AMD "Navi" GPU by Q3-2018: Report

[f22a4bandit]

I know, it puzzles me why all Ryzen CPUs are being heavily discounted at retail - their latest and greatest ever - with no successor in sight? Doesn't make a lot of sense. Please enlighten me as to why Ryzen is losing popularity so quickly. I offered to build a Ryzen system for my nephew for his birthday this November. A week later he tells me he wants an Intel system, apparently because his friends laughed at him when he told them he was getting a Ryzen rig. Has AMD lost their mojo this quickly? What gives? What do teenage gamers know about it that we don't?

Zen+ is supposedly right around the corner for early 2018. That might explain some of the discounts we've seen lately. Also likely to better position themselves against Coffee Lake.

It sounds like your nephew's friends simply go for Intel because they've been touted as the best for years without any competition from AMD. I've spoken with friends who only build Intel machines because they're the best for gaming which is pretty much the same reasoning.

 

[Vayra86]

Here we go again, a full year of hype and much ado about nothing, only to end up with a product released in March/April 2019 that performs about 25% slower than its competitor.

When nothing has a set date, there is no set date, just an intent to release something at some point. But we knew that two years ago.

Fact remains, RTG has been underperforming like nobody's business so far, GPU did better when it still had the AMD tag, and it is a fact that AMD has separated RTG from the core business because of risk concerns for their successful Zen product. Read between the lines, that'll be all thanks

IF AMD is really pushing Navi as the next best thing, this can really go only way direction, and that is pushing Nvidia harder on Volta, resulting in Volta single-die GPUs battling it out with MCM solutions from AMD. So basically where AMD needs two, three or four dies, Nvidia bakes one to match its performance. I'm sure that'll help the margins alot. AMD literally has one year now to refine GCN to a point that it remains viable WITHOUT resorting to MCM. If they do not, well, say hi to stagnation and price inflated GPUs again.

Another vital issue arises, and that is tremendous amounts of work for driver teams and engineering too, because MCM is never going to work the same as blunt ol' Crossfire, it'll talk over Infinity Fabric, and it will be utilizing VRAM differently. So by then AMD has THREE radically different architectures in the marketplace to support one stack of products. Again, not the best example of efficiency.

It'll be another year of agony.

 

[eidairaman1]

Not soon enough for this process, amd will be behind yet again.

 

[efikkan]

So "AMD reportedly…" actually means Anthony Garreffa over at Tweaktown speculating, the source has nothing to do with AMD.

Have all of you forgotten how the news outlets spread the rumor about Vega launching last October?

AMD have said they hope to tape out Navi by end of 2017, usually products arrive one year after that. It took almost 13 months from tape out until Vega FE launched, about 14 months for the consumer product. Vega was a fairly minimal design improvement over Fiji, and was produced on a mature node. Navi will have more radical changes and be produced on a brand new node. GloFo 7nm is not planned to enter volume production until H2 2018, and takes another ~4 months from a wafer to a finished product.

Actually Vega is quite power efficient. If you look at GamersNexus, they did a video where they compared Vega to the Fury X and Clock for Clock Vega is significantly more power efficient. In fact I managed to get my hands on a Vega 56 to test this. I flashed the Vega 64 bios onto the card and lowered the voltage to 950mv and left everything else alone.

Operating at such low voltages will not give you a stable product.

I know, it puzzles me why all Ryzen CPUs are being heavily discounted at retail - their latest and greatest ever - with no successor in sight? Doesn't make a lot of sense. Please enlighten me as to why Ryzen is losing popularity so quickly.

Simply because Coffee Lake is a better choice.

Most likely AMD will have 12nm GloFo shrink versions for Vega. It's kind of official.
https://www.globalfoundries.com/new...-technology-for-high-performance-applications

"12nm" GloFo is just another refresh of "14nm", similar to TSMC "12nm" used by Nvidia for Volta.

 

[Vya Domus]

Oh yes, the anti-AMD shill offered to build his imaginary nephew an AMD system. That's not even the funny part, the funny part is that you are making YOUR rig building decisions based on the bulling of school children.

"I know, it puzzles me why all Ryzen CPUs are being heavily discounted at retail "

So many people fall for this pathetic attempt of troll bait. Good to see some don't.

On topic : if they find way to distribute the workload efficiently and most importantly make it so that to an API the GPU still appears as a single pool of resource , it will be a home run. With or without high clocks or GloFo's questionable nodes.

 

[evernessince]

So "AMD reportedly…" actually means Anthony Garreffa over at Tweaktown speculating, the source has nothing to do with AMD.

Have all of you forgotten how the news outlets spread the rumor about Vega launching last October?

AMD have said they hope to tape out Navi by end of 2017, usually products arrive one year after that. It took almost 13 months from tape out until Vega FE launched, about 14 months for the consumer product. Vega was a fairly minimal design improvement over Fiji, and was produced on a mature node. Navi will have more radical changes and be produced on a brand new node. GloFo 7nm is not planned to enter volume production until H2 2018, and takes another ~4 months from a wafer to a finished product.


Operating at such low voltages will not give you a stable product.


Simply because Coffee Lake is a better choice.


"12nm" GloFo is just another refresh of "14nm", similar to TSMC "12nm" used by Nvidia for Volta.

"Operating at such low voltages will not give you a stable product."

And you would know because you've had it and tried? I stress tested for 8 hours and then gamed on it, it is stable contrary to your uneducated opinion.

 

[evernessince]

No it's not, different CPU architectures often have different CU counts which is what leads to different IPC...

First, we are not talking about CPUs. Second, CPUs don't have CUs, as in compute units that GPUs have. The only CU a CPU has is a control unit, and that's singular not plural. You obvious have no idea about CPU architecture if you think that CPUs have a bunch of CUs like GPUs. Thrid, CPU IPC is much more complicated that you put it. In fact, most modern CPUs have eeked out extra performance through better branch prediction and cache improvements, to the contrary of your imaginary clusters of CUs on CPUs.

 

[Dave65]

Love AMD but all I hear is BLAH BLAH BLAH, believe it when I see it!

 

[GorbazTheDragon]

First, we are not talking about CPUs. Second, CPUs don't have CUs, as in compute units that GPUs have. The only CU a CPU has is a control unit, and that's singular not plural. You obvious have no idea about CPU architecture if you think that CPUs have a bunch of CUs like GPUs. Thrid, CPU IPC is much more complicated that you put it. In fact, most modern CPUs have eeked out extra performance through better branch prediction and cache improvements, to the contrary of your imaginary clusters of CUs on CPUs.

CUs being a generic term for the different processors in a CPU core such as FPUs and ALUs which still have a big effect on architecture performance...

 

[Xzibit]

So "AMD reportedly…" actually means Anthony Garreffa over at Tweaktown speculating, the source has nothing to do with AMD.

Have all of you forgotten how the news outlets spread the rumor about Vega launching last October?

Well techsites tend to regurgitate what they think is news worthy even when it doesn't make sense just to get clicks

Source articles said.

Once again I have an exclusive story that AMD will have Navi ready to go sometime in July-August 2018, with a Navi-based professional card being launched at SIGGRAPH 2018. We're still waiting for AMD to launch Radeon Pro SSG, something they unveiled during SIGGRAPH 2017 that hasn't yet materialized. In the meantime, Radeon Technologies Group boss Raja Koduri has taken a sabbatical from the company until early-2018.

Not a story nor did they include one.

Navi will be made on the 7nm process, but other than that we don't know much.

Wait what happen to your exclusive?

The entire piece is speculation and TPU bite on it for what ever reason. Then we get this drawn out debate on something that's just speculative.

Sites win. Made you click and gain traffic...

 

[evernessince]

CUs being a generic term for the different processors in a CPU core such as FPUs and ALUs which still have a big effect on architecture performance...

Please point to a reference besides you that used "CU" as catch all for FPU and ALU. I took a computer architecture class and digital logic class in college and never before did I see anything in our studies of past or present architecture that uses the same terminology as you. CU is either a control unit, of which there is only one, or compute unit, of which refers specifically to graphics cards.

 

[GorbazTheDragon]

Please point to a reference besides you that used "CU" as catch all for FPU and ALU. I took a computer architecture class and digital logic class in college and never before did I see anything in our studies of past or present architecture that uses the same terminology as you. CU is either a control unit, of which there is only one, or compute unit, of which refers specifically to graphics cards.

CU in graphics cards typically refers to a SP FPU no?

Sorry for being so bad with the terms, but I think my point still stands. Radeon's chips, in particular Vega, is doing less work with more resources...

 

[Vya Domus]

A CU (talking about compute unit not control unit) is an abstract name which was given to stream processors to signify that they are capable of general purpose computing. In actual fact it doesn't really bear any other meaning or hardware requirement associated with it outside of that.

A stream processor was also also an abstract name given to a class of processors that rely heavily on data level parallelism.

Stop getting stuck in meaningless semantics.

 

[ratirt]

A CU (talking about compute unit not control unit) is an abstract name which was given to stream processors to signify that they are capable of general purpose computing. In actual fact it doesn't really bear any other meaning or hardware requirement associated with it outside of that.

A stream processor was also also an abstract name given to a class of processors that rely heavily on data level parallelism.

Stop getting stuck in meaningless semantics.

I think that's what he meant. CU aka compute units is reserved for GPUs. CU for CPUs is control unit and there's only one in CPU.
You shouldn't be like that. It's a site where people talk. If somebody struggles it's good to straighten him up with information.

 

[Slizzo]

hello terrible performance scaling
has amd not learned that nobody is going to optimize for there stupid module approach

You realize that NVIDIA will be moving to an MCM approach after Volta as well, right? Jensen himself said that was the next move.

 

[Boosnie]

You realize that NVIDIA will be moving to an MCM approach after Volta as well, right? Jensen himself said that was the next move.

On top of that, Ryzen scales near 100%, and GPUS are modular in nature, no way this could not scale as much as well.
Even if the bridge will let software access to indipendent modules, DX12 are here for indipendent multiple processing.

 

[Assimilator]

You mean 2018, 10nm isn't going into (mass?) production this year & I mean full launch not paper announcements.

Intel has made it quite clear (e.g.) that they intend to start producing and shipping 10nm chips before the end of 2017. This may not be mass production, but considering the accelerated launch of Skylake-X/Kaby Lake-X and Coffee Lake in response to Zen, they are obviously taking AMD's resurgence seriously. Yes, it's been a long haul to get to 10nm, but Intel generally doesn't make statements that they can't back up - especially considering how important 10nm is going to be for them. So if they say they'll have 10nm chips by the end of this year, I believe them.

 

[InVasMani]

MCM compared to SLI/CF it'll be very different because it'll be a standardized feature of the hardware and not a luxury after thought that only a select few opt to take advantage of utilizing. That difference alone will inherently coerce developers to pay closer attention to it than optimizing for a handful of niche configurations.

 

[Boosnie]

Intel has made it quite clear (e.g.) that they intend to start producing and shipping 10nm chips before the end of 2017. This may not be mass production, but considering the accelerated launch of Skylake-X/Kaby Lake-X and Coffee Lake in response to Zen, they are obviously taking AMD's resurgence seriously. Yes, it's been a long haul to get to 10nm, but Intel generally doesn't make statements that they can't back up - especially considering how important 10nm is going to be for them. So if they say they'll have 10nm chips by the end of this year, I believe them.

I believe that those 2017 chips were intended to be FPGA.

 

[RejZoR]

Intel has made it quite clear (e.g.) that they intend to start producing and shipping 10nm chips before the end of 2017. This may not be mass production, but considering the accelerated launch of Skylake-X/Kaby Lake-X and Coffee Lake in response to Zen, they are obviously taking AMD's resurgence seriously. Yes, it's been a long haul to get to 10nm, but Intel generally doesn't make statements that they can't back up - especially considering how important 10nm is going to be for them. So if they say they'll have 10nm chips by the end of this year, I believe them.

The thing is, as process matures, so does efficiency. Even on same node. Just look at Intel's 14nm in the beginning and now. Or how 28nm matured from the beginning of it's life till its very end with GPU's, on both, Fury X and GTX 900 series. Or even how bloody efficient latest gen 22nm Haswell-E CPU's still are even compared to 1st gen of 14nm Broadwell-E.

 

[anubis44]

Lol I love AMD. They the #1 machine that comes into my shop with problems. I love them because they make me money fixing them. HP, Gateway, Acer, Dell, Asus, Etc. I used to own AMD. I had 5 of them long time ago tweaking having funs fixing issues all the time. Then I noticed when I switch all to Intel machines My problems went away. Also my power bills dropped by 100 dollars a month. I have all Intel now some older than 15 years so Ya I perfer stable and allways works than having issues and allways having problems.

LOL, your fake story is even worse than the one about the kid being picked on because he was getting an AMD-powered gaming system. Don't you people ever learn? We're just not that stupid.

I have been building computers for people since the early 1990s. My first IBM-compatible was actually an AMD 386/40MHz, which was faster than the Intel 386/33MHz, and $60 cheaper. Guest which one I bought? With few variations, it's been the same DAMNED STORY ever since. AMD gives you more for your money. Their CPUs, with only a few varations, are just as fast for all intents and purposes. When you need benchmark software to show you you're hypothetically getting better performance with an Intel chip because you CAN'T FEEL IT yourself when you actually use the computer, that's telling you something you really need to be listening to.

I also actually game, and I have a Ryzen 7 1700 o'c'd to 3.8GHz and an R9 290 Radeon card, and they're both more than fast enough in every game I throw at them. I'm playing the very latest game, Wolfenstein II: New Colossus with this system, and it's buttery smooth. AMD CPU. AMD GPU. It's UTTER NONSENSE that you need Intel or an nVidia GPU to game really well. They charge more, but that's all. You're not getting anything for that price premium. End of story.

Intel has made it quite clear (e.g.) that they intend to start producing and shipping 10nm chips before the end of 2017. This may not be mass production, but considering the accelerated launch of Skylake-X/Kaby Lake-X and Coffee Lake in response to Zen, they are obviously taking AMD's resurgence seriously. Yes, it's been a long haul to get to 10nm, but Intel generally doesn't make statements that they can't back up - especially considering how important 10nm is going to be for them. So if they say they'll have 10nm chips by the end of this year, I believe them.

Intel has pushed 10nm to 2018. They said something they can't back up. It's official.

 

[InVasMani]

I had better reliability and stability with AMD than Intel of the course of the last 15 years or so and owned a mixture of both. If it weren't for Intel's manufacturing advantage I probably wouldn't have owned any of their CPU's ironically.

 

[Slizzo]

Intel has pushed 10nm to 2018. They said something they can't back up. It's official.

Annnnnnd then Intel turns around and says this:

https://www.techpowerup.com/forums/...e-very-limited-initial-launch-in-2017.238307/

Limited production in 2017.

 

[Deadstaar]

Here's the bit everyone seems to overlook and/ or forget, pertaining to AMD. AMD does not compete. At least not for the enthusiast market. Their main objective is to offer a cost effective solution. Nvidia just makes their job harder by introducing something like the GTX 1060 which happens to rival, if not supersede, most of AMDs flagship cards(RX480/ 580) and at a lower price point.

Notice, Radeon products have not won a generation based on a shear power and performance level since the 1000XT and 2000XT series of cards, and going back before that. This is because Radeon is no longer ATI, but AMD and AMD has a completely different marketing method. ATI was of a competitive enthusiast card back in the day, but with 3DFX it was a tough battle. Nvidia, not terribly popular at the time, acquired 3DFX and their tech and ascended to glory. Still, ATI competed.

As far as CPU goes AMD has Never won on a power/ performance level, but has won CPU generations based on a price-to-performance level and only just recently done so once again with its Ryzen line of CPUs, since the days of the Phenom. Thread Ripper, like the Phenom IIs, doesn't quite rival or beat out Intel's upper end of CPUs(in this case the i9s), but gets us close at a better cost. This is why AMD is allowed to remain. They are not a threat to either Intel or Nvidia, but a pain in the a$$ to either company's midrange market, which is the majority. Once Thread Ripper is over and done with people will be trying to buy them all up and run them for the next 8 years, just like the Phenom X6 1100T, after people will pour over to the next more powerful of the Ryzen CPUs until those are gone, and everyone will speak of the Ryzen days, just as people do the Phenom days. AMD will then fade into unpopularity and will then again produce another memorable moment in another near decade. They've done this ever since they've been around. AMD Athlon - Thunderbird, a popular CPU at its time for cost-to -performance, as well as their unlocked nature(circa 2000), AMD wouldn't repeat this popularity again until the Phenoms roughly 8 years later...hmmm

The AMD's GPU line now echos their CPU market. AMD produces a kick-a$$ piece of silicon that nearly rivals what Intel and Nvidia are doing in their respective markets, but at a lower price point.

 

[eidairaman1]

580 was never flagship