Q4 2016 GPU shipments totalled 100 million, says JPR

[watercooled]

It's good that you help other people to afford new cards by buying their old ones off them. I don't think people are idiots for buying cards though - the leap in speeds we see each generation is quite impressive. Only a few years back the power you get in something costing under £200 today used to cost closer to £1000.

Case in point, an RX 480 4GB - released in June 2016 for $199 - is generally a little faster than a GTX 780, released in May 2013 for $649. The 1080 Founder's Edition, which launched at $699, is at least twice as fast as a GTX 780, and in some cases gets close to three times as fast.

Unlike CPUs, which have had relatively small generational performance increments recently, GPUs have continued to increase performance massively (remember, 780 - 1080 is actually only 2 generations).

EDIT: in fact, the 8GB RX 480 is pretty much level with a 780 Ti, which launched 6 months later for $699...

Improved performance at a price point is expected, otherwise why bother buying the newer cards? The problem is the value improvements per generation really aren't what they used to be; what we would have historically called mid-range GPUs (based on e.g. die size) are being sold as increasingly-expensive high-end cards, with the previous high-end cards being sold as some new e-peen tier for extortionate amounts of money. Some use the slowing of Moore's law as a complete excuse for this, it isn't. While the $/transistor seemingly isn't dropping as fast as it used to due to the use of things like multiple patterning, increasing mask costs, etc, it doesn't come close to justifying the jump in price we've seen in some market segments. This is apparent from the increasing profit margins published - in one breath, Nvidia is claiming that fab costs are squeezing them, and in another they're boasting increased margins - can't have it both ways!

Nvidia just keep pushing prices higher and higher every generation to see what they can get away with, so from a consumer perspective it's perfectly reasonable to be critical of it IMHO. It's obvious the effect competition has on the market when you look at prices of cards from both companies within AMD's catchment. Meanwhile the Titan has even lost the excuse of being a Geforce-Pro mid-ground - the features that defined the original Titan cards have evaporated from all of the recent cards, and it's gotten even more expensive!

Part of the reason we're still seeing good performance gains in GPUs compared to CPUs is down to the nature of the tasks they process - graphics is an 'embarrassingly parallel' task with extremely good parallel scaling, so there is still low-hanging fruit in the form of adding more cores to the GPU die and you'll get a very good return on that. That's not to take away from the huge amount of engineering that goes into GPU design, even something as 'simple' as 'bolting on a few more cores' isn't straightforward as you then have to figure out clock and power distribution across a larger die, ensure the processor fabric and caches will not become bottlenecks, etc. And I think it was a blog post from a PowerVR graphics engineer that said there are still lots of new tricks being discovered for graphics processing to e.g. make better use of the resources you do have.

With CPUs, it's harder to make good use of that extra transistor budget. You only have to look at the size of CPUs compared to GPUs - Intel's desktop Skylake processors are barely over 120mm2 vs upwards of 500mm2 for large GPUs - and most of that CPU isn't CPU space at all, it's things like integrated graphics, Northbridge and caches. I estimated Skylake's individual core size without L3 cache to be something like 8mm2, and Skylake is a *big* CPU core.

[scaryjim]

... I estimated Skylake's individual core size without L3 cache to be something like 8mm2 ....

I believe a recent AMD slide comparing Ryzen to Skylake had Ryzen's 4 core complex @ 44mm², vs something like 49mm² for Skylake, so you're a little under the figures there.

But here's the thing - what level of generational performance increase do you expect?

As I said above, a 1080 is more than twice as fast as a 780 Ti. That's more than doubling the performance @ a given price point in 2 generations. And that hasn't just been done by going wide and throwing shaders at it - the 1080 has less shaders than the 780 Ti. It's been done by engineering a core that is both more efficient and able to run at much higher clock speeds.

The only recent time I can think of a similar performance delta between generations would be the transition through ATI/AMD's 2000/3000/4000 series cards, but that was down to them seriously messing up the original design and having to do a significant reworking to get anywhere near competitive. The 780 Ti wasn't a bad card. Nor was the 780. But in 2 generations nvidia managed to obliterate both of them with a card sporting a similar number of shaders, and retailing in the same price range.

Same price range. More than twice the performance. Seriously, what more do you want?

[nobodyspecial]

...and we're all idiots for buying them. When a company, such as nVidia, make 2 BILLION profit they really are over pricing their parts. We lap it up and then return to spend EVEN more cash. But, obviously, that's business.

P.s. I own an eBay purchased 980.

This is just silly. It's taken then a decade to get back to 800mil+ in income. A HUGE leap this TTM but that's it. Get a better job if you can't afford your toys. People pay for GREAT products (no matter who they are from), as shown by Q reports. If we don't think a product is worth it, it doesn't sell. See AMD cpus gpus vs. Intel/NV Q reports. AMD should never have paid 3x for ATI that it was worth, and compounding the problem by chasing consoles (twice now) and apu low-end crap (consoles/apu both against what Dirk Meyer wanted and was fired for in 2011 - KING CPU/GPU which they seem to be chasing 5yrs later...ROFL). That money should have been available for R&D and then they wouldn't have been in the mess they've been in for a decade.

Management is AMD's problem, not the employees. They have the ability, if management would just get the heck out of the way. NV/Intel Q reports success comes from ALWAYS chasing the high-end first. That is where most of the profits (and ridiculous margins) are made. You can't get rich off poor people, generally speaking.

[watercooled]

I believe a recent AMD slide comparing Ryzen to Skylake had Ryzen's 4 core complex @ 44mm², vs something like 49mm² for Skylake, so you're a little under the figures there.

Like I said, individual core size, not including L3 cache (core complex includes L3). RWT put it at something like 8.5mm2 IIRC.

But here's the thing - what level of generational performance increase do you expect?

I don't really expect anything in particular, and that's not my point. My point is that generational performance improvements at a given price is nothing unusual in the semiconductor market. There weren't all that many GPUs being sold for $1000 until recently, but that tier of GPUs has always existed - Nvidia had a bit of a pause with the 600 series, bumping the midrange up to high-end pricing, then portraying the (typically sized for a normal high-end GPU for Nvidia) GK110 series as a new one above that, and of course charging considerably more for it. In one generation, Nvidia went from charging $500 for their biggest, top-binned die to $1000. And that pricing structure has stuck.

As I said above, a 1080 is more than twice as fast as a 780 Ti. That's more than doubling the performance @ a given price point in 2 generations. And that hasn't just been done by going wide and throwing shaders at it - the 1080 has less shaders than the 780 Ti. It's been done by engineering a core that is both more efficient and able to run at much higher clock speeds.

Read over my comment again - I tried to make it clear I wasn't knocking the engineering achievements of GPUs, but to further clarify, I was demonstrating that there are still areas where you can make good use of extra transistors in GPUs i.e. to a much greater extent than CPUs, and this is 'part of' the reason performance has continued to scale so well. I took your original comment (perhaps mistakenly) as a dig at CPU progress, and I was just pointing out that it's become increasingly more challenging to eke extra performance out of CPU cores regardless of transistor budget, whilst GPUs can still scale performance very well given more transistors to play with. Or put another way, you *can* still improve performance by 'throwing shaders at it', I wasn't at all claiming that's all that's been happening. The improvements such as those seen in Pascal are exactly the sort of thing I meant by 'new tricks' - GPU architects haven't been under the same sort of constraints as CPU architects to wring performance out at an architectural level, while they've been able to get huge gains by making good use of Moore's law, so it makes sense that there are some core efficiency improvements not yet implemented.

Same price range. More than twice the performance. Seriously, what more do you want?

Well, since you're asking, why not a Titan XP for the same price? It's using the same tier of GPU as a 780Ti after all? That's pretty much the point I'm getting at - you can easily make a point of "well it's x% faster than last year, and all for the same price, why complain?", but when the only thing stopping it from being x+y% faster, is a company setting the price arbitrarily higher to improve nothing but their own profit margins, why exactly would I cheer that on? Sure, they're a business looking to make money - doesn't mean I have to start a fan club for their business practices.

BTW, don't confuse generations and years - e.g. the 2000-3000-4000 had 6 and 7 months between them (back when GPUs also used half-nodes from the foundries). The time between 780Ti and 1080 is about 30 months.

[scaryjim]

Like I said, individual core size, not including L3 cache (core complex includes L3). RWT put it at something like 8.5mm2 IIRC. ....

http://hexus.net/tech/news/cpu/10227...ompact-intels/

Core complex, L2 cache and L3 cache are all listed separately, so I was interpreting that as meaning the measurements are exclusive of each other - I suppose if the core measure is inclusive of all caches that'd work out about to 7.5mm² like you said! Wish people would label/explain their tables better *cough*Hexus*cough*

If the core figures in that table *are* inclusive of L3 cache that'll make 8-core Ryzen tiny ... the cores and L3 cache will only take up ~ 88mm²! Although they've got the IMC, PCIe complex and southbridge to cram in there too, of course...

To get back on topic though (this thread is about GPU sales, after all ), I guess we can all wish for the moon on a stick. Yes, nvidia probably could charge $200 less for the 1080 and still make a profit - their margins are ever widening. But they're selling many 1080s at $699 - so why would they?

It's not like we're in Intel territory here: http://www.anandtech.com/bench/product/287?vs=1543 is a comparison of the 2600k and 6700k - 4 and a half years apart, roughly the same price, generally no more than 50% faster, and never as much as 100% faster. Even the thermal design points are almost identical.

I'm not arguing that nvidia aren't, to a certain extent, profiteering with their current generation chips. But at least they're backing it up with genuine advances in performance and energy efficiency, including at the top end. You're getting genuine generational performance improvements at a given price point. You might want more, but I think it's rather naive to expect more...

[watercooled]

Oh yeah don't get me wrong, I'm really glad to see we're still seeing very good performance improvements per GPU generation, just the optimism is somewhat tempered by the profiteering we're seeing lately. And like you say I don't really expect more, especially given the lack of competition at the high end, but it's somewhat irritating to think the only thing stopping better progress is business, not technology.

CPU core sizes are surprisingly small now, so much so I think the likes of Intel and AMD almost have to use up some space with GPUs, Northbridge, etc in order to make the die large enough to accommodate the number of pads necessary for IO.

[watercooled]

If the core figures in that table *are* inclusive of L3 cache that'll make 8-core Ryzen tiny ... the cores and L3 cache will only take up ~ 88mm²! Although they've got the IMC, PCIe complex and southbridge to cram in there too, of course...

The uncore does look to take up a significant amount of die area:


It's probably not that unlike the Intel workstation parts - the uncore and last level caches seem to take up relatively more space on this sort of CPU vs the desktop/mobile ones:


Having said that, the 10C Broadwell-E die is 'only' 246.3mm2, really not that big for a workstation CPU, and based on pure guesswork looking at the two CPUs (I might have a go at pixel-counting and estimating it properly later) we're probably still looking at <200mm2 for Ryzen. TBH that's still pretty small! And another thing, is Ryzen has dual channel IMC vs the (pretty much unnecessary for desktop) quad channel one found on Broadwell-E - that could save some space too.