We have 14nm Vega until the mid-2019 arrival of Navi on 7nm - How about a "Gigahertz Edition" redo?

[Maddness]

I wouldn't buy into the "we have abandoned the gaming market" nonsense. If AMD can make a profit, even a small one, then they will be there.


Chasing the HPC market is what they need to do in order to rebuilt their R&D budgets. Lisa Su likes to compete. More importantly, she likes to win. Given the shakeups she has overseen, coupled with the wonderful success Ryzen has been, all points toward AMD coming out swinging. It might take 18-24 months but I believe it will happen.

I'm actually putting my money where my mouth is abd picking up more AMD stock. It has no where to go but up. Way up.

It was reasonably easy for AMD with RyZen to catch Intel. They have almost been stagnant since Sandy Bridge launched. The same cannot be said for Nvidia. RTG certainly has a battle on its hands there. Still, now that Raja has gone I'm really interested In the direction they take.

 

[IdiotInCharge]

Also, a lot of AMD fans seemed to sit Vega out, even when it was affordable.

You mean for the week that it was MSRP?

Generally speaking, Vega was never affordable. And being about 980Ti performance and power draw and several years later...

AMD didn't do their fans any favors

 

[IdiotInCharge]

They have almost been stagnant since Sandy Bridge launched.

I know people like to tout this, but Intel most certainly hasn't been stagnant: they've just been largely putting their tech in places other than the desktop, but they never stopped advancing.

 

[Hagrid]

I know people like to tout this, but Intel most certainly hasn't been stagnant: they've just been largely putting their tech in places other than the desktop, but they never stopped advancing.

I read it as stagnant, not going forward much, in desktop cpu's. Unless you read it as overall, not sure.
It really isn't too much of a big deal since they are becoming quite powerful enough to play most games with no problem. More cores for productivity is probably more likely.(Or if they can make games use all available cores)

 

[IdiotInCharge]

I read it as stagnant, not going forward much, in desktop cpu's. Unless you read it as overall, not sure.
It really isn't too much of a big deal since they are becoming quite powerful enough to play most games with no problem. More cores for productivity is probably more likely.(Or if they can make games use all available cores)

That's the thing, they have been putting their tech into their desktop CPUs, but they've also been hamstrung by their process delays.

Other challenge is that games are generally single-core dependent by nature. Plenty of work has been done to paralellize workloads and plenty of stuff can be parallelized and ow-level APIs like DX12 and Vulcan have definitely helped here a bit too. However, the main threads especially in multiplayer shooters remain difficult due to the dependencies involved, and further, as games advance and evolve, as we hope they do, single-core affinity will likely be needed first before said evolutions can themselves be paralellized.

Meaning that having enough cores and threads and highest single-core performance at a particular price-point remains the best recommendation for gaming. At the top-end, that's the 8700k; below that, AMD takes a fairly good chunk.

 

[Maddness]

I read it as stagnant, not going forward much, in desktop cpu's. Unless you read it as overall, not sure.
It really isn't too much of a big deal since they are becoming quite powerful enough to play most games with no problem. More cores for productivity is probably more likely.(Or if they can make games use all available cores)

That is pretty much what i meant. How long have we been stuck with 4 core CPU's. Only 6+ cores on high end chipsets. Now that RyZen has made up to 8 core in the main stream, hopefully this will kick Intel into gear. I know they have had great difficulty getting the 10nm process into production. But we have only been getting up to 10% increase since Sandy Bridge launched per generation. Not really any incentive to upgrade like GPU's.

 

[Shop_vac]

You mean for the week that it was MSRP?

Generally speaking, Vega was never affordable. And being about 980Ti performance and power draw and several years later...

AMD didn't do their fans any favors

Vega doesn't average 980ti performance (even as a meme), and 1080 performance for $399 (56 tuned) is an incredible gesture. It wasn't reality for the product, but AMD had the right idea.


I'd be more concerned that most hawaii/Polaris/fury users sat it out. Almost like they were insulted by the release or something.

 

[Pieter3dnow]

Vega doesn't average 980ti performance (even as a meme), and 1080 performance for $399 (56 tuned) is an incredible gesture. It wasn't reality for the product, but AMD had the right idea.
I'd be more concerned that most hawaii/Polaris/fury users sat it out. Almost like they were insulted by the release or something.

By the time Vega hit the market the 56 looked appealing the prices went up and the custom versions were out of reach.
You can't blame people that buy on price performance ratio to suddenly not do that.

 

[N4CR]

Navi is not a pipe-cleaner either, not too sure why the person you replied to said/thinks that, they have specifically stated that the 7nm Vega Instinct is gonna be the pipe cleaner of their GPU roadmap. I do think Navi will be something special, it's going to further expand upon the IF-minded design of Vega and I believe we will see something akin to R700's design philosophy w/o all of the baggage and bullshit that has plagued both camps mGPU setups since their inception. DX12/Vulkan mGPU lends itself to my line of thinking here as well.

I was remembering the old roadmaps where IIRC navi was supposed to be the first 7nm product.
You are right, with more recent info it seems 7nm vega will be the pipe cleaner, which makes sense given that it's still related to vega and 'known'. That information didn't exist until this year.

Navi will bring IF to the forefront. I am expecting them to make an MCM design. They already have 500gb/sec IF in Vega now for testing this... Polaris also had IF links, not sure what speed though. They were what revolutionised oil and gas exploration via SSG.

I'd be more concerned that most hawaii/Polaris/fury users sat it out. Almost like they were insulted by the release or something.

I was expecting between 1080 and Ti and Vega delivered that. If it was more Ti I would have pulled trigger.
Thing is for the few games I play, 290x at 1440p/60 is plenty of card. For now I'm happy to wait for Navi, but Vega 7nm may be tempting...

 

[Shop_vac]

By the time Vega hit the market the 56 looked appealing to the prices went up and the custom versions were out of reach.
You can't blame people that buy on price performance ratio to suddenly not do that.

If AMD needs to set unrealistic prices (that they need rebates to guarantee ), it explains why there won't be a Vega refresh. Not for gaming anyway.

 

[Maddness]

If AMD needs to set unrealistic prices (that they need rebates to guarantee ), it explains why there won't be a Vega refresh. Not for gaming anyway.

It wasn't AMD how set those prices. It was the bloody retailers.

 

[Jedibeeftrix]

The VEGAs love that fast ram though. The difference in min frame rates is insane, along with the avg fame times in regards to gaming.
As far as mining goes, going from 950 to 1100 nets me 20+% for less then 10W more power usage.

Interesting, I have been asking whether Vega perrformance was memory bound, on the assumption that it is the most obvious place to seek change for a "Gigahertz Edition", but people have so far pooh-pooh'ed the notion!

 

[Pieter3dnow]

There is a difference where you assume things and where reality starts. Check Buildzoid stuff on Vega. In the end the scaling does not come from one side alone when you do overclocking as he does it shows that the memory speed gets capped and then Vega scales with clock. He even mentions the ratio.

This does not mean that if you ramp up to 2000 memory speed and 2500 clock on Vega it still holds the same progress it hits a wall where it stops the increase by the same margin as ccityinstaller experienced and it drops quite fast.

It does by a margin where AMD does not see any urgency in releasing a 12nm or a 7nm for consumers. They must be crazy not to do this right if all it took were better clock speeds ?

 

[luminousone]

Interesting, I have been asking whether Vega perrformance was memory bound, on the assumption that it is the most obvious place to seek change for a "Gigahertz Edition", but people have so far pooh-pooh'ed the notion!

Its a shame that Vega didn't have a 4 stack(4096bit bus) version, 800GB-1TB bandwidth would have done wonders for the thing. Some of the memory optimizations in Polaris, where changed in Vega, I have wondered if AMD made memory use worse by trying to make some of those features "programmable". I coulda swore that Fiji/Polaris had a form of fixed pipeline primitive discard(aka not a bloody shader).

As well, the old bus's on most cards are based on teams of 32bit channels that can work independently, which is great for lots of little random writes, which is what a video card does, The HBM path does have the disadvantage of creating just 1-4(vega-m=1,vega64=2,Fiji/vega?7nm=4) channels that use large bit width(1024bit) per channel strides, I often wonder what the consequences of that are. When I have a 32bit write out to some location do I have to read in that location first so that I can bit mask in the change before writing(thus cutting the real bandwidth in half), Even a large l2/3 cache on the card might not be enough to obtain benefits of write combining(the feature the 486 was the first to implement!).

LC(low cost) HBM3, with 512bit channel's, I would bet performs better for most use cases when comparing similar bandwidth configs( 2stack hbm2 vs 4 stack lc hbm3).

 

[Jedibeeftrix]

I'm still not convinced that amd has anything more than a re-launched Vega10 on 14nm for thwe high-end.

But i'm becoming more and more convinced that the vega mid-range that disappeared may reappear as a 12nm product using a single stack of 2.4Gbps 8GB HBM2.

 

[N4CR]

I'm still not convinced that amd has anything more than a re-launched Vega10 on 14nm for thwe high-end.

But i'm becoming more and more convinced that the vega mid-range that disappeared may reappear as a 12nm product using a single stack of 2.4Gbps 8GB HBM2.

Doesn't explain the leaked 3dmark bench though. It was 32Gb... so there has to be something big coming through the pipeline also.

 

[Jedibeeftrix]

Doesn't explain the leaked 3dmark bench though. It was 32Gb... so there has to be something big coming through the pipeline also.

Oh, I have no doubt that that Vega20 on 7nm is coming, but i don't think consumers will see it.

7nm with the huge multi-patterning required of a non-EUV process is going to make it very expensive, as will the interproser required for four stacks of HBM2.
It will arrive as a low volume Instinct product, and may eventually make its way into a Frontier Edition semi-pro product, but not before Navi has arrived.

 

[N4CR]

Oh, I have no doubt that that Vega20 on 7nm is coming, but i don't think consumers will see it.

7nm with the huge multi-patterning required of a non-EUV process is going to make it very expensive, as will the interproser required for four stacks of HBM2.
It will arrive as a low volume Instinct product, and may eventually make its way into a Frontier Edition semi-pro product, but not before Navi has arrived.

Doesn't sound like something AMD would do, they don't have the budget and scale to treat wafer production like that, they have MOQs to meet usually. They like to reuse things as much as possible, for as many different markets in order to keep costs down.
I'd say Navi will be later 2019, hence why they are trying to get initial 7nm Vega product out in 2018, likely late q4. Gives it a year and a bit on the market, a quick fix for those who won't pull trigger for Navi or Nextgen.

 

[Jedibeeftrix]

Doesn't sound like something AMD would do, they don't have the budget and scale to treat wafer production like that, they have MOQs to meet usually. They like to reuse things as much as possible, for as many different markets in order to keep costs down.
I'd say Navi will be later 2019, hence why they are trying to get initial 7nm Vega product out in 2018, likely late q4. Gives it a year and a bit on the market, a quick fix for those who won't pull trigger for Navi or Nextgen.

it is what AMD have said they will do, i.e.
1. release an AI/ML Vega20 on 7nm as a pipecleaner (will require expensive multi-patterning as it is a non-EUV product)
2. follwed up by a consumer part we all know and love as Navi, around Q2 2019.

Now, there is nothing stopping them releasing Vega20 as a consumer prodcut (except frightening expense - for little extra gaming performance beyond Vega10).
I just don't think they'll bother. Sell it in small volumes to the people willing to pay big-dollar for it. You know, AI/ML types.

 

[N4CR]

it is what AMD have said they will do, i.e.
1. release an AI/ML Vega20 on 7nm as a pipecleaner (will require expensive multi-patterning as it is a non-EUV product)
2. follwed up by a consumer part we all know and love as Navi, around Q2 2019.

Now, there is nothing stopping them releasing Vega20 as a consumer prodcut (except frightening expense - for little extra gaming performance beyond Vega10).
I just don't think they'll bother. Sell it in small volumes to the people willing to pay big-dollar for it. You know, AI/ML types.

Probably right but this only makes sense if they can't get the clocks out of it like rumoured. AMD doesn't typically do halo, limited run chips like that even for enterprise markets typically. It doesn't follow their usual mode of operation - I would expect a similar approach like Vega was, FE version then consumer version to at least respond somehow to the 11xx series form Nvidia. That's what I would do anyway.. at least provide options and another choice to steal some thunder.
AMD is a tight ship since Raja fucked off and Lisa took over, so we probably won't know until a week or so prior.

 

[Maddness]

Probably right but this only makes sense if they can't get the clocks out of it like rumoured. AMD doesn't typically do halo, limited run chips like that even for enterprise markets typically. It doesn't follow their usual mode of operation - I would expect a similar approach like Vega was, FE version then consumer version to at least respond somehow to the 11xx series form Nvidia. That's what I would do anyway.. at least provide options and another choice to steal some thunder.
AMD is a tight ship since Raja fucked off and Lisa took over, so we probably won't know until a week or so prior.

It is what AMD have said previously, not that things don't change.

 

[N4CR]

It is what AMD have said previously, not that things don't change.

Was that the earnings call?

 

[NKD]

It is what AMD have said previously, not that things don't change.

I think that was just he said, she said rumor. I don't think I have officially heard there wont be anything consumer related in 2018. If not so be it, but If they can get decent sustained clocks from vega 20, I have a hard time believe we don't see the consumer version of it until navi comes along.

 

[Maddness]

This was one article that i read, but there was more. I took it to mean no new performance cards for the masses this year.

http://hexus.net/tech/news/graphics/113903-amd-details-2018-plans-radeon-gpus/

AMD took the opportunity before CES to brief the tech press fraternity with details on where it is going with respect to consumer graphics in 2018. Unlike the CPU announcements made at the same event, content for GPU releases was altogether lighter.

The biggest takeaway we had, which remains speculation until proven otherwise, is that AMD's not planning on releasing its next-generation Navi GPU architecture this year, and we can be reasonably certain of this due to a lack of announcements.

AMD CEO, Lisa Su, outlined the high-level trajectory for the graphics roadmap by indicating that existing Vega will transition over to 7nm - to be first seen in the machine-learning Instinct space - before Navi has its bow, likely in 2019. That's not great news for enthusiasts hoping that AMD would come back with GeForce-toppling technology in 2018.

More proof that this is indeed the case came later in the presentation. In fact, AMD wouldn't offer a firm comment that we would see the Vega architecture in the mainstream space anytime soon, either, so all we can go on is that RX Vega 56 and 64 will be the only members of the Vega club for the foreseeable future.

To that end, AMD says that all AIB partners now have access to custom designs that will filter through the market this month. Internal testing shows them to be up to 12 per cent faster than the reference card released in August last year, but we find that surprising given that the well-cooled Sapphire Nitro+ wasn't much quicker than the base card.

So what is new from the GPU front? We already know about the semi-custom project alongside Intel where AMD has supplied a 24 CU RX Vega chip to offer enhanced graphics to the Core G-series CPUs.

It is likely that a variant of that GPU will find itself as the guts of the Radeon Vega mobile discrete GPU line-up, though other than mentioning the impressive 1.7mm package z-height and HBM2 memory usage (which is the same as the Intel SKU), no details were given with respect to specifications, performance, and release date.

We will learn more in due course, but just like the desktop variant, AMD has its work cut out against the existing Pascal architecture behind the 10-series mobile chips.

Perhaps what's more telling is that AMD is positioning this unreleased chip as a mobile workstation offering as much as a gaming champion.

One bit of good news is that Radeons will support FreeSync technology over HDMI 2.1. The standard offers an impressive level of futureproofing, from oodles of bandwidth to higher refresh rates. Best of all, the hope is that TVs coming out later this year will feature the standard, meaning the potential for lag- and stutter-free gaming on the big screen, not just the computer monitor.

And that's yer lot as far as consumer Radeon graphics is concerned in 2018. That or AMD is playing the graphics card very close to its chest at the turn of the year.

 

[Pieter3dnow]

Okay but what is the need for the strategy to do this with AMD? Personal changes make it clear is that there not having the performance. If you recall AM3+ strategy after Piledriver was the same silence.
2018 is a lost year for the AMD consumer space, Nvidia will release something new and more then likely faster then their previous consumer card.

 

[NKD]

One bit of good news is that Radeons will support FreeSync technology over HDMI 2.1. The standard offers an impressive level of futureproofing, from oodles of bandwidth to higher refresh rates. Best of all, the hope is that TVs coming out later this year will feature the standard, meaning the potential for lag- and stutter-free gaming on the big screen, not just the computer monitor.

And that's yer lot as far as consumer Radeon graphics is concerned in 2018. That or AMD is playing the graphics card very close to its chest at the turn of the year.

That is true. As I mentioned I wouldn't be surprised to see no consumer card in 2018. Also if they do they are playing it close to the chest. After Lisa came along things have been quiet. No hype coming from RTG. I wouldn't be surprised if they told them to just focus on Navi and 7nm. Since mining has probably helped their sales quiet a bit I think it might have saved them from getting too hurt when it comes to sales. Use that revenue to focus on the next upcoming product and maximize your potential there, instead of hyping polaris.


this is from Anandtech,
"In our recent interview with AMD’s CEO, when asked if the GPU market will at some point have to bifurcate between gaming focused and compute focused designs, Dr. Lisa Su stated that ‘it must be the case’."

They might be moving in the direction of Nvidia and how they do their cards. Compute focused first and then consumer focused.

 

[NKD]

Okay but what is the need for the strategy to do this with AMD? Personal changes make it clear is that there not having the performance. If you recall AM3+ strategy after Piledriver was the same silence.
2018 is a lost year for the AMD consumer space, Nvidia will release something new and more then likely faster then their previous consumer card.

I think Lisa cleaned house. I think Raja was likely going to intel for a while. May be AMD made him head of RTG to have him stay who knows but I truly don't think his heart and mind were in the same place. He was a disappointment to say the least. He overhyped every release, marketing was stupid rebellious crap.

I have no idea what was going on with RTG. Lisa probably saw it the best Raja goes where he wants. I think this is what RTG needs, less fud coming across and more time making the products. I am honestly perfectly fine with nothing in 2018, they might get hurt in consumer space but they likely sold all the cards they made, due to mining. So revenue might not be hurt at all. Take this time and money to make a decent product. I think silence is what RTG needs. It was all talk no walk before. I would rather have them not talk for a while and work on decent product.

 

[DooKey]

Okay but what is the need for the strategy to do this with AMD? Personal changes make it clear is that there not having the performance. If you recall AM3+ strategy after Piledriver was the same silence.
2018 is a lost year for the AMD consumer space, Nvidia will release something new and more then likely faster then their previous consumer card.

And their market share will approach 85% as a result.

 

[Gideon]

And their market share will approach 85% as a result.

AMD increased its overall graphics chips market share by 8% quarter over quarter. Nvidia suffered a -6% decline quarter over quarter.

On discrete GPUs, AMD also posted a notable quarter over quarter increase from 27.2% in Q3 2017 to 33.7% in Q4. Nvidia’s share dipped from 72.8% to 66.3%.

Jon Peddie Research also pointed out that AMD was the primary beneficiary of the 3 million discrete GPUs, worth $776 million, sold to cryptocurrency miners in Q4 2017.

I dont think the 1180 will be that sought after product it looks to be on par with a 1080ti, if AMD uses this time to refocus their attention and delivers a competitor to what the 1180ti will be next year then it's all good for us.

 

[Pieter3dnow]

Would rather stay in the consumer space then talking about percentages that means absolutely nothing to me as a consumer. Because if AMD could improve any of their positions in the market with selling hardware that does not matter to the consumer space is a dead end either way (no need for me to buy a product that does machine learning/AI). That AMD can make money this way is good for AMD.

Even if you believe that Navi is going to be MCM that would have to meet some pretty decent specs if you compare it to the same package as Threadripper because the cooling for a gpu as Vega was pretty extreme already (Liquid Edition). And that would still have to clear hurdles as performance and power to make it feasible (no expensive HBM2 solution for mainstream product). And that is supposed to be for 2019.

 

[N4CR]

Perhaps they realized that GCN isn't going to win them the battle so have held off. But Navi is still GCN isn't it?
In another thread someone mentioned the unbalanced design and geometry bottleneck they are trying to solve, it might be 'the last stand' of GCN as such, until next architecture comes along.

 

[KazeoHin]

Perhaps they realized that GCN isn't going to win them the battle so have held off. But Navi is still GCN isn't it?
In another thread someone mentioned the unbalanced design and geometry bottleneck they are trying to solve, it might be 'the last stand' of GCN as such, until next architecture comes along.

GCN needs to be put out to pasture. Each successive GCN launch is more disappointing than the last...

If AMD wants to capture both gaming and mining markets they'll need to have two different architectures. GCN was just too good at compute operations, and that is great for overall sales, but it's doing them a disservice in the gaming market.

 

[Pieter3dnow]

GCN is a RISC SIMD (or rather SIMT) microarchitecture contrasting the VLIW SIMD architecture of TeraScale. GCN requires considerably more transistors than TeraScale, but offers advantages for GPGPU computation. It makes the compiler simpler and should also lead to better utilization.[citation needed]

https://en.wikipedia.org/wiki/Graphics_Core_Next

GCN should not be the limiting factor the way people think it is. It is just how key components in the design are located and addressed..

 

[luminousone]

I generally think AMD GPU's are more limited by issues in the memory subsystem, the execution engine is fantastic its why they hashed so well in mining until proof of work algs went deep into memory limited problems to combat ASIC's.

 

[IdiotInCharge]

I generally think AMD GPU's are more limited by issues in the memory subsystem, the execution engine is fantastic its why they hashed so well in mining until proof of work algs went deep into memory limited problems to combat ASIC's.

Memory is where they are limited least; AMD's challenge is that they seem to only build one architecture up and down their range, which tends to err more toward compute capability than consumer graphics rendering.

This sounds like a reasonable choice for their limited resources, but it does have the effect of their products being less efficient per transistor for gaming, and since they tend to clock lower, they also remain behind in gaming performance despite selling far larger and far more power hungry GPUs on similar process nodes.

 

[luminousone]

I would have to disagree, The Polaris line clearly has better utilization of memory bandwidth then Vega, It just has much less of it to throw around. You get better performance per GB/s with Polaris. In fact the most common Mining config for Vega involves undervolting and underclocking the GPU while overclocking the memory.

Vega 64 does fantastically well on certain games as well, Doom is a good example.

I am fairly certain that Vega has a memory stride size problem, namely its to large, so write combining hit rates are bad. A Vega 64 has 2 1024 bit memory channels, A Polaris or 10 series card has 8 to 12 32 bit memory channels, If I am writing a 32 bit pixel out to memory the probability of it occupying a memory channel is 100%, meaning I am not losing any overhead to having to read the memory line in and mask the original memory line and the new 32 bit pixel bytes together.

Vega 56/64, has a large 4MB L2 cache, this isn't a coincidence, I would bet you this is primary used as an eviction cache with the hope of being able to write combine as often as possible. writing 32 bits on a 1024 bit bus isn't remotely efficient, waiting for a bundle of writes to the same line and shooting them out at the same time, is a vast improvement over what the other options would be.

Most GPU's don't have a cache, I would bet you that the cache on the 7nm Vega gpu's ends up being larger, I also bet future versions of HBM end up including features to help with this(per quad bit masking/write enable bits) or something of the like.

 

[Jedibeeftrix]

but back to the topic in hand - will they or won't they re-release vega 10?
either:
1. nothing more than a new stepping of vega 10, with some optimised power circuitry
2. slight revision of vega 10 (ala Ryzen 2), with faster HMB2 and revised power circuitry
3. ta-da! fooled you; you thought we'd cancelled 12nm Vega redux, but we were talking about uarchs, not product sku's!
4. sneaky, we told you vega 20 was for data-centres, but it runs Crysis just fine too. see you in Q4 2018!
5. absolutely nothing, just continueing to drip-feed the current vega 10 product in the market, as already doing
6. none of the above

Nothing exciting at Computex, looks like it is option 5 after all.

 

[Pieter3dnow]

Nothing exciting at Computex, looks like it is option 5 after all.

AMD did promise that 7nm gpu would be there for the gamers. Now it is only waiting for how much there going for in the shops between launch price and availability that seems to be somewhat of a problem.

 

[Jedibeeftrix]

AMD did promise that 7nm gpu would be there for the gamers. Now it is only waiting for how much there going for in the shops between launch price and availability that seems to be somewhat of a problem.

i'm not sure that was a reference to anything more than Navi in Q2/Q3 2019.

 

[Pieter3dnow]

i'm not sure that was a reference to anything more than Navi in Q2/Q3 2019.

Yeah you are right a rather generic comment, I could have sworn it was related to 2018.