Support NeoGAF

[Ascend]

New leaks... Here's a screenshot of how the information was retrieved...

Short version;

Navi 21 - 80CUs, maxing out at 2.2GHz = 22.5TFLOPS
Navi 22 - 40CUs, maxing out at 2.5GHz = 12.8TFLOPS
Navi 23 - 32CUs, No power tables yet
Navi 31 - 80CUs with identical configuration as Navi 21; possibly RDNA3?
Van Gogh APU - 1 SA with 8 RDNA CUs = 8 CUs
Rembrandt APU - 2 SAs with 6 RDNA CUs each = 12 CUs

Source

Edit:
Newegg has some additional information on bandwidth and bus configuration. See this post;

Navi 21 possibly runs at 2.2GHz with 80CUs, Navi 22 at 2.5GHz with 40 CUs

as you can see below, the rtx 2080 ti is on average 56% faster than the 5700 xt. It's 46% at 1440p (the resolution people actually use these cards for). With a minimum of 30% clock speed increase and some IPC improvements, Navi 22 should be on a 2080Ti level.

www.neogaf.com

 

[BattleScar]

Navi 22 is absolutely flying, damn. 30% higher clockspeeds than the 5700XT, at 10W less power on the same node.
That could get awfully close to the 2080Ti, with 70% fewer shaders.

 

[pullcounter]

22.5 Tflops, is that FP16 or FP32?

edit: its fp32

 

[Soodanim]

I know too little to know how these numbers compare to anything on the market. Would anyone care to enlighten myself and the millions like me?

 

[martino]

i like what i read .
i'm glad i've followed my

 

[JohnnyFootball]

And the hype begins. I have little doubt it would be a good video card, but AMD needs it to be a game changer and not just a water treader like the 5700s were.

 

[SantaC]

I know too little to know how these numbers compare to anything on the market. Would anyone care to enlighten myself and the millions like me?

Navi 21 likely to perform between 3070 and 3080

 

[00_Zer0]

I was planning on a RTX 3080 this time around, but with all the screw ups like shortages, and black screens because of some AIB manufacturers cheaping out on quality parts, I will wait for these cards to be revealed before I make up my mind on what to purchase.

 

[Ascend]

I know too little to know how these numbers compare to anything on the market. Would anyone care to enlighten myself and the millions like me?

Basically it should be an RTX 3080 contender.

 

[Soodanim]

Navi 21 likely to perform between 3070 and 3080

Thanks. That, if it goes as well as Ryzen, could be a huge upset for NVidia given that the 70-80 market is their biggest area as far as I'm aware. The right price point and they're laughing, right?

Of course there's drivers and all that, NVidia still do better on that front from what I hear.

 

[Orta]

A 3070ti in AMD clothing then?

 

[sircaw]

Where does this leave them though?

What will they hope to be competing with on the same level with Nvidia.

A 3080?

 

[Caio]

Could a Custom Navi 31 (80CUs with identical configuration as Navi 21; possibly RDNA3) be the GPU of a hypothetical PS5 PRO coming out in Nov 2022 ? Just wondering if this would be feasible in terms of cost and heating dissipation in two years from now.

 

[SlimySnake]

Navi 22 is absolutely flying, damn. 30% higher clockspeeds than the 5700XT, at 10W less power on the same node.
That could get awfully close to the 2080Ti, with 70% fewer shaders.

It's not the same node.

And it will be roughly 13 tflops which will only equal 2080 super. 2080 ti is around 17 tflops at in game clocks.

 

[Ascend]

Where does this leave them though?

What will they hope to be competing with on the same level with Nvidia.

A 3080?

I did an estimation a while back. Basically, if the 80CU Big Navi runs at 2.2 GHz, it should be about 10% faster than a 3080, provided there are no bandwidth bottlenecks or other instabilities/limitations. See here;

Why you shouldn't underestimate Radeon 6000/Big Navi

Introduction With nVidia announcing their RTX 3000 series, everyone is jumping up and down from excitement. The cards are extremely fast, relatively well-priced, and everyone is already planning which card they will be getting. Many are saying that there is no way AMD can catch up, and well, I'm...

www.neogaf.com

 

[onQ123]

The good news is that PS5 isn't pushing the limits of the clockrate.

 

[cucuchu]

This will put them between a 3070 and 3080 on the high end. Probably closer to a 3070 in practical performance...assuming a 3070 is a bit better vs a 2080ti. I imagine they wiIl aggressively price is below the 3070 though. It will be a solid choice for mid-high range GPU's but NVIDIA will continue to dominate the high end.

I wouldn't take the 'AMD said they are going to do to NVIDIA what they did to intel' bit seriously. AMD has a history of overhyping.

 

[Ascend]

It's not the same node.

And it will be roughly 13 tflops which will only equal 2080 super. 2080 ti is around 17 tflops at in game clocks.

I don't think the 2080Ti could reach 17TF under anything but LN2. I'd say 15 is accurate. Feel free to prove me wrong.

This will put them between a 3070 and 3080 on the high end. Probably closer to a 3070 in practical performance...assuming a 3070 is a bit better vs a 2080ti. I imagine they wiIl aggressively price is below the 3070 though. It will be a solid choice for mid-high range GPU's but NVIDIA will continue to dominate the high end.

I wouldn't take the 'AMD said they are going to do to NVIDIA what they did to intel' bit seriously. AMD has a history of overhyping.

The 2080Ti is a 'mere' 35%-40% faster than the 5700XT on average. Do you really think that a 100% increase in the CUs (from 40 to 80) and a 15% (1.9 to 2.2) clock boost is going to net just 40%-50% in performance increase...? And that is ignoring the rumored 15% increase in IPC...

And before someone does it... You definitely can't compare TFLOPS from the RTX 3000 series to Navi. You can't even compare them to Turing (2000 series). In order to compare it to Turing, you basically have to multiply the TFLOPS by 0.7, and you get the Turing equivalent TFLOPS. So a 30 TFLOP Ampere GPU is about a 21 TFLOP Turing GPU.

 

[BattleScar]

It's not the same node.

And it will be roughly 13 tflops which will only equal 2080 super. 2080 ti is around 17 tflops at in game clocks.

It is the same node. Its slightly tweaked for performance, but it is fundamentally the same node. Which is why TSMC stopped calling it N7+ and now just call it N7. The design rules, tools and so on are fundamentally exactly the same. So again. Its the same node.

And here we go with the TeraFLOP game.
The 5700XT with 40CU has 2560 FP32 ALUs. Most AIB's can run it pretty steady at 1905MHz on average. So when you crunch the numbers you get 9.75 TeraFLOPS
The 2080TI with 68SMs has 4352 FP32 ALUs (CUDA cores) which typically runs at 1850MHz on average. Crunch the numbers and you get 16.1 TeraFLOPS

• If we do the math in terms of vector ALUs: (4352/2560) * 100 = 170%. So the 2080Ti has 70% more shaders than the 5700XT
• If we do the math using FLOPS: (16.1/9.75) * 100 = 165%. So the 2080Ti completes 65% more floating point arithmetic calculations per second than the 5700XT
• If we use your, more generous FLOP numbers for the 2080TI: 17/9.75 * 100 = 174%. Making the 2080Ti capable of performing a whopping 75% more FLOPs than the 5700XT

The challenge I set for you, is to find me any benchmark that has the 2080Ti as being 70% faster, on average than a 5700XT.

 

[JohnnyFootball]

This will put them between a 3070 and 3080 on the high end. Probably closer to a 3070 in practical performance...assuming a 3070 is a bit better vs a 2080ti. I imagine they wiIl aggressively price is below the 3070 though. It will be a solid choice for mid-high range GPU's but NVIDIA will continue to dominate the high end.

I wouldn't take the 'AMD said they are going to do to NVIDIA what they did to intel' bit seriously. AMD has a history of overhyping.

Im doubtful this will happen. They will likely be more expensive than the 5700 and 5700XT card

 

[Ascend]

Im doubtful this will happen. They will likely be more expensive than the 5700 and 5700XT card

Agreed. If it is more power efficient and performs the same while having 16GB instead of 10GB, they could technically charge more. But I still think they will undercut nVidia by $50, which is not much. I expect the lowest they would go is $100 below the RTX 3080 price, if that... I think $50 below is more realistic.

 

[Allandor]

Not legit. 2.2 Ghz @ 238W. No way with 80 CUs
And 2.5 GHz @170W ...

Really those numbers are not legit. 2.5 Ghz is also much to high.

 

[Ascend]

Not legit. 2.2 Ghz @ 238W. No way with 80 CUs
And 2.5 GHz @170W ...

Really those numbers are not legit. 2.5 Ghz is also much to high.

That's the GPU itself, not counting the RAM (or other components like fans, RGB etc). It does seem a bit low though. But we'll see.

As for the clocks, I don't consider 2.5 GHz unrealistic for a 40CU on PC, if a console with 40CU (with 4 disabled) can reach 2.2 GHz.

 

[McHuj]

Looks like RDNA2's main focus was clock and power improvements. This is consistent with what the consoles are showing: a reasonable power consumption with a perf of 10-12 TF.

I think Navi 21 would be over 2X raw perf over Navi 10, I don't think it will scale perfectly, but if this is likely so let's assumer only a 1.8x increase in real world performance. That would likely put this chip right around a 3080. Not bad, AMD.

The power consumption is most impressive and I think really shows the comprises Nvidia made with Ampere by going with Samsungs 8nm process.

 

[CrustyBritches]

Navi 22 40CU is your XSX equivalent, while the 36CU version if it exist would be the same for PS5. IIRC the 5700/5700XT launched at $349 and $399, respectively. Could see the same for these GPUs.

 

[ethomaz]

These overclocks are not sustainable...

I’m doing right?

 

[Ascend]

Navi 22 40CU is your XSX equivalent, while the 36CU version if it exist would be the same for PS5. IIRC the 5700/5700XT launched at $349 and $399, respectively. Could see the same for these GPUs.

The XSX has 52 CUs, from a 56 CU chip. So there is currently nothing in the Navi line-up that reflects the XSX GPU. The PS5 would be the 40 CU equivalent, with 4 CUs disabled.

 

[CrustyBritches]

The XSX has 52 CUs, from a 56 CU chip. So there is currently nothing in the Navi line-up that reflects the XSX GPU. The PS5 would be the 40 CU equivalent, with 4 CUs disabled.

Performance-wise. 1.8GHz is slow comparatively.

 

[Senua]

It's not the same node.

And it will be roughly 13 tflops which will only equal 2080 super. 2080 ti is around 17 tflops at in game clocks.

Fucking tlop numbers have ruined all GPU discussion lol

 

[NoviDon]

Sucks we gotta wait another month to see what this bad boy can really do. I'm most interested in the ray tracing solutions they've implemented .

 

[StateofMajora]

With those clocks, And how disappointing ampere is (TF rating completely misleading) this could actually dethrone Nvidia.

its no Vega.

with extra clock speed every part of the pipeline gets faster.

 

[StateofMajora]

Fucking tlop numbers have ruined all GPU discussion lol

Yeah. Relative to bit wars, we’ve hit the 128bit mark.

was the Xbox weaker than ps2 because 32 bit? Lol no.

Nvidia has finished muddying the waters with ampere.

 

[McHuj]

One more thing that popped into my head is that Navi 21 would be an interesting data point for the feasibility of a PS5 Pro. It's already roughly 2x the TF of PS5 and at a decent power consumption.

A node shrink and continued power consumption optimizations could cut this number down in half and make it suitable for a console. I didn't think a Pro was feasible simply due to the poor power scaling of nodes, but it looks like AMD is making great strides in the architecture optimizations to over come that.

 

[Ascend]

Performance-wise. 1.8GHz is slow comparatively.

That's very true. Maybe 40CUs at 2.5GHz is indeed not that far off from the XSX.

One more thing that popped into my head is that Navi 21 would be an interesting data point for the feasibility of a PS5 Pro. It's already roughly 2x the TF of PS5 and at a decent power consumption.

A node shrink and continued power consumption optimizations could cut this number down in half and make it suitable for a console. I didn't think a Pro was feasible simply due to the poor power scaling of nodes, but it looks like AMD is making great strides in the architecture optimizations to over come that.

I'm not sure we will get Pro versions this generation. The consoles aren't even out and people already want pro versions lol.

 

[McHuj]

I'm not sure we will get Pro versions this generation. The consoles aren't even out and people already want pro versions lol.

I don't know either, but to get a Pro console in 3-4 years, the decision has to be made now. It takes that long to develop that kind of SOC and surrounding product.

 

[wordslaughter]

I don't trust AMD GPU leaks. They are usually wrong and only in one direction.

But wouldn't it be nice if they were true and Nvidia vs AMD was actually a choice you had to think about for once.

If getting a 20GB 3080 wasn't a no brainer and AMD actually brings worthy competition to the table, I would buy AMD GPU for no other reason than that. My last AMD GPU was a 7970 GHz Edition. ( or a 280X for my wife ).

 

[Tchu-Espresso]

Could a Custom Navi 31 (80CUs with identical configuration as Navi 21; possibly RDNA3) be the GPU of a hypothetical PS5 PRO coming out in Nov 2022 ? Just wondering if this would be feasible in terms of cost and heating dissipation in two years from now.

No, for the simple fact that Sony won’t do anything of the sort for at least 4 years.

 

[wordslaughter]

With those clocks, And how disappointing ampere is (TF rating completely misleading)

Ampere TFLOPS are not misleading in any way. You just don't understand what a TFLOP is or how it's calculated ( or it's real world significance ).

An Ampere TFLOP is calculated in the EXACT same way that an RDNA, or Playstation or Xbox TFLOP is calculated.

The only problem here is your understanding.

 

[McHuj]

If getting a 20GB 3080 wasn't a no brainer and AMD actually brings worthy competition to the table, I would buy AMD for no other reason than that.

I think for that to happen AMD needs to show a DLSS and RTX competitive equivalent. I hope to be surprised, but I don't expect to that to really show up until RDNA3 type GPU's.

For now, I'll be sticking with a 3080 (if I can pre-order the fucking thing),

 

[wordslaughter]

I think for that to happen AMD needs to show a DLSS and RTX competitive equivalent. I hope to be surprised, but I don't expect to that to really show up until RDNA3 type GPU's.

For now, I'll be sticking with a 3080 (if I can pre-order the fucking thing),

For sure. If it's equal to a 3080 in rasterization but inferior in Raytracing and DLSS then the price has to be a good bit lower.

I'd really like to see AMD have a response to DLSS because it really IS a game changer. If you have 2 equal GPUs ( but one with DLSS and one without ) it allows one GPU to perform at 4K like the other performs at 1440p. Normally you'd have to buy a completely different tier of GPU to get that kind of performance bump.

AMD has its work cut out for it.

But what if Big Navi has worse raytracing, no DLSS answer but higher rasterization performance and at the same price. Then my answer is .... Hmmmmmmmmm.....

 

[Psajdak]

Can someone translate those numbers into English?

 

[wordslaughter]

Can someone translate those numbers into English?

Basically, if true, Big Navi will clock significantly higher than ANY other GPU. Even on PC with liquid cooling you're not going to get higher than maybe 2.1GHz. So if this is true, Big Navi AMD will have a GPU that clocks at 2.5GHz and that's presumably BEFORE people get to overclocking it!

This really would be significant. Imagine if AMD or Intel released a new CPU that clocked at 6GHz ( out of the box on air ). This would be like that.

It would result is a significant bump in performance, and clock vs performance is almost linear. ( And when it comes to calculating TFLOPS it absolutely is linear. )

 

[duhmetree]

Basically, if true, Big Navi will clock significantly higher than ANY other GPU. Even on PC with liquid cooling you're not going to get higher than maybe 2.1GHz. So if this is true, Big Navi will have a GPU that clocks at 2.5GHz and that presumably BEFORE people get to overclocking it!

This really would be significant. Imagine if AMD or Intel released a new CPU that clocked at 6GHz ( out of the box on air ). This would be like that.

It would result is a significant bump in performance, and clock vs performance is almost linear. ( And when it comes to calculating TFLOPS it absolutely is linear. )

I thought 21 was Big Navi? 2.2ghz at 228w? am I reading that right?

 

[wordslaughter]

I thought 21 was Big Navi? 2.2ghz at 228w? am I reading that right?

21 is Big Navi yes. My bad. FIxed my post. I just meant to say that it's significant that ANY GPU could clock so high. Even 2.2GHz is higher than we see on PC GPUs. And we can't really trust wattage ratings. They can be calculated in a lot of different ways. Same goes for CPU wattages. We won't really know how much power these GPUs use until they are released and people actually test them while running games.

 

[BluRayHiDef]

Why is there such a huge gap in CU count between Navi 21 and Navi 22 (80 CUs vs 40 CUs)? A 60CU model could fit between them. Also, does anyone know how a terfalop of RDNA2 compares to one from Ampere? How does 22.5 teraflops of RDNA2 compare to 36 teraflops of Ampere?

 

[Andodalf]

I think for that to happen AMD needs to show a DLSS and RTX competitive equivalent. I hope to be surprised, but I don't expect to that to really show up until RDNA3 type GPU's.

For now, I'll be sticking with a 3080 (if I can pre-order the fucking thing),

RDNA 2 will have Hardware RT, and support for Direct ML

 

[semicool]

That's very true. Maybe 40CUs at 2.5GHz is indeed not that far off from the XSX.


I'm not sure we will get Pro versions this generation. The consoles aren't even out and people already want pro versions lol.

The only PRO talk that I've seen is the Playstation pro. Sony fans aren't truthfully happy with the variable performance with the top of the variation at the lower 10 TFs, that must be why they are excited at the idea IMO.

 

[StateofMajora]

Ampere TFLOPS are not misleading in any way. You just don't understand what a TFLOP is or how it's calculated ( or it's real world significance ).

An Ampere TFLOP is calculated in the EXACT same way that an RDNA, or Playstation or Xbox TFLOP is calculated.

The only problem here is your understanding.

How bout that Vega 64 vs. 1080ti eh?

I understand perfectly fine, however tflops don’t tell the whole story. Continue to shill.

 

[Bernd Lauert]

Big if true, unironically. 2x the CUs, 15% more clock frequency and likely 10ish% more IPC performance compared to the 5700XT... if true, it should be quite a bit faster than the 3080.

Price it at $599 or less and it will be a very good deal.

 

[SlimySnake]

It is the same node. Its slightly tweaked for performance, but it is fundamentally the same node. Which is why TSMC stopped calling it N7+ and now just call it N7. The design rules, tools and so on are fundamentally exactly the same. So again. Its the same node.

And here we go with the TeraFLOP game.
The 5700XT with 40CU has 2560 FP32 ALUs. Most AIB's can run it pretty steady at 1905MHz on average. So when you crunch the numbers you get 9.75 TeraFLOPS
The 2080TI with 68SMs has 4352 FP32 ALUs (CUDA cores) which typically runs at 1850MHz on average. Crunch the numbers and you get 16.1 TeraFLOPS

• If we do the math in terms of vector ALUs: (4352/2560) * 100 = 170%. So the 2080Ti has 70% more shaders than the 5700XT
• If we do the math using FLOPS: (16.1/9.75) * 100 = 165%. So the 2080Ti completes 65% more floating point arithmetic calculations per second than the 5700XT
• If we use your, more generous FLOP numbers for the 2080TI: 17/9.75 * 100 = 174%. Making the 2080Ti capable of performing a whopping 75% more FLOPs than the 5700XT

The challenge I set for you, is to find me any benchmark that has the 2080Ti as being 70% faster, on average than a 5700XT.

But I never said that it is 70% faster. I said it has 17 tflops based on the in game clocks ive seen in benchmarks though I concede that they might have been overclocking it. from the flight sim benchmarks I just searched on youtube hovers in the 1950s though ive seen some that are in the 1890s. it seems to be either silicon lottery or cooling solutions that allow it to high higher clocks.

regardless, tflops are tflops. higher clocks will give you more performance, and the same logic should apply when comparing a 12.8 tflops rdna gpu to a 12-13 tflops to a 16 tflops rtx 2080 ti. the point is that a 12.8 tflops rdna gpu is not gonna match a 2080ti unless AMD has managed to improve IPC gains by another 25%.

as you can see below, the rtx 2080 ti is on average 56% faster than the 5700 xt. you are not going to make that up by just increasing tflops by 25%-30%. there is still 25-30% performance left.

What's interesting is that we do indeed see that even nvidia was struggling to get 1:1 performance scaling when increasing shader processors. 56% is not 65% by your tflops calculations and yet here we are. so while AMD hit that 64 cu ceiling with the vega 56, it seems that Nvidia has gotten there with the 2080 ti and now the entire 3080 which has 8700 shader processors, 3x that of 2080 and is only offering 2x more performance.

it will be interesting to see if the 80 cu rdna 2.0 can offer double the performance of the 40 cu 5700xt. as for the 40 cu 2.5 ghz navi 22, it will need a whole lot more than 2.5 ghz to come even close to the 2080 ti. IPC gains are a must if that were to happen.