Rumor: AMD Radeon Big Navi has 16GB VRAM (NAVI21), and 12GB for NAVI22

#5833869

icedman

2020-09-23 06:45

I'll wait to see how this compares to the 3070 thats where I'm aiming performance/costwise.

#5833871

Venix

2020-09-23 07:06

poornaprakash:

Big Navi aka Navi 21 die size of 505mm² is highly unlikely to compete with RTX 3080 die size of 625mm². Unless some revolutionary technology present with Navi 21, its very unlikely to beat a 3080.

Die size is just half the equation Transistor density missing so let's do it proper shall we ? Samsung 8nm Density is 61.18 Million Transistor per 1mm² so die size of 625x61.18 = 38.237.500.000 or 38.25 billion Transistors Tsmc 7nm Density is 96.5 Million Transistor per 1mm² so die size of 505 x 95.3 = 48.732.500.000 or 48.75 billion Transistors ! if anything they have 22% more transistors to do it ! https://semiwiki.com/wp-content/uploads/2019/05/img_5cd106e96a81a.jpg.webp https://fuse.wikichip.org/wp-content/uploads/2018/06/samsung-density-14nm-10nm-8nm.png those are the sources i used to take the numbers i know they are around there not sure if 100% accurate ... that said i think my point comes across fine either way !

#5833881

Darren Hodgson

2020-09-23 08:25

If 12 GB and 16 GB are true for AMD's forthcoming "Big Navi" GPUs then I would argue that those amounts are better for futureproofing than NVIDIA's worryingly low 8 GB and 10 GB for its current high end *affordable* Ampere GPUs. I personally think that the 3080 should have had at least 12 GB and the 3070 10 GB with the 3080 Ti (if that model even exists) having 16 GB. I suspect NVIDIA have used lower amounts of VRAM to keep the prices down after the criticism of the RTX 20000 series launch. I'm interested to see what AMD's RNDA2 cards can do but I am locked into NVIDIA cards, unfortunately, due to owning a ASUS ROG SWIFT PG279Q monitor that only supports G-SYNC and not FreeSync.

#5833897

AlmondMan

2020-09-23 09:12

It takes around 5 years to turn around and build a good chip design, so we're still maybe a generation from the one where they truly catch up. Till then, I don't consider my 5700 XT Red Devil a flop card. The performance is perfect for my 3440x1440 monitor - the RTX 2070 Super would do just as well, but the first gen RT cards just can't run it as well as you'd like, so I wasn't going to pay an extra 25-30% for one. Value, that's all that matters. No brand loyalty or other things. The RTX part just wasn't enough value yet to be worth it. I will certainly consider the next gen stuff now that it's out, but I'll wait and see where AMD positions itself. Because there's value to be had in waiting.

#5833935

Fediuld

2020-09-23 10:47

poornaprakash:

Big Navi aka Navi 21 die size of 505mm² is highly unlikely to compete with RTX 3080 die size of 625mm². Unless some revolutionary technology present with Navi 21, its very unlikely to beat a 3080.

Nvidia is on Samsung 8nm (8LLP) which is actually 10nm process with 10% lower power consumption and 10% smaller die. No other enhancements or benefits. AMD is on TSMC N7P not even N7 any more, not to be confused with N7+, having 33% better transistor density than the Samsung 8LLP and lower power consumption and higher clocks than the N7 (5700XT) Also arc matters. Whole Xbox X Series APU is 360mm2 that includes GPU, I/O and 8 Zen 2 cores. Yet is expected to be faster than the 2080Ti so around 3070 perf level. In addition RTX cards have a lot of space taken for Tensor Cores to work with DLSS.

#5834049

JamesSneed

2020-09-23 15:42

poornaprakash:

Big Navi aka Navi 21 die size of 505mm² is highly unlikely to compete with RTX 3080 die size of 625mm². Unless some revolutionary technology present with Navi 21, its very unlikely to beat a 3080.

I wouldn't count them out yet. The TSMC 7nm process should be about 15% more efficient than the 8nm Samsung process which will help close the gap on 20% less die size. They should be able to run at higher frequencies with AMD's approach and stay within a reasonable TDP. Anyhow I personally am waiting to see what AMD has because whats a couple months.

#5834111

CPC_RedDawn

2020-09-23 17:58

Are some people here not reading this correctly? Navi Flounder is the lower end model with Sienna Cichlid being the big boy with 80CU's AMD promised 50% performance per watt when compared to RDNA1......... rumours point towards it actually being 60% Rumours also point towards a massive 128MB cache AMD are calling the "infinity cache" which will help to increase memory bandwidth and narrow the gap between GDDR6 and 6X Couple this with a better 7nm+ process, better IPC, Proven high clockspeeds in "relatively" low powered APU console systems. A "leaked" GPU picture with a heatsink that covers the entire GPU board and requires 2x8pin connectors..... It seems AMD are trying their best to give the best performance they can, even if this falls 10-15% short of the RTX3080 but costs $100-$150 less I know which one I am picking up.

#5834222

Fediuld

2020-09-23 22:40

CPC_RedDawn:

Are some people here not reading this correctly? Navi Flounder is the lower end model with Sienna Cichlid being the big boy with 80CU's AMD promised 50% performance per watt when compared to RDNA1......... rumours point towards it actually being 60% Rumours also point towards a massive 128MB cache AMD are calling the "infinity cache" which will help to increase memory bandwidth and narrow the gap between GDDR6 and 6X Couple this with a better 7nm+ process, better IPC, Proven high clockspeeds in "relatively" low powered APU console systems. A "leaked" GPU picture with a heatsink that covers the entire GPU board and requires 2x8pin connectors..... It seems AMD are trying their best to give the best performance they can, even if this falls 10-15% short of the RTX3080 but costs $100-$150 less I know which one I am picking up.

Yep apparently 80CU exists https://www.pcgamer.com/amp/a-linux-update-may-have-let-slip-amd-big-navis-mammoth-core/

#5834231

Agonist

2020-09-23 23:06

CPC_RedDawn:

Are some people here not reading this correctly? Navi Flounder is the lower end model with Sienna Cichlid being the big boy with 80CU's AMD promised 50% performance per watt when compared to RDNA1......... rumours point towards it actually being 60% Rumours also point towards a massive 128MB cache AMD are calling the "infinity cache" which will help to increase memory bandwidth and narrow the gap between GDDR6 and 6X Couple this with a better 7nm+ process, better IPC, Proven high clockspeeds in "relatively" low powered APU console systems. A "leaked" GPU picture with a heatsink that covers the entire GPU board and requires 2x8pin connectors..... It seems AMD are trying their best to give the best performance they can, even if this falls 10-15% short of the RTX3080 but costs $100-$150 less I know which one I am picking up.

If your close to how it performs, I will be all over the SC navi. my 5700xt is nice and does well but damn would the big navi top dog fly. Currently on 3440x1440 100hz. And since I upgrade monitors every 5 years. I still have 4 years on this one and its nice as hell. Only missing HDR but its not a deal breaker.

#5834351

panogr

2020-09-24 09:45

i`m all in for big navi , just like that 😉

#5834369

Maddness

2020-09-24 10:14

if the rumored 6900xt big Navi for $499 is true, it'll be impossible to ignore. https://www.tweaktown.com/news/75291/amd-radeon-rx-6900-xt-specs-navi-21-gpu-with-16gb-could-cost-499/index.html

#5834381

Fediuld

2020-09-24 10:59

Maddness:

if the rumored 6900xt big Navi for $499 is true, it'll be impossible to ignore. https://www.tweaktown.com/news/75291/amd-radeon-rx-6900-xt-specs-navi-21-gpu-with-16gb-could-cost-499/index.html

Surely something is wrong. AMD selling a 80CU graphic card for $500? That would be over twice as powerful than the 5700XT to start with.

#5834389

alanm

2020-09-24 11:26

Maddness:

if the rumored 6900xt big Navi for $499 is true, it'll be impossible to ignore. https://www.tweaktown.com/news/75291/amd-radeon-rx-6900-xt-specs-navi-21-gpu-with-16gb-could-cost-499/index.html

The only way they would sell it for $499 is if its performance was no better than a 2080ti (3070).

#5834399

Maddness

2020-09-24 11:51

alanm:

The only way they would sell it for $499 is if its performance was no better than a 2080ti (3070).

Who knows, but it's still a rumor. AMD has been very tight lipped about this release. I'm hoping in a good way.

#5834404

Undying

2020-09-24 11:56

alanm:

The only way they would sell it for $499 is if its performance was no better than a 2080ti (3070).

AMD always offered more for you buck. Card could be very well between 3070 and 3080 with double the frame buffer. I know i will want one over nvidia offerings.

#5834423

Embra

2020-09-24 12:42

Let's not runway with rumors. It can only lead to disappointments that should not exist. A lot of saltiness going around. Patience.

#5834473

H83

2020-09-24 14:58

500€ for their top of the line card? Don´t believe it unless the card underperforms the ones from Nvidia considerably.

#5834694

CPC_RedDawn

2020-09-24 21:24

Agonist:

If your close to how it performs, I will be all over the SC navi. my 5700xt is nice and does well but damn would the big navi top dog fly. Currently on 3440x1440 100hz. And since I upgrade monitors every 5 years. I still have 4 years on this one and its nice as hell. Only missing HDR but its not a deal breaker.

Sounds good, if you are happy with it then make it last as long as you can. An 80CU card would be double the performance of your 5700XT. But then you have to add in architecture improvements, IPC gains, higher clockspeeds, and faster memory. It would be a lot more than double.... if all the stars align that is.

Maddness:

if the rumored 6900xt big Navi for $499 is true, it'll be impossible to ignore. https://www.tweaktown.com/news/75291/amd-radeon-rx-6900-xt-specs-navi-21-gpu-with-16gb-could-cost-499/index.html

If it is $500 then this worries me just like the 256bit bus does. On an 80CU chip with these rumoured clock speeds it would seriously bottleneck the chip. Going to a 384bit bus or higher would help greatly but the easiest way to solve the issue would be to just go with HBM2 then a chip like this would really fly. But everytime HBM2 is mentioned it gets shot down so I doubt it. Some form a caching system is the only other way around it, makes sense if AMD want to keep costs down as going above 256bit adds more expense and HBM2 is still expensive to make. But they did it with Radeon VII and that had 60CU's so why not a 6900XT unless they were keeping HBM2 for their prosumer cards or an AMD TITAN?

Embra:

Let's not runway with rumors. It can only lead to disappointments that should not exist. A lot of saltiness going around. Patience.

I agree, but speculating is fun! 😀

#5834723

Neo Cyrus

2020-09-24 22:46

CPC_RedDawn:

If it is $500 then this worries me just like the 256bit bus does. On an 80CU chip with these rumoured clock speeds it would seriously bottleneck the chip. Going to a 384bit bus or higher would help greatly but the easiest way to solve the issue would be to just go with HBM2 then a chip like this would really fly. But everytime HBM2 is mentioned it gets shot down so I doubt it. Some form a caching system is the only other way around it, makes sense if AMD want to keep costs down as going above 256bit adds more expense and HBM2 is still expensive to make. But they did it with Radeon VII and that had 60CU's so why not a 6900XT unless they were keeping HBM2 for their prosumer cards or an AMD TITAN?

It's not going to have memory bandwidth issues. It has a new caching system, with a mountain of L3 cache (vs the whopping zero in the past), which dramatically increases effective RAM throughput so even with a 256 bit bus width it'd be fine. Also, I don't know what the final product will be, but initially it was 384 bit, with yes, 16GB, 384 bit apparently with 16GB of RAM. Don't ask me how that works. If they cut it down to 256 bit it'll be because the new caching system was more than enough.

#5835002

CPC_RedDawn

2020-09-25 17:48

Neo Cyrus:

It's not going to have memory bandwidth issues. It has a new caching system, with a mountain of L3 cache (vs the whopping zero in the past), which dramatically increases effective RAM throughput so even with a 256 bit bus width it'd be fine. Also, I don't know what the final product will be, but initially it was 384 bit, with yes, 16GB, 384 bit apparently with 16GB of RAM. Don't ask me how that works. If they cut it down to 256 bit it'll be because the new caching system was more than enough.

Yea I've heard of this "infinity cache" rumour. Its starting to make more sense now though as we have basically had confirmation of an 80CU card and a standard 256bit bus would seriously hold it back but with this caching system it would help a lot.