NVIDIA shows signs ... [2008 - 2017]

[Alexko]

I think I prefer his free analysis - much more entertaining....it is after all a satire site isn't it?

The schedule was obviously silly, and I don't know why Charlie expected the high end to go away in 2012, but he did get one thing right: low-end GPUs are dead. It took longer than a year or two, but it did happen. Obviously, any finite and ordered set of elements has a minimum, thus both AMD and NVIDIA still have "low-end" GPUs, but they're bigger (in mm²) and have wider buses than the low-end GPUs of yesteryear. If I'm not mistaken, NVIDIA's smallest desktop GPU for this generation is GM107, which is almost 150mm² and has a 128-bit bus.

Beyond this, there are adverse trends for discrete GPUs. Attach rates are falling, in part due to the market's shift towards ever more mobile solutions, thus towards more constrained form factors, but that's not all. If you look at Intel's APUs for example (I'm using AMD's terminology here to refer to mainstream CPUs with integrated graphics) they have been "stuck" at 4 cores since Sandy Bridge. So that's Sandy Brdige, Ivy Bridge, Haswell, Broadwell, and almost certainly Skylake. There's nothing to suggest that the latter's replacement will have more CPU cores. In the meantime, processes have evolved from 32nm to 22nm and then 14nm. Once again, it's unlikely that future 10nm APUs will exceed 4 cores. The inevitable consequence of this is that the share of silicon real estate devoted to graphics increases at every generation, especially with shrinks. Meanwhile, discrete GPUs do not benefit from such increases in available space.

Since the first APUs and, for the most part, up to this day, the major problem has been bandwidth. For an equal amount of graphics silicon and a similar power budget, discrete GPUs always do better because they have access to fast GDDR5, and sometimes to wider buses; plus they don't have to share their bandwidth with CPU cores. But I think 2016/2017 will be the real test for the future viability of discrete GPUs as a mainstream product. This is when I expect stacked memory to become viable and affordable enough to be featured on mainstream APUs (be it HBM, stacked eDRAM, HMC or whatever) and all of a sudden, the bandwidth advantage of discrete GPUs should vanish. At that point, even ~150mm² GPUs will be difficult to justify, and it's only at 200mm² or above that they will start making some sense.

So, what does that mean for NVIDIA? If they can maintain their current (and huge) market share, and if they still dominate the professional market with Quadros and Teslas, they should be fine. But if their market share falls back to the 40~60% range and AMD keeps gaining share in the professional market, then Tegra had better start making money or they will indeed be in trouble. They have deep coffers, and if this is how they die, it will be a long, slow death, but I think the risk is very real.

 

[silent_guy]

What you're saying is: "Charlie may have been completely, laughably wrong about his grandiose statements, but at least he noticed that Intel iGPU are getting a larger part of the die with each new generation."

Quite the analyst indeed.

 

[Alexko]

Well, yeah, but my post wasn't really about Charlie; he was just the starting point.

 

[silent_guy]

It's one thing to point out general trends in the silicon industry (the end of Moore's law, the superiority of Intel), it's another to point to Nvidia as being the one who suffers from it more than anyone else.

I have no idea what will happen 10 years from now, the end of Moore's law makes that too hard. What I do know is that HBM will not make discrete GPUs niche just 2 years from now, 2017.

I don't understand the thinking of those who claim than bandwidth is the key issue that seems to hold back everything else: it's not. Bandwidth doesn't do anything transformative. It just feeds the units that do the transformation. Make BW infinite on a Hawaii or gm200 or Fiji and watch the GPU performance skyrocket into the stratosphere... not.
Sure, performance will go up, and that's great, but it will only move the bottleneck to other units.

As pointed out on TechReport/hardware.fr back in the day, the bandwidth bottleneck for the lower 3.5GB of a GTX970 is limited by the SM's ability to feed it, not by the raw BW of the MC itself.

In 2016, we'll see Pascal. These days, flagship GPUs are 50% faster than the previous one. That will make it 2.25 faster than a GTX980. Which will be just enough to play all games at 4K in ultra, forget about surround or anything. That chip will easily make it through 2017 as well. A 150mm2 will only be a fraction of that performance. That idea that that will be difficult to justify is ridiculous: it will not even match a GTX980. An APU will be much worse.

 

[dbz]

The schedule was obviously silly, and I don't know why Charlie expected the high end to go away in 2012, but he did get one thing right: low-end GPUs are dead. It took longer than a year or two, but it did happen. Obviously, any finite and ordered set of elements has a minimum, thus both AMD and NVIDIA still have "low-end" GPUs, but they're bigger (in mm²) and have wider buses than the low-end GPUs of yesteryear.

In the era of the APU (2010 - ), the minimum discrete GPU size for Nvidia has increased from 57mm²/64-bit (GT218) to 79mm²/64-bit (GK208 and GF119 as found on entry level systems). I'd presume the Maxwell low end would be GM108 (presently used for mobile), but has only shown up as a mobile SKU (840M)

If I'm not mistaken, NVIDIA's smallest desktop GPU for this generation is GM107, which is almost 150mm² and has a 128-bit bus.

If you're talking about Maxwell, then I guess you'd have to wait until the rest of the series is fleshed out, but given that a 118mm² GK107 and 90mm² Oland kick AMD's strongest R7 APU senseless, the present/ short term future isn't as dire as it would seem for some.

At that point, even ~150mm² GPUs will be difficult to justify, and it's only at 200mm² or above that they will start making some sense.

Undoubtedly, and the "good enough" requirement obviously plays a significant part, but isn't also likely that PC gaming might see a levelling out of falling dGPU sales if APU/IGP fail to keep pace with the lowest dGPU options if gaming graphics requirements/screen resolution raise the bar faster than APUs evolve- at least in the medium term. It seems like Gaming Evolved and TWIMTBP both have a vested interest in keeping the discrete market afloat. Not so long ago 1280x1024 and then 1366x768 used to overwhelming choices for gaming, today 1980x1080 is de rigueur with 2560x1440 starting to gain traction in the mainstream. So, is there a cut off point in PC gaming where APUs simply don't cut it?
There seems to have already been a decrease in the dGPU sales slide in the wake of the initial APU offerings from both Intel and AMD

 

[swaaye]

nevermind.

 

[Alexko]

It's one thing to point out general trends in the silicon industry (the end of Moore's law, the superiority of Intel), it's another to point to Nvidia as being the one who suffers from it more than anyone else.

I have no idea what will happen 10 years from now, the end of Moore's law makes that too hard. What I do know is that HBM will not make discrete GPUs niche just 2 years from now, 2017.

I don't understand the thinking of those who claim than bandwidth is the key issue that seems to hold back everything else: it's not. Bandwidth doesn't do anything transformative. It just feeds the units that do the transformation. Make BW infinite on a Hawaii or gm200 or Fiji and watch the GPU performance skyrocket into the stratosphere... not.
Sure, performance will go up, and that's great, but it will only move the bottleneck to other units.

As pointed out on TechReport/hardware.fr back in the day, the bandwidth bottleneck for the lower 3.5GB of a GTX970 is limited by the SM's ability to feed it, not by the raw BW of the MC itself.

In 2016, we'll see Pascal. These days, flagship GPUs are 50% faster than the previous one. That will make it 2.25 faster than a GTX980. Which will be just enough to play all games at 4K in ultra, forget about surround or anything. That chip will easily make it through 2017 as well. A 150mm2 will only be a fraction of that performance. That idea that that will be difficult to justify is ridiculous: it will not even match a GTX980. An APU will be much worse.

That's not what I'm arguing.

But if you look at the fastest integrated GPU currently available, namely that of Kaveri, you may notice that it's identical to that of the Radeon HD 7750, or R7 250 if you prefer. In practice, however, the latter is massively faster, and the reason for it is simple: the APU is limited to shared DDR3 on a 128-bit bus when the discrete GPU has exclusive access to much faster GDDR5 on a bus of the same size.

So what happens if you throw HBM into the mix? Simple: the APU catches up to the discrete GPU. In that scenario, does that specific discrete GPU still make any sense? Perhaps it does if you have an older APU with slower graphics, but on a new build it would be pointless. This does not mean, of course, that bigger GPUs will go away, or that something like Hawaii would benefit nearly as much from HBM, for the simple reason that Hawaii is not significantly bandwidth-limited.

The R7 250 is based on Cape Verde, a 123mm² GPU. Granted, 20% of its compute units (which may amount to 10% of the die) are disabled. That's about as high as current APUs can reach.

The argument I'm making can be boiled down to this: if say, two years from now, you're buying an APU with about 150mm² of graphics silicon in it, why would you buy a 150mm² discrete GPU on top of it? If you have need of it, then you certainly would still buy a 250mm² GPU or bigger, but a 150mm² one would be pointless. Now, I don't know how NVIDIA will handle this transition, but since they are the only company whose main business is discrete GPUs, they are the most exposed to this particular risk.

 

[pharma]

you're buying an APU with about 150mm² of graphics silicon in it, why would you buy a 150mm² discrete GPU on top of it? If you have need of it, then you certainly would still buy a 250mm² GPU or bigger, but a 150mm² one would be pointless..

Are there any bandwidth limitations for an APU that might not be present for a GPU? When Pascal is launched in 2016 it is supposed to support 1 TB/s bandwidth in 3D Stacked memory ...

 

[Silent_Buddha]

Are there any bandwidth limitations for an APU that might not be present for a GPU? When Pascal is launched in 2016 it is supposed to support 1 TB/s bandwidth in 3D Stacked memory ...

For the market segment he's talking about, it's unlikely to matter. Discrete cards costing ~100-150 USD are unlikely to need that much bandwidth for many generations to come.

His point is that once APU's get access to enough bandwidth where the integrated GPU to memory bandwidth (even with CPU contention) allows it to be not be memory bandwidth starved, what would be the point of the discrete card of similar capabilities? Right now discrete still wins handily if capability are similar because any integrated GPU that is reasonably powerful will be massively bandwidth starved.

That is unlikely to happen with the enthusiast or high end cards. But it should allow integrated to overtake budget (making them worthless except for legacy systems) and start seriously encroaching on midrange (putting their value into question).

Regards,
SB

 

[Kaotik]

Are there any bandwidth limitations for an APU that might not be present for a GPU? When Pascal is launched in 2016 it is supposed to support 1 TB/s bandwidth in 3D Stacked memory ...

Fiji is coming in near future (between tomorrow and computex by the looks of it) and sports the same HBM-memories Pascal will use, unless Pascal will use 2nd gen HBM which might or might not be available by the time Pascal should launch

 

[dbz]

The argument I'm making can be boiled down to this: if say, two years from now, you're buying an APU with about 150mm² of graphics silicon in it, why would you buy a 150mm² discrete GPU on top of it?

If performance-per-watt were equal you would not. Is this likely to be the case?
The current ~150mm² GPUs ( GM107/GTX 750Ti and Bonaire/R7 260X) are 65-75W boards. Probably safe to assume that a process node shrink won't do wonders to reduce the envelope - it hasn't before. RV730 was a 60W GPU (HD 4670), Juniper at 85W for 166mm² scales at the same rate as Bonaire (W/mm²). Can an APU devote that much power to its graphics and remain competitive as a package even deducting a few watts for HBM rather than DDR?

 

[Alexko]

If performance-per-watt were equal you would not. Is this likely to be the case?
The current ~150mm² GPUs ( GM107/GTX 750Ti and Bonaire/R7 260X) are 65-75W boards. Probably safe to assume that a process node shrink won't do wonders to reduce the envelope - it hasn't before. RV730 was a 60W GPU (HD 4670), Juniper at 85W for 166mm² scales at the same rate as Bonaire (W/mm²). Can an APU devote that much power to its graphics and remain competitive as a package even deducting a few watts for HBM rather than DDR?

Why not? What matters is platform power, not the specific envelope of individual chips. There's no law of physics that says APUs must remain below 65W or whatever. If you have a 65W APU and a 100W discrete graphics card, you're not better off than you would be with a 165W APU. If anything, having everything on the same die with fast, low-power communication and a few common components is likely to save you a few watts at a given performance point.

 

[A1xLLcqAgt0qc2RyMz0y]

Why not? What matters is platform power, not the specific envelope of individual chips. There's no law of physics that says APUs must remain below 65W or whatever. If you have a 65W APU and a 100W discrete graphics card, you're not better off than you would be with a 165W APU. If anything, having everything on the same die with fast, low-power communication and a few common components is likely to save you a few watts at a given performance point.

Sorry but no way would I ever buy a 165 watt APU.

I would rather have the split 65 watt CPU (internal graphics disabled) and a discrete 100 watt GPU that can be upgraded over time to a more powerful one.

 

[swaaye]

The Sony and MS consoles essentially have 165W APUs and they sell pretty well and also demonstrate the many advantages to high integration over discrete components. I can't imagine it happening on PC though simply because the increased cost of such an APU and its supporting platform needs like 256-bit GDDR5 would be wasteful when most people don't game. There would be no benefit for such people.

 

[Silent_Buddha]

More realistic is a 95-100 watt APU that features a 15-30 watt CPU core and a 65-85 watt GPU core. There's already CPU's by Intel (haven't checked AMD side in a while) that hit 95-135 watt TDPs, so that'd still be less than what is used in many people's computers.

Minus the power requirements for the board and memory that likely gets you into the ~100 watt class of GPUs. Which is getting into midrange territory. With games coded for DX12, you'd likely do as well (probably better) as you would with a modern i3 +100 watt discrete GPU in DX11 and possibly DX12. Rendering lower cost discrete cards of little use other than for legacy system and midrange questionable for use in anything other than legacy systems. Once you hit the high end and enthusiast level's thats where discrete will hold its own for quite a while.

For example, it isn't out of the question to see something more powerful than the PS4/XBO SOCs come to the PC. And that is already solidly in midrange territory. The only thing holding something like that back on PC compared to console is the relatively thick DX11. DX12 should go a long way to sorting that out for games coded for DX12. Just replace the GDDR5 in PS4 and ESRAM+DDR3 in XBO with HBM or some other high speed memory technology.

Nvidia can see this coming which is why they are working hard on maximizing their margins and attempting to diversify their portfolio. Tegra is their attempt to have an option to the integrated options that will be coming from Intel and AMD. And their hope is that when those become performant enough to displace low end and potentially midrange GPUs that Tegra (and Tegra devices like Shield) will be able to replace that lost revenue.

Regards,
SB

 

[Silent_Buddha]

The Sony and MS consoles essentially have 165W APUs and they sell pretty well and also demonstrate the many advantages to high integration over discrete components. I can't imagine it happening on PC though simply because the increased cost of such an APU would be wasteful when most people don't game. There would be no benefit for such people.

They are closer to 100 watt APUs. The XBO for example attempts to stay below 100 watts at all times. Which is borne out by total system power when gaming only slightly exceeding 100 watts. (http://www.extremetech.com/gaming/1...to-sony-advantage-and-future-efficiency-gains ).

Regards,
SB

 

[swaaye]

I imagine XBO and PS4 APUs benefit from those tiny Jaguar cores being low power. Most of the power is surely going into the GPU.

I'm sure that HBM will kick APU performance up a notch just as APUs in general have done so compared to previous chipset based graphics. But HBM will need to be cheap for it to happen in volume, and APU die size also needs to be kept reasonable. I don't know if a premium product would sell well. Intel's Iris graphics is very expensive for example, making it poor value, and only common in some Apples.

 

[Alexko]

Sorry but no way would I ever buy a 165 watt APU.

I would rather have the split 65 watt CPU (internal graphics disabled) and a discrete 100 watt GPU that can be upgraded over time to a more powerful one.

Yes, for DIY desktop computers, upgradability is a big factor. For OEM machines and especially for laptops, it doesn't matter.

 

[xpea]

Nvidia can see this coming which is why they are working hard on maximizing their margins and attempting to diversify their portfolio. Tegra is their attempt to have an option to the integrated options that will be coming from Intel and AMD. And their hope is that when those become performant enough to displace low end and potentially midrange GPUs that Tegra (and Tegra devices like Shield) will be able to replace that lost revenue.

Regards,
SB

I think NV is betting much more on research/scientific/enterprise/high margin solutions like GRID, TESLA, QUADRO than Tegra to keep afloat in the distant future

 

[madyasiwi]

NV tops R&D/sales ratio among top 10 chip R&D spenders in 2014, despite also the least sales compared to the rest of the group.

Source.