Vega Rumors

[razor1]

So what exactly are you hoping for? A dual GPU card based on Vega?

And we will see quite a few houses burn down if such a product ever comes to light.

He wants to see a dual Vega that functions like a single GPU. And when we say the tech for that isn't there, he points to infinity fabric, which has nothing to do mGPU's and the needed architectural changes in GPU's to make that happen.

 

[Dayaks]

And we will see quite a few houses burn down if such a product ever comes to light.

He wants to see a dual Vega that functions like a single GPU. And when we say the tech for that isn't there, he points to infinity fabric, which has nothing to do mGPU's and the needed architectural changes in GPU's to make that happen.

Let's pretend it did exist and was functional. You still have the problem of Titan Volta being faster and cheaper... (high confidence of this anyways). I just wanted to make sure I was reading this all right.

 

[razor1]

Sorry, that is not true.

I am not dreaming about anything, I am TRYING to discuss AMD's infinity fabric and the virtues that could possible bring to the high end Gamer. But it seems, that topic keeps getting trolled, and threadcrapped on by people who would rather talk about Nvidia, while badmouthing AMD.

Understand, it is OK to admire and seek knowledge and discussion about Vega technology. Or to even theorize about AMD's technologies and possible path in an AMD titled thread. (It is a Vega Rumor thread, after all.)

If you are trying to discuss Infinity Fabrics and its virtues, leave mGPU out of it, it has nothing to do with mGPU's or seeing multiple GPU's as one, the only thing is its an interconnect between the GPU's. That is all it is, that job can be done with ANY TYPE OF bi directional interconnect. Doesn't need to be fabric.

OK understand that? That is not minimalising it that is just straight out fact.. What AMD is using fabric for is to reduce its over all footprint for different types of system interconnects and reduce COST. OK do you see where I'm coming from here? COST is a big thing for AMD ATM, their margins have been piss poor, so anything to reduce cost by reducing complexity is great for them. That means if they keep the same prices as BD and its platform, their margins go up, and that is what they are doing. What did I say about money walks and BS walks? yeah its money thing for AMD ATM. And AMD even STATED This about infinity fabric when they launched it!

Secondly, I already have Nvidia stuff. I want more. Hence my hunger for HEDT GPU.

But what is troubling and tiring is the amount of negative and derogatory remarks being fired off by the same people, ad nauseam, while others are trying to have a legit discussion here. It is annoying and has to stop. Glad Kyle stepped in, because I am oldschool and believe People should have to show their driver's license (or CC#), to open an account. (That way cellphone kiddies need not apply)

yeah And I'm 40 years old with over 28 years of programming expereince and I don't sit around and talk about post counts, so what gives with you and 50 and post counts?

Allowing this anti-AMD bashing to continue is just negative experience here at HardOCP. I know some of you get off on the giggles, but your actions are pushing away real readers, with real incomes and real interests. I am not afraid of a face to face, or a meet & greet. But being attacked for pushing an AMD subject, with posts trying to dissuade my buying habits, when ALL I am doing is speaking about technology is getting old. I buy whatever I want, whenever I want. I don't fret it. Not everyone here is a poor college student who dreams about owning SLI Titans. I remember wanting things back then too. The difference now is I can afford what I want ! Subsequently, Volta is insignificant and is 10 months away. because GDDR6 doesn't start production until Q2 2018. Coincidentally, the fact that Nvidia is locked out the console market is a death blow, only to Nvidia. It can only been seen as a boon for AMD, but nice try on your spin..

How is it bashing AMD infinity fabric when AMD infinity fabric doesn't do what you are proposing it does?

Come on, know the tech before you try to say an interconnect can make mGPU look like one GPU.... interconnects aren't for that, shit they don't even do that for CPU's lol. Multple CPU's with infinity fabric function like dual socket CPU without NUMA.

I am nearing 50 years of age and about to drop some serious cash on some full-on gaming rigs. I have good reason to believe AMD is going to come threw on their Infinity fabric and Vega, & offer us high end Gamer's, a TitanRipper like card later this year.

Good for you, but that has nothing to do with the discussion at hand you proposed. Dual Vega has no merit purely from a power conumption point of view. Also again Money walks and BS walks, from this point of view it has no merits either. HBM2 with two large dies, will make a dual Vega card extremely expensive no less than 1k, at 1k it will be going up against Titan Xp, while only functioning as a single GPU solution 25% of the time? With much higher power consumption like double the power consumption of the Titan Xp. Its not a product that will look good in any light, it can't take the performance crown, unless you love running 3dmark all day long.

If Someone is not all for that goodness, then I would wager they are not a Gamer, or are peddling an agenda.

So stop making things personal then, no need to talk about post counts, what peoples preferences are, what you think a person likes or dislikes. Stick with what is being talking about the pertinent facts.

Its been pointed out to you numerous times by people with experience with these things, and by AMD's on Raja Koduri with his push towards mGPU and approach to it, that there will be no such thing as a fully transparent solution for mGPU in the near future. So trying to make Infinity fabric as the end all be all of that, is a HUGE ass mistake man.

 

[JustReason]

Dude AMD told Anandtech they can only improve it like 10 to 15%

Here man this what AMD stated about it again.

http://www.anandtech.com/show/11544/intel-skylake-ep-vs-amd-epyc-7000-cpu-battle-of-the-decade/18



What the hell man you think I make things up about what AMD says?

that is just one test and received a bit of criticism being the data set fit in l3 instead of larger sets that would utilize ram and i/o. But as Gideon said depending on the work load, some times AMD looked good, especially FP workloads where they buried Intel. However I have tried not to put much stock in that review as it was preliminary and likely would require a lot more in depth look as even they felt some finding were at best "ballpark" accurate.

 

[razor1]

that is just one test and received a bit of criticism being the data set fit in l3 instead of larger sets that would utilize ram and i/o. But as Gideon said depending on the work load, some times AMD looked good, especially FP workloads where they buried Intel. However I have tried not to put much stock in that review as it was preliminary and likely would require a lot more in depth look as even they felt some finding were at best "ballpark" accurate.

Its not only going to happen in one test, trust me, with bigger databases, that require localization of data, this problem is persistent and will be worse.

Yes I stated it a while back too, Epyc can keep up with Skylake for the most part, but for anything to do with localization of data, another words Databases, any type, its screwed, there is no way around that. And if you look at any large business infrastructure, they can't use Epyc cause most of their databases run their company from sales, to inventory to even HR. Shit anything on the internet, sales sites, resume sites, anything that collects user information, etc, all these are databases, large and small.

And this problem is all due to Infinity fabric. AMD would have known this problem well ahead of release right? They would have know of it during design too. But the reason they went down this path was because its cost effective for them.

PS only one way to go for Epyc's test and that is down, the Anandtech's tests were done with AMD's supervision. So best cases?

No Gideon stated I pretty much made things up. What he doesn't remember I linked him to that article before? he should remember he was part of that conversation.

now back to the main point how the hell can one say infinity fabric is cure all to mGPU when it can't even keep multiple cpu's without performance pitfalls, where GPU's need more bandwidth and less latency?

Just doesn't make any sense how anyone can connect those two together they are polar opposites. One is dropping performance by increasing latency, and the other needs lower latency and higher performance!

That right there should hit sine wave over the head to understand infinity fabric isn't going to be what GPU's need for his proposed dual vega functioning like a single vega theory.

 

[razor1]

Didn't Raja try to peddle 2 x 480 as being faster than 1080 early on? Yeah he's totally honest and forthright. AMD fans are funny, they realize Vega is a disaster so now they start coming up with a "TitanRipper" fantasy--how long until Navi saves the world? I'ts always the same story with AMD and their fans..."just wait.."

P.S. Since we're making shit up, don't be shocked to see consumer Volta by the end of this year which would all but bury AMD for good.

Navi will save the world and cure cancer and make light speed space flight happen. Saw it on Star Trek man. Its going to happen! Gotta top nV's holodeck lol. Or lets see AMD marketing copy that.....

 

[razor1]

Sorry, that is not true.

I am not dreaming about anything, I am TRYING to discuss AMD's infinity fabric and the virtues that could possible bring to the high end Gamer. But it seems, that topic keeps getting trolled, and threadcrapped on by people who would rather talk about Nvidia, while badmouthing AMD.

Understand, it is OK to admire and seek knowledge and discussion about Vega technology. Or to even theorize about AMD's technologies and possible path in an AMD titled thread. (It is a Vega Rumor thread, after all.)




Secondly, I already have Nvidia stuff. I want more. Hence my hunger for HEDT GPU.

But what is troubling and tiring is the amount of negative and derogatory remarks being fired off by the same people, ad nauseam, while others are trying to have a legit discussion here. It is annoying and has to stop. Glad Kyle stepped in, because I am oldschool and believe People should have to show their driver's license (or CC#), to open an account. (That way cellphone kiddies need not apply)

Allowing this anti-AMD bashing to continue is just negative experience here at HardOCP. I know some of you get off on the giggles, but your actions are pushing away real readers, with real incomes and real interests. I am not afraid of a face to face, or a meet & greet. But being attacked for pushing an AMD subject, with posts trying to dissuade my buying habits, when ALL I am doing is speaking about technology is getting old. I buy whatever I want, whenever I want. I don't fret it. Not everyone here is a poor college student who dreams about owning SLI Titans. I remember wanting things back then too. The difference now is I can afford what I want ! Subsequently, Volta is insignificant and is 10 months away. because GDDR6 doesn't start production until Q2 2018. Coincidentally, the fact that Nvidia is locked out the console market is a death blow, only to Nvidia. It can only been seen as a boon for AMD, but nice try on your spin..

I am nearing 50 years of age and about to drop some serious cash on some full-on gaming rigs. I have good reason to believe AMD is going to come threw on their Infinity fabric and Vega, & offer us high end Gamer's, a TitanRipper like card later this year.

If Someone is not all for that goodness, then I would wager they are not a Gamer, or are peddling an agenda.

ALL BS aside

You tell me how infinity fabric solves this problem

https://cvw.cac.cornell.edu/codeopt/datalocality?AspxAutoDetectCookieSupport=1

Not only does it not solve this problem it makes the problem worse!

even in a unified cache THIS PROBLEM STILL IS THERE. There is no work around to this problem unless the architecture changes. The only way we can do it right now is through programming different models. Sorry but that is not transparency.

Data locality becomes even more important for coprocessors and accelerators (e.g., GPUs) than it is for host CPUs, simply because every item on the above list becomes individually more critical to performance. A fetch from main memory to a coprocessor or GPU cache is usually slower than it would be for a CPU fetch, and cache sizes are smaller than on CPUs. Furthermore, coprocessors and GPUs rely more heavily on vector processors for performance. Finally, data alignment is not so much an advantage as a requirement for making use of coprocessors and GPUs, since vector operations require it. While each device is different, it is fairly crucial to adhere to the principle of isolating data into boundary-aligned sets of contiguous vectors, when using coprocessors and accelerators. For more details related to coprocessors and GPUs, see the modules on the MIC coprocessor and GPGPU programming.

I am a blind person who doesn't know what I'm talking about, but they are teaching this in Cornell University? One of the best ivy league colleges in the world?

Yeah go figure.

Anyone that tires to put Infinity Fabric = to the demise of multi processing programming in any type of processor, doesn't know what the hell they are talking about. Its the opposite of that, it makes programming models for multi processor a step up from what it is now! It increases the the need for this type of programming model not decreases.

Now anyone that thinks unified caching and memory systems do this automagically.

As discussed earlier and pointed out by others too nV has this implemented but....

https://devblogs.nvidia.com/parallelforall/beyond-gpu-memory-limits-unified-memory-pascal/

Unified Memory Profiling
The NVIDIA Visual Profiler is one of the most important tools that every GPU developer who cares about performance should get familiar with. It can help you identify hot spots and performance problems in your application and guide you through the optimization process. CUDA 8 adds several new features related to Unified Memory profiling. First, a new segmented mode groups Unified Memory events on the timeline to present a high-level view (see Figure 8). The segments are colored according to their weight highlighting migrations and faults that take most of the time.

Do you know why they have this? It tells you were the segments are and if there is a slow down you can code around that to increase performance.

As I stated, even in a unified system, it doesn't matter.

So sitting here and pulling up things like unified cache because of infinity fabric is utter BS. infinity fabric doesn't give the capability of unified cache nor does unified cache solve the problems of data locality and latency incurred by that.

 

[Ocellaris]

Y'all look like you are in an endless battle to always have the last post in this thread.

It's possible to disagree with someone and not reply 5 seconds later, try it out.

 

[OutOfPhase]

The reality is you won't ever realistically solve an inherent performance/power problem in a single unit by scaling to multiple units.
You can of course brute force your way to a higher performance bracket that way, but at a power and die size (factory) cost. Both performance per watt and performance per factory cost $ (silicon space) decline. Generally speaking: "you're going the wrong way".

An interconnect design isn't going to fix this, and there is no magic interconnect which just solves inherent problems. There are some better choices and some worse ones, but all teams can play competently here.

 

[Algrim]

Fun fact: nVidia is part of the HyperTransport Consortium too. Whatever steroids AMD can infuse HyperTransport (aka Infinity Fabric) with, so can nVidia.

 

[Tup3x]

Navi will save the world and cure cancer and make light speed space flight happen. Saw it on Star Trek man. Its going to happen! Gotta top nV's holodeck lol. Or lets see AMD marketing copy that.....

I truly hope that Navi isn't just GCN @ 7nm. If it is things aren't going to be pretty.

 

[razor1]

I truly hope that Navi isn't just GCN @ 7nm. If it is things aren't going to be pretty.

Navi won't be GCN, tell ya what GCN should have died out before Fiji, that was the regular timing for a new architecture for AMD, typical 3 gens of products before a new architecture design. But it didn't happen, most likely because of lack of R&D money. So they stretched it out for another 2 years over their cycle times. With a stretched out time line, a deceased R&D budget, will accumulate to equal the similar cycle R&D budget.

This is exactly what they did with Ryzen too. They decided not to shrink and modify BD or Pile driver and focus purely on Ryzen. End results their lower R&D per quarter, over time, piles up to equal or surpass previous typical cycle amounts. But as we can see they could only compete with Intel's current crop with some advantages and disadvantages, so they had to price it accordingly. All the while Intel was doing mediocre upgrades to their lines. Essentially Intel was standing still.

GPU side isn't like that, that is why AMD had to switch to different nodes, they couldn't save that money but the end results of not having a completely new architecture saves money in the short term.

Pretty much they are weathering the storm so they can get the financials to keep up with nV, but in this case, unlike Intel, nV is not stalling, they are making big strides every generation, and this is where the power and efficiency discrepancies have come into GCN vs nV's different architecture. With Keplar AMD's HD 7xxx line was close right around there in performance, one can say beats Keplar as well, and right around the same power envelope. But as nV's focus shifted to increasing its perf/watt both sides up with performance and down with wattage, AMD's meager R&D could not have two teams working both GCN and Navi, to deliver both in the regular time lines while trying to stay competitive. This is also why we saw Vega come out so soon (relative to Polaris its only one year). At the end AMD's relative delays to higher end SKU's has hurt their marketshare greatly, Initially starting from the r3xx line delay.

Every gen to gen comparison from Kepler to GCN, we can see GCN is falling further and further behind, once equal now its at a point where top end Vega seems to compete with the gtx 1080.

So lets look at this way, HD7xxxx series could compete with the best Keplar, Fiji could compete with the second best Maxwell, now with Vega, its competing with the third best Pascal and with power consumption, its higher then all the cards its competing against, and power consumption differential got worse ever generation of GCN as well with respect to the cards its competing against.

 

[c3k]

Y'all look like you are in an endless battle to always have the last post in this thread.

It's possible to disagree with someone and not reply 5 seconds later, try it out.

No, I don't think it's possible.

 

[_mockingbird]

There will NOT be a livestream of Vega's pseudo launch event at SIGGRAPH2017.

 

[Elios]

There will NOT be a livestream of Vega's pseudo launch event at SIGGRAPH2017.

huh that doesn't sound good

 

[_mockingbird]

huh that doesn't sound good

Everything "sounds" bad for months now.

I said months ago on Anandtech that liquid-cooled Vega is barely faster a Geforce GTX 1080 and got banned.

 

[Nuby1Canuby]

PCGamesN has the following info on their site: https://www.pcgamesn.com/amd/amd-vega-gpu-specifications/#specs

Is this confirmed yet?

• AMD Vega specs
  The Vega GPU is built on 14nm silicon with the lower-end RX Vega XL sporting 3,584 GCN cores and the RX Vega XT and RX Vega XTX each rocking 4,096.
• AMD Vega price
  The air-cooled Vega Frontier Edition is on sale for $999, with rumours of the RX Vega cards ranging between $599 for the top GPU and $399 for the basic Vega.
• AMD Vega performance
  The pro-level Frontier Edition delivers between GTX 1070 and GTX 1080 gaming performance, but AMD is targeting ~GTX 1080 speeds for the RX Vega.
• The RX Vega XTX is the reportedly top-spec, water-cooled version, rocking 4,096 GCN cores and a TDP of 375W. The RX Vega XT uses the same 4,096 core GPU, but has air-cooling and a lower TDP of 285W. That kind of suggests it might have a lower clockspeed than the liquid-chilled chip too.
  At the bottom end of the AMD Vega tech tree is the AMD RX Vega XL, with a 3,584 GCN core GPU and the same 285W TDP. All three cards seem like they're going to be released with 8GB of HBM2 memory.
  

 

[Anarchist4000]

Advanced copy of the SIGGRAPH live-stream everyone can appreciate!

Spoiler: SIGGRAPH Livestream Recap

 

[NKD]

Everything "sounds" bad for months now.

I said months ago on Anandtech that liquid-cooled Vega is barely faster a Geforce GTX 1080 and got banned.

I don't think saying that will get you banned lol. I am sure it was more than that. Doesn't really work like that lol

 

[NKD]

PCGamesN has the following info on their site: https://www.pcgamesn.com/amd/amd-vega-gpu-specifications/#specs

Is this confirmed yet?

• AMD Vega specs
  The Vega GPU is built on 14nm silicon with the lower-end RX Vega XL sporting 3,584 GCN cores and the RX Vega XT and RX Vega XTX each rocking 4,096.
• AMD Vega price
  The air-cooled Vega Frontier Edition is on sale for $999, with rumours of the RX Vega cards ranging between $599 for the top GPU and $399 for the basic Vega.
• AMD Vega performance
  The pro-level Frontier Edition delivers between GTX 1070 and GTX 1080 gaming performance, but AMD is targeting ~GTX 1080 speeds for the RX Vega.
• The RX Vega XTX is the reportedly top-spec, water-cooled version, rocking 4,096 GCN cores and a TDP of 375W. The RX Vega XT uses the same 4,096 core GPU, but has air-cooling and a lower TDP of 285W. That kind of suggests it might have a lower clockspeed than the liquid-chilled chip too.
  At the bottom end of the AMD Vega tech tree is the AMD RX Vega XL, with a 3,584 GCN core GPU and the same 285W TDP. All three cards seem like they're going to be released with 8GB of HBM2 memory.

Lol. Looks like they justness recycled everything from last month and put it in one page.

 

[Elios]

PCGamesN has the following info on their site: https://www.pcgamesn.com/amd/amd-vega-gpu-specifications/#specs

Is this confirmed yet?

• AMD Vega specs
  The Vega GPU is built on 14nm silicon with the lower-end RX Vega XL sporting 3,584 GCN cores and the RX Vega XT and RX Vega XTX each rocking 4,096.
• AMD Vega price
  The air-cooled Vega Frontier Edition is on sale for $999, with rumours of the RX Vega cards ranging between $599 for the top GPU and $399 for the basic Vega.
• AMD Vega performance
  The pro-level Frontier Edition delivers between GTX 1070 and GTX 1080 gaming performance, but AMD is targeting ~GTX 1080 speeds for the RX Vega.
• The RX Vega XTX is the reportedly top-spec, water-cooled version, rocking 4,096 GCN cores and a TDP of 375W. The RX Vega XT uses the same 4,096 core GPU, but has air-cooling and a lower TDP of 285W. That kind of suggests it might have a lower clockspeed than the liquid-chilled chip too.
  At the bottom end of the AMD Vega tech tree is the AMD RX Vega XL, with a 3,584 GCN core GPU and the same 285W TDP. All three cards seem like they're going to be released with 8GB of HBM2 memory.

so...
XTX = Vega FE $650
XT =1070 $500
XL = 1060 $400

 

[Dayaks]

so...
XTX = Vega FE $650
XT =1070 $500
XL = 1060 $400

It's ok, just lock yourself into FreeSync and hope to god AMD doesn't get further behind!

 

[NKD]

so...
XTX = Vega FE $650
XT =1070 $500
XL = 1060 $400

When was the last time you saw amd over price their cards if they weren't up to par with performance? Never! So no you are not going to see a gtx 1070 competitor for $100 more then gtx 1070. Not happening!

and gtx 1060 for $400 in vega xl? Common man you should really check things before posting. rx 580 competes with gtx 1060 for same price or cheaper. I mean you really think AMD needs more expensive card to compete with 1060? lol

 

[_mockingbird]

I don't think saying that will get you banned lol. I am sure it was more than that. Doesn't really work like that lol

You should try it.

Literally said anything negative about AMD on Anandtech and you're banned.

 

[Dayaks]

When was the last time you saw amd over price their cards if they weren't up to par with performance? Never! So no you are not going to see a gtx 1070 competitor for $100 more then gtx 1070. Not happening!

and gtx 1060 for $400 in vega xl? Common man you should really check things before posting. rx 580 competes with gtx 1060 for same price or cheaper. I mean you really think AMD needs more expensive card to compete with 1060? lol

... Fury X and Fury Nano. Wasn't [H]'s criticism of the price what started the whole fiasco between AMD and H? RX480 launched too high.

I actually expect AMD to be $50 too expensjve anymore.

 

[_mockingbird]

When was the last time you saw amd over price their cards if they weren't up to par with performance? Never! So no you are not going to see a gtx 1070 competitor for $100 more then gtx 1070. Not happening!

and gtx 1060 for $400 in vega xl? Common man you should really check things before posting. rx 580 competes with gtx 1060 for same price or cheaper. I mean you really think AMD needs more expensive card to compete with 1060? lol

2015

 

[SickBeast]

You should try it.

Literally said anything negative about AMD on Anandtech and you're banned.

I got banned at ATF last night for pointing out the fact that AMD's viral marketers do not disclose their affiliation. Thank God Kyle has not sold out to AMD or a large media conglomerate. This site and forum is the last of a dying breed.

 

[Dayaks]

I got banned at ATF last night for pointing out the fact that AMD's viral marketers do not disclose their affiliation. Thank God Kyle has not sold out to AMD or a large media conglomerate. This site and forum is the last of a dying breed.

He has a patreon now! Just imagine.. a site for the people funded by the people.

 

[Maverick177]

I got banned at ATF last night for pointing out the fact that AMD's viral marketers do not disclose their affiliation. Thank God Kyle has not sold out to AMD or a large media conglomerate. This site and forum is the last of a dying breed.

Funny, because that's not what I read. You raged and accused some posters of shilling without any hard evidence.

 

[SickBeast]

Funny, because that's not what I read. You raged and accused some posters of shilling without any hard evidence.

Raged? I simply pointed out the truth. I never directly accused anyone of anything. I simply pointed out the fact that there have been AMD shills posting at ATF for a long time, and that the AMD shills do no disclose their affiliation. I can post proof of this if you like. But I really think you're sidetracking this discussion with inflammatory non truths. Like I said, I never "raged" and I never accused anyone of anything. If you can't handle the truth stop reading my posts.

 

[Elios]

When was the last time you saw amd over price their cards if they weren't up to par with performance? Never! So no you are not going to see a gtx 1070 competitor for $100 more then gtx 1070. Not happening!

and gtx 1060 for $400 in vega xl? Common man you should really check things before posting. rx 580 competes with gtx 1060 for same price or cheaper. I mean you really think AMD needs more expensive card to compete with 1060? lol

no but i think that is what happened

something didn't work out as expected for AMD part of which i think was bad HBM2 yields that caused the massive delay

the high prices are imo do to HBM2 costing a fuck ton it's just not worth it on consumer cards maybe for a Titan / FE level card but Nv provided you don't need it even for that with GDDR5X and now GDDR6 coming

even my 1070 with GDDR5 and a narrower bus gets near 0 gain from ram overclocks so it's clear for gaming its not bandwidth starved

 

[FrgMstr]

Raged? I simply pointed out the truth. I never directly accused anyone of anything. I simply pointed out the fact that there have been AMD shills posting at ATF for a long time, and that the AMD shills do no disclose their affiliation. I can post proof of this if you like. But I really think you're sidetracking this discussion with inflammatory non truths. Like I said, I never "raged" and I never accused anyone of anything. If you can't handle the truth stop reading my posts.

Post the proof.

 

[Elios]

well then ...

getting the feeling something really bad happened with Vega wile in development ...

 

[bildad]

well then ...

getting the feeling something really bad happened with Vega wile in development ...

They pinched the loaf a little too short. Needs more fiber.

 

[sine wave]

Navi won't be GCN, tell ya what GCN should have died out before Fiji, that was the regular timing for a new architecture for AMD, typical 3 gens of products before a new architecture design. But it didn't happen, most likely because of lack of R&D money. So they stretched it out for another 2 years over their cycle times. With a stretched out time line, a deceased R&D budget, will accumulate to equal the similar cycle R&D budget.

This is exactly what they did with Ryzen too. They decided not to shrink and modify BD or Pile driver and focus purely on Ryzen. End results their lower R&D per quarter, over time, piles up to equal or surpass previous typical cycle amounts. But as we can see they could only compete with Intel's current crop with some advantages and disadvantages, so they had to price it accordingly. All the while Intel was doing mediocre upgrades to their lines. Essentially Intel was standing still.

GPU side isn't like that, that is why AMD had to switch to different nodes, they couldn't save that money but the end results of not having a completely new architecture saves money in the short term.

Pretty much they are weathering the storm so they can get the financials to keep up with nV, but in this case, unlike Intel, nV is not stalling, they are making big strides every generation, and this is where the power and efficiency discrepancies have come into GCN vs nV's different architecture. With Keplar AMD's HD 7xxx line was close right around there in performance, one can say beats Keplar as well, and right around the same power envelope. But as nV's focus shifted to increasing its perf/watt both sides up with performance and down with wattage, AMD's meager R&D could not have two teams working both GCN and Navi, to deliver both in the regular time lines while trying to stay competitive. This is also why we saw Vega come out so soon (relative to Polaris its only one year). At the end AMD's relative delays to higher end SKU's has hurt their marketshare greatly, Initially starting from the r3xx line delay.

Every gen to gen comparison from Kepler to GCN, we can see GCN is falling further and further behind, once equal now its at a point where top end Vega seems to compete with the gtx 1080.

So lets look at this way, HD7xxxx series could compete with the best Keplar, Fiji could compete with the second best Maxwell, now with Vega, its competing with the third best Pascal and with power consumption, its higher then all the cards its competing against, and power consumption differential got worse ever generation of GCN as well with respect to the cards its competing against.

Stop trolling us with Nvidia hyperbole. There is already a thread section for your incessant rants. (btw, nobody cares about Fiji, Kepler, Maxwell, Tahiti, Pascal, etc..)



@ Everyone else:

AMD is following a different path than both Intel and Nvidia. I've mentioned this many times and that over the better part of the last decade AMD has been in pursuance of their HSA (ie: heterogeneous Computing) for longer than both Companies. AMD has been acquiring and developing it's building blocks and it's backbone, to push their idea and vision forward. Whatever stumbling steps Intel and Nvidia are having now, AMD had 5 years ago.

As you can see, AMD has an APU, Intel and Nvidia don't.

As you can see, AMD has the One X and PS4 Pro, Intel and Nvidia don't.

As you can see, AMD has ThreadRipper and Intel is on it's heels trying to patch holes in their platforms... chasing AMD.

As you can see, AMD has EPYC enterprise chips and Intel is also on their heels, looking for a response.


AMD has their first batch of their "Secret Sauce" ready and are now tasting it on all the chips & kibbles. But Infinity fabric 1.0 technology will tick-tock in odd cadence with Vega/Navi & Zen/2.
They will advance off each other in this fashion. And 7nm Navi will make this easily known, perhaps sitting on IF 2.0.

Nothing in the Industry that AMD has done, or trying to do, gives anyone the right to demean them. You don't see Jen-Hsun Huang mocking Dr Su, do you ??


RX Vega can easily sit on top of AMD's fabric to solve many of the inherent problems that typically goes with x2 packaging. It doesn't take a brain surgeon to figure out that no matter how good Croosfire (or, SLI for that matter) really is, or was. That two side by side GPUs chips are better and cheaper and cooler than 2 separate cards. Win/Win/Win



Strictly within AMD's IP, play with the idea that AMD has fabric and could work magic with two HBCC connected (coherently) and possibly allow two GPU to share the same memory space. Thus eliminating AFR all together ?

GPU GPU
| |
HBCC - HBCC
\ /
HBM2

It is just an idea (?).
But, no more facetious than anything anyone else has tossed up. There are endless ways to combine pieces of the puzzle. AMD Engineers have probably played with their own technology, like lego blocks playing with ideas. And as such, AMD "Chiplet" patent looks like a big WIN !

It's obviously using a more advanced form of "Infinity Fabric", as the back drop to their Chiplet design. AMD's back room is probably a mess with secret sauce splattered all over the place.



Lastly, this is a Rumor thread, but Raja did say (RX) Vega being the first to use infinity fabric on their GPUs. So even in the simplest terms, it will be better than SLI/Xfire, and I see no reason why AMD can not have their HEDT Vidcard out before xmas. I understand Nvidia's fear, because if AMD is sandbagging and fabric allows vega to x2 without afr, then you might as well nickname it VoltaRipper !!


Thanks Dr Su for thinking about the Gamer!

 

[razor1]

top trolling us with Nvidia hyperbole. There is already a thread section for your incessant rants. (btw, nobody cares about Fiji, Kepler, Maxwell, Tahiti, Pascal, etc..)

lets ignore the useless crap shell we, still haven't given me an answer to the latency problem, why? Why are you ignoring the problem that is straight in your face?

@ Everyone else:

AMD is following a different path than both Intel and Nvidia. I've mentioned this many times and that over the better part of the last decade AMD has been in pursuance of their HSA (ie: heterogeneous Computing) for longer than both Companies. AMD has been acquiring and developing it's building blocks and it's backbone, to push their idea and vision forward. Whatever stumbling steps Intel and Nvidia are having now, AMD had 5 years ago.

They are following the same path, in different ways, Intel is part of the HSA lol forgot about that?

As you can see, AMD has an APU, Intel and Nvidia don't.

Both APU and IGP's are bandwidth limited, so what does that mean to anyone? Nothing in the long run.

As you can see, AMD has the One X and PS4 Pro, Intel and Nvidia don't.

What happened to Nintendo Wii? Why did AMD loose it?

As you can see, AMD has ThreadRipper and Intel is on it's heels trying to patch holes in their platforms... chasing AMD.

They aren't chasing anyone, Intel's top Skylake parts all core boost over 4.0 GHZ, same and more core counts as Thread ripper for now.

As you can see, AMD has EPYC enterprise chips and Intel is also on their heels, looking for a response.

Did you watch the video I posted, its done by a person that knows server tech not much difference between the two platforms. Variations yes but each line from both Intel and AMD solve the same problems. Eypc is not going to be a success in any software that requires data localization, which is pretty much all transaction based servers, that is like 75% of the market if not more. So even if the option if available through HP, Dell, or Lenovo, sales will be slow.

AMD has their first batch of their "Secret Sauce" ready and are now tasting it on all the chips & kibbles. But Infinity fabric 1.0 technology will tick-tock in odd cadence with Vega/Navi & Zen/2.
They will advance off each other in this fashion. And 7nm Navi will make this easily known, perhaps sitting on IF 2.0.

yes its so secret that only people with imaginary friends can see it

Nothing in the Industry that AMD has done, or trying to do, gives anyone the right to demean them. You don't see Jen-Hsun Huang mocking Dr Su, do you ??

That was Intel being Stupid, AMD being stupid, we have seen that many times too, they have more stupid excuses and marketing tactics than both Intel and nV combined, Poor Volta was the recent idiotic marketing ploy by AMD, we can go through ever launch AMD has done, any product and find similar mistakes.

RX Vega can easily sit on top of AMD's fabric to solve many of the inherent problems that typically goes with x2 packaging. It doesn't take a brain surgeon to figure out that no matter how good Croosfire (or, SLI for that matter) really is, or was. That two side by side GPUs chips are better and cheaper and cooler than 2 separate cards. Win/Win/Win

No it doesn't take a brain surgeon, it takes a good EE and any EE doesn't even need to be good to see that is BS.

Strictly within AMD's IP, play with the idea that AMD has fabric and could work magic with two HBCC connected (coherently) and possibly allow two GPU to share the same memory space. Thus eliminating AFR all together ?

GPU GPU
| |
HBCC - HBCC
\ /
HBM2

It is just an idea (?).
But, no more facetious than anything anyone else has tossed up. There are endless ways to combine pieces of the puzzle. AMD Engineers have probably played with their own technology, like lego blocks playing with ideas. And as such, AMD "Chiplet" patent looks like a big WIN !

AMD went to HBM first and it hurt them, nV went to HBM 2 and they had success..... Why the difference?

It's obviously using a more advanced form of "Infinity Fabric", as the back drop to their Chiplet design. AMD's back room is probably a mess with secret sauce splattered all over the place.

Lastly, this is a Rumor thread, but Raja did say (RX) Vega being the first to use infinity fabric on their GPUs. So even in the simplest terms, it will be better than SLI/Xfire, and I see no reason why AMD can not have their HEDT Vidcard out before xmas. I understand Nvidia's fear, because if AMD is sandbagging and fabric allows vega to x2 without afr, then you might as well nickname it VoltaRipper !!


Thanks Dr Su for thinking about the Gamer!

Again imaginary friends will help in that yes

 

[OutOfPhase]

So it's now not actually the infinity fabric at all but "secret sauce"?

Come on man. This is getting pretty far off the rails of reality.

 

[razor1]

So it's now not actually the infinity fabric at all but "secret sauce"?

Come on man. This is getting pretty far off the rails of reality.

Everything AMD does is a secret sauce lol, they could have been smoking a J and doing shrooms for the last year and what ever product comes out its all magically delicious.

Sine Wave please answer my point on data locality then we might be able to have a serious discussion about this topic, otherwise, your posts are becoming a rambling sonnet of not even rumors.......Its just made up crap.

Rumors usually have some basis of truth to tihem, not just fictitious ramblings.

 

[Sodapopjones]

Everything AMD does is a secret sauce lol, they could have been smoking a J and doing shrooms for the last year and what ever product comes out its all magically delicious.

Sine Wave please answer my point on data locality then we might be able to have a serious discussion about this topic, otherwise, your posts are becoming a rambling sonnet of not even rumors.......Its just made up crap.

Rumors usually have some basis of truth to tihem, not just fictitious ramblings.

That's generally what happens when someone is trying to prove a point and won't admit they're wrong. I noticed he likes to plaster paragraphs of crap, rather than a concise statement.

 

[razor1]

That's generally what happens when someone is trying to prove a point and won't admit they're wrong.

I don't care if he admits he is wrong, just the simple thought process will suffice. I can counter the latency issue, but it requires a fundamental change in architecture and how it communicates with its self and other chips, and I have a pretty good idea of how to do it too even with current architectures to maintain scalablity. But it will introduce a new type of control silicon specific for GPU's. Now I'm not an EE, but theoretically I can figure it out actual implementation lol, hell no.

The general theory is part of what came about by my posts with Anarachist, close to a year ago. It was a heated discussion but mutually I understood where he was coming from, and modified what I was thinking.

This guy wants to discuss possible rumors, but he isn't willing to understand what the current pitfalls are, which is kinda what we have to look at when the problem is there. Can't fix something without looking at the problem first :/