guys I have to fight to get a damn graphics card, lol.
 
So what is your real expectation for Big Navi? I mean gaming wise?

I hope for a RX6900 with 80CU + 16GB HBM2e as the flagship for $699, a RX6800 with 62 CU + 16GB GDDR6 for $499 and a RX6700 with 40CU + 12GB GDDR6 for $399. Maybe a RX6600 with 36CU + 8GB GDDR6 for $299. Now that I think about it, it sounds too good to be true.
 
So what is your real expectation for Big Navi? I mean gaming wise?

I hope for a RX6900 with 80CU + 16GB HBM2e as the flagship for $699, a RX6800 with 62 CU + 16GB GDDR6 for $499 and a RX6700 with 40CU + 12GB GDDR6 for $399. Maybe a RX6600 with 36CU + 8GB GDDR6 for $299. Now that I think about it, it sounds too good to be true.
I seriously doubt they going to keep doing HBM for consumer products.
 
So what is your real expectation for Big Navi? I mean gaming wise?

I hope for a RX6900 with 80CU + 16GB HBM2e as the flagship for $699, a RX6800 with 62 CU + 16GB GDDR6 for $499 and a RX6700 with 40CU + 12GB GDDR6 for $399. Maybe a RX6600 with 36CU + 8GB GDDR6 for $299. Now that I think about it, it sounds too good to be true.


The huge problem with those stats and prices,is Dr Su wants to make money,increase margins and get more huge rocks on her fingers,those big diamonds and rubies are not cheap.
 
Back
Top