News AMD Says its Upcoming RDNA 2 and Navi 2x Will Boost Performance per Watt by 50%

[Admin]

AMD plans to launch its RDNA 2 architecture and Navi 2x GPUs by the end of 2020, bringing ray tracing support and up to 50% more performance per Watt to its consumer graphics products.

AMD Says its Upcoming RDNA 2 and Navi 2x Will Boost Performance per Watt by 50% : Read more

 

[Giroro]

"by the end of this year."

He was stumbling on his words a bit, he also said they would be announced late this year.

 

[willgart]

my problem: if they reduces the watt by 50%... there is no performance improvement at all at the end.
so the perf per watt is fine for data centers.
but for gamers... we dont care.
so give me a 300watt GPU (33% more than the 5700XT) + 50%/watt = 19.5TFlops

 

[JarredWaltonGPU]

my problem: if they reduces the watt by 50%... there is no performance improvement at all at the end.
so the perf per watt is fine for data centers.
but for gamers... we dont care.
so give me a 300watt GPU (33% more than the 5700XT) + 50%/watt = 19.5TFlops

I suspect it will be a 250W level that’s 20-30% faster than the 5700 XT. But a lot of that might be in specific workloads. I’m certainly curious to see actual specs though.

 

[escksu]

Performance per watt is very ambiguous...... But its extremely unlikely we will see a GPU that has same performance as 5700XT but consumes 1/3 less power.....5700XT is 225W is 1/3 less is 150W. Neither will we see a new 225W GPU thats 50% faster.

I think its probably somewhere in between. AMD and Nvidia loves to make GPUs that pretty comparable so as not to upset the market.

I am inclined to believe we will see 175W-200W GPU. Then we get around 20% higher in performance. So I think we will see 2080 Super or maybe 2080ti performance for 175-200W GPU. Maybe it will be called 6700xt.

 

[Deicidium369]

AMD always shooting for last year's Nvidia product and still coming up short. The full-screen ray tracing on consoles will be CLOUD-BASED... I hope your internet connection is fast, low latency and without caps.
I hope they come out with something that can actually compete with Nvidia at the high end - but that is highly unlikely. I haven't seen anything to rival my dual 2080Ti cards - and the next upgrade will be dual Ampere "3080Ti".

Maybe AMD could start making outdoor furniture to make some money, it seems like they are not selling much of anything - yeah those exascale supercomputers are great - but they are a 1 shot and done thing - and the #s they sell is a rounding error at Intel.

 

[Deicidium369]

Performance per watt is very ambiguous...... But its extremely unlikely we will see a GPU that has same performance as 5700XT but consumes 1/3 less power.....5700XT is 225W is 1/3 less is 150W. Neither will we see a new 225W GPU thats 50% faster.

I think its probably somewhere in between. AMD and Nvidia loves to make GPUs that pretty comparable so as not to upset the market.

I am inclined to believe we will see 175W-200W GPU. Then we get around 20% higher in performance. So I think we will see 2080 Super or maybe 2080ti performance for 175-200W GPU. Maybe it will be called 6700xt.

Lets not get crazy with the optimism - 2080 maybe - super and Ti probably not. I really don't care whether a card pulls 10W or 500W - if it is the top performer at the high end - then it doesn't matter. IF you are that concerned about your electric bill then the high end is not for you.

6700XT is likely - 2080 levels at around the same 200-250W envelope - but that great and all in a world without Nvidia releasing a major update (architecturally and process) sometimes this year as well. AMD needs to invest in someone with experience in drivers to shore up the terrible video drivers. Greatest hardware in the universe with a total crap software environment is nothing.

 

[Deicidium369]

my problem: if they reduces the watt by 50%... there is no performance improvement at all at the end.
so the perf per watt is fine for data centers.
but for gamers... we dont care.
so give me a 300watt GPU (33% more than the 5700XT) + 50%/watt = 19.5TFlops

And that "19.5TFLOPs" without good drivers might as well not even exist - both sides have to be upto the task - best HW in the world is meaningless without SW that can unlock that performance.

 

[King_V]

my problem: if they reduces the watt by 50%... there is no performance improvement at all at the end.

That's not how the math works.

50% more performance per watt would mean either:
- Same wattage, but 150% (ie: 3/2) of the performance of an existing Navi
OR
- Same performance, but 66% (ie: 2/3) of the power draw of an existing Navi. Basically, getting 150% of your "performance units" (whatever they may be) per watt of consumption.

As to the details, well, obviously we can't know now.

 

[JarredWaltonGPU]

Lets not get crazy with the optimism - 2080 maybe - super and Ti probably not. I really don't care whether a card pulls 10W or 500W - if it is the top performer at the high end - then it doesn't matter. IF you are that concerned about your electric bill then the high end is not for you.

6700XT is likely - 2080 levels at around the same 200-250W envelope - but that great and all in a world without Nvidia releasing a major update (architecturally and process) sometimes this year as well. AMD needs to invest in someone with experience in drivers to shore up the terrible video drivers. Greatest hardware in the universe with a total crap software environment is nothing.

RTX 2080 Super seems highly plausible, considering the 2080 Super is only 15-25 percent faster than RX 5700 XT (24% faster at 4K ultra to be precise, and 14% faster at 1080p ultra). I'm still curious how AMD expects to do 50% better perf per watt than Navi 1x with Navi 2x unless there are actually significant architectural updates. 25% better performance by adding more cores? That's totally feasible. But doing 25% more performance than RX 5700 XT and only using 188W is surprising with the same 7nm node. I hope AMD is serious about this result, but I'm very skeptical of it getting there in general and not just in select scenarios.

 

[Gurg]

AMD always shooting for last year's Nvidia product and still coming up short.

AMD still can't match the performance of the three year old 1080ti. Nothing in the presentation said anything about gaming performance, just performance per watt. efficiency. How long will it take for AMD to work the bugs out of the software to seamlessly run two of its slow GPUs stuck together?

 

[JarredWaltonGPU]

AMD still can't match the performance of the three year old 1080ti. Nothing in the presentation said anything about gaming performance, just performance per watt. efficiency.

Technically, by my numbers, the RX 5700 XT is basically tied with the GTX 1080 Ti at 1080p medium and ultra, and only 3-6% behind at 1440p/4K ultra. And it uses less power. I'd say AMD can at least match the GTX 1080 Ti now -- and the Radeon VII does it as well. But yeah, that's a card from three years back. I was really hoping RDNA 2 would be out by summer rather than fall, but it doesn't sound like that's happening.

 

[drivinfast247]

AMD always shooting for last year's Nvidia product and still coming up short. The full-screen ray tracing on consoles will be CLOUD-BASED... I hope your internet connection is fast, low latency and without caps.
I hope they come out with something that can actually compete with Nvidia at the high end - but that is highly unlikely. I haven't seen anything to rival my dual 2080Ti cards - and the next upgrade will be dual Ampere "3080Ti".

Maybe AMD could start making outdoor furniture to make some money, it seems like they are not selling much of anything - yeah those exascale supercomputers are great - but they are a 1 shot and done thing - and the #s they sell is a rounding error at Intel.

Of only AMD had another product for the PC and data center markets.

 

[derekullo]

Nvidia is saying 50% more performance at half (50%) of the power consumption.

If we get a hypothetical 100 fps at 100 watts with turing that comes out to 1 fps per watt. (For easy math)

With Ampere they are saying 150 fps at 50 watts or 3 fps per watt.

https://www.tomshardware.com/news/n...ter-than-turing-at-half-the-power-consumption

 

[cryoburner]

Nvidia is saying 50% more performance at half (50%) of the power consumption.

According to that link, "Yuanta Securities Investment Consulting Co" was saying that, not Nvidia. Also, the way it was described in that Taipei Times article does not make it clear whether they actually mean 50% more performance at 50% less power. In fact, that seems pretty unlikely, as it would mean a tripling of efficiency. If that were true, we would be looking at 2080 Ti level performance in a graphics card drawing less than 90 watts.

That should clearly be treated as a rumor, and probably not a very reliable one. We'll undoubtedly see big efficiency gains from Nvidia's next cards, but it doesn't seem likely to be much more than 50% more performance at the same power, at least if we are talking about general gaming performance.

 

[digitalgriffin]

And that "19.5TFLOPs" without good drivers might as well not even exist - both sides have to be upto the task - best HW in the world is meaningless without SW that can unlock that performance.

Early leaks from about 4 months ago on early engineering samples suggest it's 20% faster than a 2080ti. Hardly coming up short. While late to the party, it is still faster. And competition is a good thing.

 

[digitalgriffin]

That's not how the math works.

50% more performance per watt would mean either:
- Same wattage, but 150% (ie: 3/2) of the performance of an existing Navi
OR
- Same performance, but 66% (ie: 2/3) of the power draw of an existing Navi. Basically, getting 150% of your "performance units" (whatever they may be) per watt of consumption.

As to the details, well, obviously we can't know now.

It was my understanding, that there was a corruption issue with data caching on the 5700XT couldn't resolved before release. AMD needed to get something out. Thus the feature was turned off resulting than a lower performance and higher power rating than AMD intended. This problem has been resolved from the rumblings I've been hearing.

 

[digitalgriffin]

When a process tech says X% savings on power, what they mean is "At a given Mhz a single transistor will require (1-x%) * original voltage to flip a transistor."

So lets say it takes 1V to flip a transistor on the old node, and they claim a 30% savings. (1-.3 ) * 1V = 0.7V on the new node.

There's a certain minimum current that goes through each transistor flip state. As speed increases, it gets harder and harder to satisfy that current demand at a given voltage. So you ramp the voltages again to push the current. (Voltage = push. Current = how much at a time.) High current demands also lead to voltage sag, which makes it look like a voltage drop. (The voltage across a short circuit is 0) This is why we have things like load line calibration. As power = V * V / R, you start to get an exponential increase in power as you ramp voltages. Hence why you quickly get into thermal runaway with voltage overclocking.

Also smaller features allow smaller distances between structures, thus you don't have to worry about electron propagation (travel time) as much. This is one of the reasons you can ramp higher clock rates.

So you can have power savings, or increased clock rate, but not both without some trade off.

 

[derekullo]

When a process tech says X% savings on power, what they mean is "At a given Mhz a single transistor will require (1-x%) * original voltage to flip a transistor."

So lets say it takes 1V to flip a transistor on the old node, and they claim a 30% savings. (1-.3 ) * 1V = 0.7V on the new node.

There's a certain minimum current that goes through each transistor flip state. As speed increases, it gets harder and harder to satisfy that current demand at a given voltage. So you ramp the voltages again to push the current. (Voltage = push. Current = how much at a time.) High current demands also lead to voltage sag, which makes it look like a voltage drop. (The voltage across a short circuit is 0) This is why we have things like load line calibration. As power = V * V / R, you start to get an exponential increase in power as you ramp voltages. Hence why you quickly get into thermal runaway with voltage overclocking.

Also smaller features allow smaller distances between structures, thus you don't have to worry about electron propagation (travel time) as much. This is one of the reasons you can ramp higher clock rates.

So you can have power savings, or increased clock rate, but not both without some trade off.

This holds true if the lithography becomes smaller but the architecture stays mostly the same. (See Intel's 14nm++++)

If both the lithography becomes smaller and the architecture becomes more efficient then you can indeed have power savings and increased clock rate.

For instance my current ancient gaming computer has an Intel i7-4930k which is a 6C/12T 3.9 gigahertz max turbo, a 130 watt TDP and a lithography of 22 nano-meters.

The i9-9900k which is an 8C/16T 5 gigahertz max turbo, at a controversial but still comparable since it is also Intel 95 watt TDP and a lithography of 14 nano-meters is better in all regards.

The i9-9900k is indeed more power efficient and has increased clock rate due to not only its 35% smaller lithography but also through architectural improvements.

 

[Gurg]

Technically, by my numbers, the RX 5700 XT is basically tied with the GTX 1080 Ti at 1080p medium and ultra, and only 3-6% behind at 1440p/4K ultra. And it uses less power. I'd say AMD can at least match the GTX 1080 Ti now -- and the Radeon VII does it as well. But yeah, that's a card from three years back. I was really hoping RDNA 2 would be out by summer rather than fall, but it doesn't sound like that's happening.

GPU Performance Hierarchy 2020: Video Cards Ranked

https://www.tomshardware.com/reviews/gpu-hierarchy,4388.html

 

[King_V]

Right, so you're just proving @JarredWaltonGPU 's point. Unless you're trying to argue that numbers like 0.6% or 1.4% are enormous shortcomings? In that case, you're just nitpicking.

 

[JarredWaltonGPU]

Early leaks from about 4 months ago on early engineering samples suggest it's 20% faster than a 2080ti. Hardly coming up short. While late to the party, it is still faster. And competition is a good thing.

Anything "leaked" four months ago claiming to show performance is just making stuff up. AMD has never publicly shown anything that suggests Navi 2x has actually taped out. At this point, it probably has (maybe?) but the launch is slated for the end of the year. Usually first silicon is about 6 months before launch. We'll hopefully hear about actual specs sometime in the coming months, but AMD has stated that it is trying to control leaks of data on consumer products, because the market is so competitive. Basically trying to take a page out of Nvidia's book, which means the REAL specs may not be known until the final month or so before launch.

Which is not to say that Navi 2x won't be fast or competitive, but best-case says 50% faster than Navi 1x (by AMD's own numbers, assuming it's the same power and 50% faster -- unlikely). But we've got a while to wait still before we have anything concrete.

 

[digitalgriffin]

Anything "leaked" four months ago claiming to show performance is just making stuff up. AMD has never publicly shown anything that suggests Navi 2x has actually taped out. At this point, it probably has (maybe?) but the launch is slated for the end of the year. Usually first silicon is about 6 months before launch. We'll hopefully hear about actual specs sometime in the coming months, but AMD has stated that it is trying to control leaks of data on consumer products, because the market is so competitive. Basically trying to take a page out of Nvidia's book, which means the REAL specs may not be known until the final month or so before launch.

Which is not to say that Navi 2x won't be fast or competitive, but best-case says 50% faster than Navi 1x (by AMD's own numbers, assuming it's the same power and 50% faster -- unlikely). But we've got a while to wait still before we have anything concrete.

Well Computex is in June. Given the respins of the original RDNA, I wouldn't put it past an 8 month lead time. (November->June (Computex)) to iron out the kinks (heat & power gating/cache corruption issues). AMD has a history of respins with it's architecture since GCN. Plus they have to spin up for Ryzen 4000 which is far more profitable. A faster than 2080ti might be a halo product that gets them recognition, but not with high availability.

BTW: The early benchmarks were based on a VR benchmark. But you are right I'm taking it with a grain of salt.