Support NeoGAF

[tuxfool]

According to the Anandtech article, the major expenditure in the extra transistor budget went into increasing clock speeds. I can't really say that they were all that successful.

One interesting observation I guess I could make is that speculation that the tiled rasterizer being the reason Nvidia manages to keep their power budgets under control is a somewhat simplistic explanation, clearly there is more to it than that.

 

[Bolivar687]

They have us over a barrel. They had us over a barrel with Pascal that's why they jacked the prices up another $50 and we still bought them in droves because, well, what are we going to buy? A Fury X? Competitive with 980 Ti performance and only 2/3 the VRAM! But it was $50 cheaper than the 1080! Woo! AMD! AMD!

They cut the price of the Fury X during Pascal's launch, so that it was cheaper than what the 1070 was selling for at the time by a decent margin. It also outperformed the 1070 in a number of new games that were coming out around then like DOOM and Deus Ex.

Nvidia apologists talking about year old architectures and pricing should be taken with a grain of salt.

 

[Sir Abacus]

They cut the price of the Fury X during Pascal's launch, so that it was cheaper than what the 1070 was selling for at the time by a decent margin. It also outperformed the 1070 in a number of new games that were coming out around then like DOOM and Deus Ex.

And if the 1070 was the top of the line I'd be less pissed. The complaint is we have nothing, nothing on Team Red to keep the top end in line and my prediction is the prices are going to be crazy.

 

[saskuatch]

Now I am debating about going into Canada computers at lunch and buying an msi 1080ti gaming x. Good idea bad idea? Should I wait till holiday season?

 

[Thraktor]

According to the Anandtech article, the major expenditure in the extra transistor budget went into increasing clock speeds. I can't really say that they were all that successful.

Going from ~1.3GHz on Polaris to ~1.6GHz on Vega on the same manufacturing process is a decent enough jump, even if they're still well behind Nvidia on clocks.

Realistically, though, most of the transistor budget likely went to that double-rate FP16 and quad-rate INT8 support. That kind of extra compute capability in every ALU adds up.

One interesting observation I guess I could make is that speculation that the tiled rasterizer being the reason Nvidia manages to keep their power budgets under control is a somewhat simplistic explanation, clearly there is more to it than that.

It helps, but it's likely a relatively small part of the picture. The main difference is that AMD are factory clocking their cards very close to their clock ceiling, while Nvidia aren't. Power efficiency goes to hell when you get that close to the chip's limits, which you can see when you look at the TDP jump going from the Vega 56 to the Vega 64 Air to the Vega 64 Liquid. The Vega Nano will actually probably be a pretty power efficient chip, but will have to clock down quite a bit from the other models to achieve it.

 

[ethomaz]

GCN is probably showing its age with how inefficient it scales at high CU/SPs and clocks.

AMD sure needs something new for GPUs like they did with Zen.

 

[DieH@rd]

Etherum mining has to be done on the GPUs [no ASICs], and even nvidia hardware is this time viable for it. Plus price of Bitcoin remains to be high.
https://qz.com/1039809/amd-shares-a...-boeing-747s-to-ship-graphics-cards-to-mines/

So no matter what, I expect that prices will continue to be wild for considerable time for everything between midrange [RX460] and top of the line [1080Ti].

 

[tuxfool]

Going from ~1.3GHz on Polaris to ~1.6GHz on Vega on the same manufacturing process is a decent enough jump, even if they're still well behind Nvidia on clocks.

Realistically, though, most of the transistor budget likely went to that double-rate FP16 and quad-rate INT8 support. That kind of extra compute capability in every ALU adds up.

Fishing out the quote here:

Talking to AMD’s engineers, what especially surprised me is where the bulk of those transistors went; the single largest consumer of the additional 3.9B transistors was spent on designing the chip to clock much higher than Fiji. Vega 10 can reach 1.7GHz, whereas Fiji couldn’t do much more than 1.05GHz. Additional transistors are needed to add pipeline stages at various points or build in latency hiding mechanisms, as electrons can only move so far on a single clock cycle; this is something we’ve seen in NVIDIA’s Pascal, not to mention countless CPU designs. Still, what it means is that those 3.9B transistors are serving a very important performance purpose: allowing AMD to clock the card high enough to see significant performance gains over Fiji.

Suggests otherwise. Doubtless those are big consumers, but the quote is quite explicit.

 

[Thraktor]

Fishing out the quote here:



Suggests otherwise. Doubtless those are big consumers, but the quote is quite explicit.

Both are intertwined, though, they had to redesign the ALUs for both purposes at the same time, and at the launch of the gaming cards they're more likely to tout the improvement which benefits games than the one that doesn't.

 

[icecold1983]

Going from ~1.3GHz on Polaris to ~1.6GHz on Vega on the same manufacturing process is a decent enough jump, even if they're still well behind Nvidia on clocks.

Realistically, though, most of the transistor budget likely went to that double-rate FP16 and quad-rate INT8 support. That kind of extra compute capability in every ALU adds up.



It helps, but it's likely a relatively small part of the picture. The main difference is that AMD are factory clocking their cards very close to their clock ceiling, while Nvidia aren't. Power efficiency goes to hell when you get that close to the chip's limits, which you can see when you look at the TDP jump going from the Vega 56 to the Vega 64 Air to the Vega 64 Liquid. The Vega Nano will actually probably be a pretty power efficient chip, but will have to clock down quite a bit from the other models to achieve it.

its highly unlikely to be fp16. the chip has 45MB of sram. thats one of the big contributors

 

[llien]

The first Gsync module was Jan '14. ROG Swift was August '14. The first Freesync monitors didn't come onto the market until April '15 and they was garbage next to a Swift with ghosting all over the place as Benq and LG screwed the pooch on overdrive timings.

So yeah, while it was only months behind the gold standard Gsync monitor, Freesync was AMD's trademark late to the party, worse, but it was at least cheap.

I understand your point, but still think it is unfair to AMD to say they saw GSync and did nothing. In fact, on the opposite, they responded very quickly.

I'm not sure why people bring up ghosting with adaptive sync (different things), does it somehow only affect adaptive sync monitors? Most of my life I've gamed on monitors with no motion compensation whatsoever, can't complain about my experience.

A. 1070 being on a new process should be compared to 670, not 970. 670 was $400 at launch.

That's basically "no true Scotsman".

Generally speaking, Vega is performing exactly like a GCN(3) GPU would perform when being pushed to such TFlops.

Vega performs nearly exactly as Fury X, when downclocked to 1050Mhz: (so yeah)
http://www.gamersnexus.net/guides/2977-vega-fe-vs-fury-x-at-same-clocks-ipc

Why is GCN so bad at scaling up from midrange?

Well, except it isn't even an upscaling of the kind you mean (if you meant more CUs), it's Fiji (we see that in gamer nexus review) aimed at higher clocks (as kindly shared by Anand).

I wonder, wether they aimed for higher clocks or expected higher gains from higher clocks.

Suggests otherwise. Doubtless those are big consumers, but the quote is quite explicit.

Well, note how well it worked for nVidia (although it needed a smaller jump), which even reduced IPC (1080 vs 980Ti) if I am not mistaken.

I have expected Vega's IPC to be 10%-ish higher than Fiji's, based on number of architectural improvements they have promised. (in which case we would get 1080Ti's competitor). But alas... =(

 

[dgrdsv]

I'd like to think that but what are our other options? Buy an AMD card? Sit on our hands? They have us over a barrel. They had us over a barrel with Pascal that's why they jacked the prices up another $50 and we still bought them in droves because, well, what are we going to buy? A Fury X? Competitive with 980 Ti performance and only 2/3 the VRAM! But it was $50 cheaper than the 1080! Woo! AMD! AMD!

I'll believe it when I see a 1180 or whatever they call it at a "mere" $649.

Other option is obviously not buying a new card and just continue using your old one. We're at this point of the current console generation where even a midrange card provides enough power to be able to play all new AAA games at 60 fps in 1080p. Any new card coming to market must offer a very good value for such users to consider upgrading from their current setups. This is further exacerbated by the significant turmoil in the CPU/platform market brought on by Ryzen due to which I'd expect a lot of PC h/w buyers budgets to be re-allocated from GPUs to CPUs this and probably next year as well. They will have to fight for these money.

According to the Anandtech article, the major expenditure in the extra transistor budget went into increasing clock speeds. I can't really say that they were all that successful.

GP104 clock has risen by some 33% compared to GM200 while its transistor complexity when down 0.8B. This seems like a really odd thing to spend transistors on, especially when your architecture is mostly power limited already. Using these transistors to make the chip wider would probably net a better perf/watt result which in turn would allow them to up the clocks a bit and get better overall performance. All in all Vega looks like AMD was looking at Maxwell and then Pascal and saying "hey, we need to increase the clocks as well!" even though this doesn't really suit their architecture that well.

One interesting observation I guess I could make is that speculation that the tiled rasterizer being the reason Nvidia manages to keep their power budgets under control is a somewhat simplistic explanation, clearly there is more to it than that.

I don't know why people assumed that tiled rasterizer was the main reason of NV's perf/watt advantage - it wasn't. TBIR of Maxwell/Pascal mostly helps with memory bandwidth as can be clearly seen from what they achieve against competing GCN3/4 cards. NV's power benefits come from their SM architecture more than anything else.

Going from ~1.3GHz on Polaris to ~1.6GHz on Vega on the same manufacturing process is a decent enough jump, even if they're still well behind Nvidia on clocks.

Realistically, though, most of the transistor budget likely went to that double-rate FP16 and quad-rate INT8 support. That kind of extra compute capability in every ALU adds up.

All Pascal GP10x chips have INT8x4 support. They also support pretty much all of new Vega features, with FP16x being the only notable omission.

Now, I'm not the one who's going to argue that double rate FP16 does in fact increase a GPU complexity but I very much doubt that we're talking about +28% of transistors here. That's nearly a third of the chip which they could've spent on adding 50% more CUs for example which would result in +50%(ish) performance gain on the same clock above Fiji - this is a much better transistor investment than RPM which is given an estimate of +15% of performance in FM Serra demo made specifically to demonstrate it.

Thus it's highly unlikely that FP16x2 is the reason for such complexity gain.

 

[riflen]

I'm not sure why people bring up ghosting with adaptive sync (different things), does it somehow only affect adaptive sync monitors? Most of my life I've gamed on monitors with no motion compensation whatsoever, can't complain about my experience.

They bring it up because traditional displays only refresh at fixed rates and the overdrive is tuned for the main refresh rate the display is designed to support. Most people never change the rate (or most displays don't support anything other than 60 Hz) and never see significant ghosting at different rates.

Obviously with variable refresh, the rate can potentially be any value inside the supported range. This means the voltage to the pixels must be changed on-the-fly to avoid ghosting and provide the clearest picture under motion. Apparently G-Sync also predicts upcoming frame times to do a better job at this.
There were some bad early FreeSync displays that didn't achieve this properly and motion resolution was very poor.

 

[Sir Abacus]

Other option is obviously not buying a new card and just continue using your old one.

Hey man let's not say things we can't take back here.

I mean next you'll be telling me it's going to be ok to buy an 1170.

 

[FinKL]

So the gist I'm getting is that the value proposition of a Vega + Samsung Freesync monitor + games push it to a great value vs a 1080 + G-Sync monitor.

I would consider it if they dropped the 49" 32:16 CHG90 down hehe.

 

[Crash331]

Etherum mining has to be done on the GPUs [no ASICs], and even nvidia hardware is this time viable for it. Plus price of Bitcoin remains to be high.
https://qz.com/1039809/amd-shares-a...-boeing-747s-to-ship-graphics-cards-to-mines/

So no matter what, I expect that prices will continue to be wild for considerable time for everything between midrange [RX460] and top of the line [1080Ti].

Does anyone know if it would be possible to split the market? Like if Nvidia somehow made a hardware change that would make bitcoin mining inefficient without breaking gaming performance? Is that possible? And then selling "cryptocurrency" versions of their cards that are marketed towards miners?

 

[saskuatch]

Does anyone know if it would be possible to split the market? Like if Nvidia somehow made a hardware change that would make bitcoin mining inefficient without breaking gaming performance? Is that possible? And then selling "cryptocurrency" versions of their cards that are marketed towards miners?

Why would they spend money to do this

 

[Sir Abacus]

Does anyone know if it would be possible to split the market? Like if Nvidia somehow made a hardware change that would make bitcoin mining inefficient without breaking gaming performance? Is that possible? And then selling "cryptocurrency" versions of their cards that are marketed towards miners?

Not really. Ethereum for instance made their hashing algorithm ASIC resistant by requiring it to use a fuckton of memory and memory bandwidth. The working set for ethash is a freaking gigabyte. Guess what has that in spades?

Now if they didn't want to wreck the GPU market they would have had the hashing algorithm lean heavily on having working branch prediction but either they didn't expect mining to hit the GPU market so hard or they did and they just plain didn't care.

Either way it's infuriating competing against speculators in another market.

 

[franklyshankly]

Not really. Ethereum for instance made their hashing algorithm ASIC resistant by requiring it to use a fuckton of memory and memory bandwidth. The working set for ethash is a freaking gigabyte. Guess what has that in spades?

Now if they didn't want to wreck the GPU market they would have had the hashing algorithm lean heavily on having working branch prediction but either they didn't expect mining to hit the GPU market so hard or they did and they just plain didn't care.

Either way it's infuriating competing against speculators in another market.

Nvm you meant the Ethereum people. I guess I don't see why they would care either.

 

[ISee]

Here's a look at a Vega with a non-reference cooler:
https://www.pcper.com/news/Graphics...PUs-Headlined-ROG-STRIX-RX-Vega-64-OC-Edition

1677 MHz boost clock sounds great. Wonder if 1700MHz are possible.

Of topic, because of your avatar and name: I recently rewatched space beyond and above. It's still a good show (mostly).

 

[Durante]

They bring it up because traditional displays only refresh at fixed rates and the overdrive is tuned for the main refresh rate the display is designed to support. Most people never change the rate (or most displays don't support anything other than 60 Hz) and never see significant ghosting at different rates.

Obviously with variable refresh, the rate can potentially be any value inside the supported range. This means the voltage to the pixels must be changed on-the-fly to avoid ghosting and provide the clearest picture under motion. Apparently G-Sync also predicts upcoming frame times to do a better job at this.
There were some bad early FreeSync displays that didn't achieve this properly and motion resolution was very poor.

That's exactly it. Good explanation.

 

[Sir Abacus]

Nvm you meant the Ethereum people. I guess I don't see why they would care either.

Bitcoin has long been overtaken by ASIC mining. Ethereum and its "hash random shit in this gigabyte of data" is what's doing the damage right now.

 

[Sulik2]

Generally speaking, Vega is performing exactly like a GCN(3) GPU would perform when being pushed to such TFlops. So it's not really a question of "what went wrong" but more of a couple of different questions:

- Why is GCN so bad at scaling up from midrange? Both Fiji and now Vega are quite disappointing, and even the relatively successful Hawaii was a bit overboard with power consumption.

- Why all the architectural changes brought what is essentially a zero performance gain? Was it always their idea to not improve the existing code performance at all and concentrate on adding new features instead?

Ok that makes sense. Thank you for answering my question.

 

[Thraktor]

its highly unlikely to be fp16. the chip has 45MB of sram. thats one of the big contributors

Adding double-rate FP16 support is effectively the same as significantly expanding the instruction set of the ALUs. There has to be a transistor cost for that, and over 4096 ALUs that cost will add up. I'm certainly not saying it's all the INT8/FP16 support, but I'd wager a good chunk of it is.

The SRAM certainly takes up a lot of transistors (2.16 billion assuming 6T SRAM and ignoring any attached logic for caches, etc.), but I can't find any comparable figure for Fiji, so it's hard to say how much that's expanded.

GP104 clock has risen by some 33% compared to GM200 while its transistor complexity when down 0.8B.

GP104 is on a smaller process than GM200, which basically accounts for the entire clock difference (the SMs are virtually identical). It's also got fewer SMs than GM200, so it should use fewer transistors. Here we're seeing a decent clock speed jump over Polaris at the same manufacturing process, so there's clearly been a focus on clock speeds from AMD's part.

This seems like a really odd thing to spend transistors on, especially when your architecture is mostly power limited already. Using these transistors to make the chip wider would probably net a better perf/watt result which in turn would allow them to up the clocks a bit and get better overall performance. All in all Vega looks like AMD was looking at Maxwell and then Pascal and saying "hey, we need to increase the clocks as well!" even though this doesn't really suit their architecture that well.

I think it's a pretty sensible thing to spend transistors on, and I'd argue they should go further. At the same clock and similarly sized dies, Pascal and Polaris actually pretty comparable. The major difference is that the power curve on Pascal allows for much higher clocks at reasonable power draw, allowing Nvidia to use smaller die chips and use more conservative stock clocks/voltages, resulting in lower power draw and more overclocking headroom. For the past few generations, though (since the R9 290, anyway), AMD has had to crank up stock clock speeds and voltages to compete with Nvidia, trying to squeeze as much out of them as possible and accepting lousy power efficiency and no overclocking headroom in order to sell their cards for a decent price.

If AMD were taking Nvidia's more conservative approach to clocks & voltages we'd probably be looking at ~1GHz stock and ~1.2GHz boost clocks, for a relatively power efficient card with good overclocking capabilities. Of course even the Vega 64 would barely compete with the 1070 by that point, so they'd have to price the cards a lot lower to actually sell.

Vega is AMD's first notable architectural clock jump in a long time. And while it's still not allowing them to compete with Nvidia at similar die sizes yet, it's a lot better than nothing.

I don't know why people assumed that tiled rasterizer was the main reason of NV's perf/watt advantage - it wasn't. TBIR of Maxwell/Pascal mostly helps with memory bandwidth as can be clearly seen from what they achieve against competing GCN3/4 cards. NV's power benefits come from their SM architecture more than anything else.

It does save power (on-die cache accesses use less power than DDR accesses), but I'd imagine it's a pretty minuscule saving for a desktop GPU. It is a major reason behind the use of TBR in mobile GPUs, though, which is probably where people got the idea.

All Pascal GP10x chips have INT8x4 support. They also support pretty much all of new Vega features, with FP16x being the only notable omission.

I didn't know that about INT8. In any case, Nvidia's ALUs seem to have occupied a larger die area/more transistors than AMD's pre-Vega, with them now appearing roughly equivalent with Vega, so perhaps the INT8 support has something to do with it

Now, I'm not the one who's going to argue that double rate FP16 does in fact increase a GPU complexity but I very much doubt that we're talking about +28% of transistors here. That's nearly a third of the chip which they could've spent on adding 50% more CUs for example which would result in +50%(ish) performance gain on the same clock above Fiji - this is a much better transistor investment than RPM which is given an estimate of +15% of performance in FM Serra demo made specifically to demonstrate it.

Thus it's highly unlikely that FP16x2 is the reason for such complexity gain.

The thing is that I don't think the bolded is true. AMD's GCN cards don't show a linear performance increase with more ALUs at the same clocks (unlike Nvidia, whose Maxwell and Pascal cards, at least, do). There seems to be some kind of architectural bottleneck in AMD's current GPU tech which prevents them from efficiently using larger chips, and the fact that Vega doesn't show a performance per clock increase over Fiji suggests this is still the case. The GCN front-end doesn't seem to have been substantially changed since the first generation, and I suspect it's something they're going to have to look at soon.

Incidentally, back when Navi was first announced they used the word "scalability" to describe it, so it may be a hint that they're taking a TR/Epyc style multi-die solution for larger GPUs, which would almost certainly require a substantially re-designed front-end to work efficiently.

 

[llien]

They bring it up because traditional displays only refresh at fixed rates and the overdrive is tuned for the main refresh rate the display is designed to support. Most people never change the rate (or most displays don't support anything other than 60 Hz) and never see significant ghosting at different rates.

Obviously with variable refresh, the rate can potentially be any value inside the supported range. This means the voltage to the pixels must be changed on-the-fly to avoid ghosting and provide the clearest picture under motion. Apparently G-Sync also predicts upcoming frame times to do a better job at this.
There were some bad early FreeSync displays that didn't achieve this properly and motion resolution was very poor.

It makes perfect sense, thank you.

Vega is AMD's first notable architectural clock jump in a long time. And while it's still not allowing them to compete with Nvidia at similar die sizes yet, it's a lot better than nothing.

I might be missing something, do you mean high end cards?
Confused, why this would be called progress:

290x - 438mm^2 vs 780Ti's 561mm^2 (+28% die size for nVidia)
Fury X - 596mm^2 vs 980Ti's 601mm^2 (basically the same, not that big a difference at stock)

Now, on 14nm:

480(580) - 230mm^2 vs 1060's - 200mm^2 (+15% die area for AMD)
Vega 64 - 484mm^2 vs 1080 - 314mm^2 (+54% die size for AMD, ouch)

 

[snap]

So the gist I'm getting is that the value proposition of a Vega + Samsung Freesync monitor + games push it to a great value vs a 1080 + G-Sync monitor.

I would consider it if they dropped the 49" 32:16 CHG90 down hehe.

Those games are roughly $30 each on CDKeys, so they're not really a big factor.

I think the biggest question here is if the $200 discount applies to MSRP or street price of that Samsung monitor--if it's MSRP, street price is already $200 lower so it's not really a discount, and since the Ryzen 7 + MoBo discount is only $100 (and you're paying $100 more for the bundle over the GPU itself), it's unarguably a bad deal.

 

[thelastword]

https://youtu.be/51SsMsUXTt8



If there was any doubt, confirmation that Vega was rushed out for Q2. So that's good news for RX.

Funny that so many are not giving this post more feedback. The gamernexus breakdown is probably the most thorough and levelheaded one describing the specs and offers....on Vega...


So we're looking at 150-165 watts on Vega 56 and 220 watts on Vega 64. Whilst people are soley focused on the maximum power of these cards, the truth is the cards won't be using max power as average.....The vega cards was designed to run much higher clocks than GCN3 and they've accomplished that, so the power being used here is not out of focus or bad design as some people will try to prop, it's not that much higher than NV in any case....

For what it's worth TSBR was disabled and will be enabled for All Vega models including FE and RX, so we will see how that affects things. Apparently, these cards had to meet a deadline, so expect lots of driver level performance improvements as is typical for AMD GPU's, the fine wine aspect of AMD GPU's is defintely valid here despite my expectations of some nice/competitive benchmarks results for these cards on the 14th of August....

1700Mhz is not out of the ordinary for the Vega, so we can expect that and even faster clocks from AIB partners....AMD has the raw power in their GPU's, it's all a matter of developers using it, especially it's rapid packed math algos, it's enhanced pixel engine, reduced latency and much higher minimums on frames per second.


Speaking of higher minimums, I think that's genius tbh, which is what they're doing with HBCC (high bandwidth cache). Vega caters to much larger and bigger worlds, higher detailed worlds as a result so all this pop-in issues and loading issues and see-saw details will become something of the past or severly reduced. I think this is where Vega will shine over the competition. As I've said the vega features is what I'm most interested in and the vega features look very impressive to me.....

So back to minimums, Ryzen already has the edge over intel with better minimum framerates in many titles, therefore when you have Vega+HBCC and Ryzen working in tandem, this only means your minimums will improve even moreso with that combo. So it's not odd at all that AMD is offering bundle discounts on Ryzen + Vega + Freesync.

HBCC and rapid packed math seems to be the genius features here tbh, one reduces or eliminates the stutter/frame lag in high pressure, intense or high bandwidth moments in a game, whilst the other offers you the benefit of two shader instructions for the price of one, which in essence means higher fps. So if a game is developed with Vega in mind, I can easily see some significant performance issues with the competition.

Take BF1 for example, when all the explosions start ramping up, Vega will be able to maintain a high framerate with HBCC on. Nvidia may have slightly higher max framerates in current games, but if games have highs of 97fps on NV and lows of 46fps, whils Vega has highs of 93 fps and lows of 63, I think that will be a huge boon for the AMD GPU.....

Which brings me to the games that are in development which will use some of Vega's features. Far Cry 5 looks amazing to me, especially it's foliage and textures, theese guys did good work using compute on the PS4 and I can't wait to see what they do with this new technology. Prior to yesterday, I had no idea wolfenstein 2 was going to use Vulkan, or maybe I forgot, nonetheless it's always great to hear that. I wish there were much more vulkan games coming up that I knew of.......

So I'm interested in how developers leverage this new architecture and see what they come up with. I know Bethesda has a lot in the pipeline with AMD, maybe an expansion to Fallout 4 utilizing the vega features. Imagine a fallout game that does not stutter when loading large areas or new sections, A game like Deus EX MD tends to benefit greatly here too. I think AMD is tackling the right constraints for GPU's, I think too many are just hamstrung on max fps even over how smoothly a game plays....That needed to change for the better imo....Can't wait for benchmarks on the 14 to see current performance with the latest drivers......

 

[Thraktor]

I might be missing something, do you mean high end cards?
Confused, why this would be called progress:

290x - 438mm^2 vs 780Ti's 561mm^2 (+28% die size for nVidia)
Fury X - 596mm^2 vs 980Ti's 601mm^2 (basically the same, not that big a difference at stock)

Now, on 14nm:

480(580) - 230mm^2 vs 1060's - 200mm^2 (+15% die area for AMD)
Vega 64 - 484mm^2 vs 1080 - 314mm^2 (+54% die size for AMD, ouch)

Surely you agree it would be much worse if it only clocked to ~1.35GHz like Polaris?

As I said, I don't think the increase in die size has that much to do with the increased clocks. With the exception of TBR, almost all of the improvements in Vega seem to have been focussed on compute (where the card does seem to perform well). AMD don't have the R&D budget to put out pure compute cards like the GP100/GV100, and they want to start competing better in the workstation/server/neural net market, so we've got a die which has to do that and work as a gaming GPU. It's resulted in a larger die than if they went with 64CU Polaris, but at least we've got increased clocks to go along with those compute improvements.

 

[dgrdsv]

GP104 is on a smaller process than GM200, which basically accounts for the entire clock difference (the SMs are virtually identical). It's also got fewer SMs than GM200, so it should use fewer transistors. Here we're seeing a decent clock speed jump over Polaris at the same manufacturing process, so there's clearly been a focus on clock speeds from AMD's part.

You're not really suggesting that NV got +50% clock boost in Pascal simply due to a new process, are you? Because this isn't how it works these days, and there were efforts when building Pascal to make sure that it will clock higher - they've talked about it during Pascal's launch.

All in all Pascal has the exact same emphasis on higher clocks compared to Maxwell as Vega has compared to either Polaris or Fiji (do note that AMD themselves are constantly comparing Vega to Fiji instead of Polaris in their own official slides). But in case of Pascal this clock gain came without a significant complexity increase over Maxwell. So either this was already implemented in Maxwell (which was clocked significantly higher than Kepler mind you) or AMD did something weird in Vega.

I think it's a pretty sensible thing to spend transistors on, and I'd argue they should go further. At the same clock and similarly sized dies, Pascal and Polaris actually pretty comparable. The major difference is that the power curve on Pascal allows for much higher clocks at reasonable power draw, allowing Nvidia to use smaller die chips and use more conservative stock clocks/voltages, resulting in lower power draw and more overclocking headroom. For the past few generations, though (since the R9 290, anyway), AMD has had to crank up stock clock speeds and voltages to compete with Nvidia, trying to squeeze as much out of them as possible and accepting lousy power efficiency and no overclocking headroom in order to sell their cards for a decent price.

Go further where? They are launching a 345W card on 14th of August in consumer space. They need to spend all their resources on getting power consumption down in Navi - something which clearly wasn't done in Vega as it's actually much worse than Fiji compared to GM200 here.

I'd wager that them doubling P10 (88 CUs / basically 1.5x of Vega) and spending the effort on optimizing its power consumption would net them better results than increasing Fiji's clock as was done in Vega. Such chip even when running on Polaris clocks, without any power optimization, would already be faster.

GCN is not a high frequency architecture, and this clearly shows in Vega results where they had to spend a shitload of transistors and power to reach what is essentially last gen clocks for NV. If they really wanted to go high clocks they had to make an architectural update in vein of what NV did between Kepler and Maxwell.

If AMD were taking Nvidia's more conservative approach to clocks & voltages we'd probably be looking at ~1GHz stock and ~1.2GHz boost clocks, for a relatively power efficient card with good overclocking capabilities. Of course even the Vega 64 would barely compete with the 1070 by that point, so they'd have to price the cards a lot lower to actually sell.

Not sure what you mean by NV's "conservative" approach here considering that NV chips are clocked 33% higher than Vega and are running on 1.0V at this while Vega is 1.2V.

Vega is AMD's first notable architectural clock jump in a long time. And while it's still not allowing them to compete with Nvidia at similar die sizes yet, it's a lot better than nothing.

You seem to completely miss the point that this ~50% clocks increase cost them ~30% die size increase as well. This is not a good investment of die area IMO.

It does save power (on-die cache accesses use less power than DDR accesses), but I'd imagine it's a pretty minuscule saving for a desktop GPU. It is a major reason behind the use of TBR in mobile GPUs, though, which is probably where people got the idea.

Mobile GPUs are TBDRs mostly because of who are making them (Imagination) and less due to power savings of tiling. I don't know even, are Adrenos TBDR these days? Tegra sure aren't as they use the same form of tiling rasterization as desktop GeForces.

I didn't know that about INT8. In any case, Nvidia's ALUs seem to have occupied a larger die area/more transistors than AMD's pre-Vega, with them now appearing roughly equivalent with Vega, so perhaps the INT8 support has something to do with it

Unlikely as ALUs themselves do not change much for any type of packed math support. What is changing are the surrounding storage and control logic.

The thing is that I don't think the bolded is true. AMD's GCN cards don't show a linear performance increase with more ALUs at the same clocks (unlike Nvidia, whose Maxwell and Pascal cards, at least, do). There seems to be some kind of architectural bottleneck in AMD's current GPU tech which prevents them from efficiently using larger chips, and the fact that Vega doesn't show a performance per clock increase over Fiji suggests this is still the case. The GCN front-end doesn't seem to have been substantially changed since the first generation, and I suspect it's something they're going to have to look at soon.

Incidentally, back when Navi was first announced they used the word "scalability" to describe it, so it may be a hint that they're taking a TR/Epyc style multi-die solution for larger GPUs, which would almost certainly require a substantially re-designed front-end to work efficiently.

Well, yeah, this may be true but here's what's bothering me here: Vega did get a new frontend and a new backend already. If they were still rebuilding them then they had the option of making sure that they would work good with a much wider shader core - something which Fiji didn't get presumably because GCN3's scope wasn't big enough.

But you may be right, if we consider that GCN's utilization goes down with more CUs being idle each clock in a wider GPU.

Surely you agree it would be much worse if it only clocked to ~1.35GHz like Polaris?

As I said, I don't think the increase in die size has that much to do with the increased clocks. With the exception of TBR, almost all of the improvements in Vega seem to have been focussed on compute (where the card does seem to perform well). AMD don't have the R&D budget to put out pure compute cards like the GP100/GV100, and they want to start competing better in the workstation/server/neural net market, so we've got a die which has to do that and work as a gaming GPU. It's resulted in a larger die than if they went with 64CU Polaris, but at least we've got increased clocks to go along with those compute improvements.

Vega 10 is not a compute chip, they don't position it as such, and it's missing the main feature relevant for HPC -- fast FP64 math. It's also missing ECC support which means that Vega 10 cannot be compared to HPC chips like GP100 and GV100. They do support FP16x2 which is obviously targeted at DL/AI market but that's the extent of its non-gaming applications, and as I've said already it's very unlikely that FP16x2 has added more than 10% of die size to the chip.

 

[tuxfool]

Surely you agree it would be much worse if it only clocked to ~1.35GHz like Polaris?

As I said, I don't think the increase in die size has that much to do with the increased clocks. With the exception of TBR, almost all of the improvements in Vega seem to have been focussed on compute (where the card does seem to perform well). AMD don't have the R&D budget to put out pure compute cards like the GP100/GV100, and they want to start competing better in the workstation/server/neural net market, so we've got a die which has to do that and work as a gaming GPU. It's resulted in a larger die than if they went with 64CU Polaris, but at least we've got increased clocks to go along with those compute improvements.

Yeah, this is kind of the thing. If they get any share of the compute market, that segment is worth a lot more than the equivalent share of the gaming market.

 

[E-Cat]

Here is the Capsaicin event at Siggraph 2017 for anyone who's interested:

https://www.youtube.com/watch?v=nuFGUVg0U9U

 

[thelastword]

Naturally, they take more adventurous path, as they need to roll out a hit to regain market share. So they bet on HBM/HBM2 and it didn't work out/is a disaster (48x mm^2 chip vs 314 mm^2 chip with cheaper memory and much lower power consumption).
Ryzen's infinity fabric, on the other hand, worked great.

HBM2 along with HBCC is a genius combo, I don't see how it's a disaster, especially since the product is not on shelves yet and you have not seen games and apps using the strengths of said architecture....

Dat TDP.
Welp my hopes of a budget AMD build for a ITX case just died.

Sigh....Vega 56 is 150-165 watts.......Moreover, there will be vega cards hitting polaris levels of performance in vein of the Rx 470/480/580 and even 460 eventually. There's also raven ridge which should hit later this year (the ryzen+Vega APU).....

Damn 1000w psu I only have 850, guess my 980ti will have to last me until Nvidia next batch of cards.
Assuming Nvidia releases any, at this point they could just keep the exact same cards at the exact same prices till holiday 2018

Where is this coming from? Why would you believe this....Is it that people would rather latch unto negative things or untruthful statements.? There's no way you need 1000w PSU even for the 350TDP vega part.....I mean if you're buying a PSU now, I'll advise you to get a 1000w, but your 850w PSU is just fine. I have an 850w EVGA Supernova btw and I'm getting Vega....

 

[JaseC]

Funny that so many are not giving this post more feedback. The gamernexus breakdown is probably the most thorough and levelheaded one describing the specs and offers....on Vega...

There's not much to glean from the current FE drivers not having power efficiency measures or TSBR enabled when even AMD is comparing the 64 to the 1080. The logical conclusion to draw here is that there's no ace up AMD's sleeve that will kick the 64 into overdrive and make it competitive with the 1080 Ti.

Sigh....Vega 56 is 150-165 watts.......Moreover, there will be vega cards hitting polaris levels of performance in vein of the Rx 470/480/580 and even 460 eventually. There's also raven ridge which should hit later this year (the ryzen+Vega APU).....

210, actually, but I'd imagine he could accommodate that, too.

Where is this coming from? Why would you believe this....Is it that people would rather latch unto to negative things or untruthful statements.? There's no way you need 1000w PSU even for the 350TDP vega part.....I mean if you're buying a PSU now, I'll advise you to get a 1000w, but your 850w PSU is just fine. I have an 850w EVGA Supernova btw and I'm getting Vega....

He didn't pluck the figure out of thin air and he's not perpetuating an anti-Vega rumour. 1000W is what Gigabyte recommends. Official PSU recommendations are always overstated, though.

 

[Sir Abacus]

HBM2 along with HBCC is a genius combo, I don't see how it's a disaster, especially since the product is not on shelves yet and you have not seen games and apps using the strengths of said architecture....

Blast Processing.

 

[wbEMX]

That's wrong. A good 1080 is €550, 19% of which are taxes (which are always excluded when US prices are quoted). Removing that tax and converting to USD, you get $540.

That's why I later said this:

Damn, I always forget the VAT thing. Didn't keep that in mind, my bad.

I do think that the better 1080s start at about 600 €, though. Had some bad experiences with Palit, which is the only non-reference card that retails for this price on Mindfactory right now.

 

[thelastword]

There's not much to glean from the FE not having power efficiency measures and TSBR enabled when even AMD is comparing the 64 to the 1080. The logical conclusion to draw here is that there's no ace up AMD's sleeve that will kick the 64 into overdrive and make it competitive with the 1080 Ti.

We'll see, but AIB cards with higher clcoks and better coolers should be interesting. As I said though. I'm more interested in how beneficial the vega features are to gamers while they play, higher minimums, less stress and stutter under big explosiosn and alpha, less stutter and lod issues etc...Of course the potential of this card can't all be seen from day one since games will have to utilize vega functions, but I'm stoked by the games being developed with Vega inmind and I'm stoked for vulkan games in general..

You may not believe, but I think Vega performance will increase a bunch by release and thereafter. I guess it's the AMD way. Launching the product at this time was paramount, they'll sort aout any niggling issues in due course, but I'm already sold oin what they're offering. I'm not so much enthused as to high max fps with low minimums, I'd much prefer for my games to play smooth under high bandwidth moments....

He didn't pluck the figure out of thin air and he's not perpetuating an anti-Vega rumour. 1000W is what Gigabyte recommends. Official PSU recommendations are always overstated, though.

Gigabyte? Are they going to push 2Ghz clocks on Vega, in any case I don't expect it, so yes, 1000w is overstating requirements there....

 

[Thraktor]

You're not really suggesting that NV got +50% clock boost in Pascal simply due to a new process, are you? Because this isn't how it works these days, and there were efforts when building Pascal to make sure that it will clock higher - they've talked about it during Pascal's launch.

All in all Pascal has the exact same emphasis on higher clocks compared to Maxwell as Vega has compared to either Polaris or Fiji (do note that AMD themselves are constantly comparing Vega to Fiji instead of Polaris in their own official slides). But in case of Pascal this clock gain came without a significant complexity increase over Maxwell. So either this was already implemented in Maxwell (which was clocked significantly higher than Kepler mind you) or AMD did something weird in Vega.

It's 50% now? It was 33% in your last post. And yes, I do think the jump from 28nm to 16nm was the major driver in allowing them to hit higher clocks with Pascal.

In any case this is beside the point. I'm specifically arguing that I don't think increased clock speeds were the primary driver behind Vega's jump in transistors count. Maxwell is actually a perfect example, with a notable clock speed jump over Kepler (plus a variety of other architectural improvements) with little to no increase in transistor count or die size.

Go further where? They are launching a 345W card on 14th of August in consumer space. They need to spend all their resources on getting power consumption down in Navi - something which clearly wasn't done in Vega as it's actually much worse than Fiji compared to GM200 here.

I'd wager that them doubling P10 (88 CUs / basically 1.5x of Vega) and spending the effort on optimizing its power consumption would net them better results than increasing Fiji's clock as was done in Vega. Such chip even when running on Polaris clocks, without any power optimization, would already be faster.

GCN is not a high frequency architecture, and this clearly shows in Vega results where they had to spend a shitload of transistors and power to reach what is essentially last gen clocks for NV. If they really wanted to go high clocks they had to make an architectural update in vein of what NV did between Kepler and Maxwell.

Increasing maximum clock speeds and improving power efficiency are effectively the same thing, it's all about moving the power curve down and to the right, allowing them to hit the same clocks at lower voltages or higher clocks at the same voltages. The reason they're launching a 345W card is that they're ramming right up against the end of that power curve, releasing a card with a max OC of probably 1.7GHz with a stock clock of 1.67GHz.

I would bet very good money that Vega at the same ALU count and clock speed is comfortably more power efficient than Polaris. In fact, I wouldn't be surprised to see Vega Nano draw less power than RX580 while significantly outperforming it. We may even see Vega Nano outperform the 1070 at a similar power draw (although almost certainly at a higher price).

Not sure what you mean by NV's "conservative" approach here considering that NV chips are clocked 33% higher than Vega and are running on 1.0V at this while Vega is 1.2V.

That's precisely what I mean. Nvidia could have pushed up boost voltages on Pascal to ~1.1V for a bit of extra performance and said "to hell with power consumption", but they didn't need to. AMD have basically said "to hell with power consumption", at least with the liquid cooled Vega 64.

You seem to completely miss the point that this ~50% clocks increase cost them ~30% die size increase as well. This is not a good investment of die area IMO.

As I've said, I don't think the die size increase had that much to do with the increased clocking ability. I certainly don't think it's cost them more than it's gained.

Mobile GPUs are TBDRs mostly because of who are making them (Imagination) and less due to power savings of tiling. I don't know even, are Adrenos TBDR these days? Tegra sure aren't as they use the same form of tiling rasterization as desktop GeForces.

I didn't say TBDR, I said TBR. As far as I'm aware all mobile GPUs these days use tile-based rasterisation of one form or another.

Unlikely as ALUs themselves do not change much for any type of packed math support. What is changing are the surrounding storage and control logic.

I was joking (a little). When I say ALUs I am talking about control logic, registers, etc., as well, though.

Well, yeah, this may be true but here's what's bothering me here: Vega did get a new frontend and a new backend already. If they were still rebuilding them then they had the option of making sure that they would work good with a much wider shader core - something which Fiji didn't get presumably because GCN3's scope wasn't big enough.

But you may be right, if we consider that GCN's utilization goes down with more CUs being idle each clock in a wider GPU.

They've talked about front-end improvements, but I from what I've read they haven't gone into much detail (although I may have missed it), indicating that they were relatively minor revisions. We're still seeing a 64CU Vega GPU underperform older, narrower GPUs on a pure performance/GF metric, which suggests that there's still something up.

 

[JaseC]

We'll see, but AIB cards with higher clcoks and better coolers should be interesting. As I said though. I'm more interested in how beneficial the vega features are to gamers while they play, higher minimums, less stress and stutter under big explosiosn and alpha, less stutter and lod issues etc...Of course the potential of this card can't all be seen from day one since games will have to utilize vega functions, but I'm stoked by the games being developed with Vega inmind and I'm stoked for vulkan games in general..

I don't think non-ref SKUs are going to push the 64 much further if at all as it's already water-cooled, and even if they did, the point would be moot as the 1080 Ti has a generous amount of overclocking headroom. It'd be disingenuous to compare a well-overclocked 64 to a stock 1080 Ti.

You may not believe, but I think Vega performance will increase a bunch by release and thereafter. I guess it's the AMD way. Launching the product at this time was paramount, they'll sort aout any niggling issues in due course, but I'm already sold oin what they're offering. I'm not so much enthused as to high max fps with low minimums, I'd much prefer for my games to play smooth under high bandwidth moments....

I'm sure AMD will fine-tune things to the best of its ability.

 

[DSN2K]

is the Vega 56 any better then a 1060 ?

 

[RenegadeXV8]

Seems like all future Bethesda games so far will be implementing FP16 along with Far Cry 5 (Ubisoft really likes to flip-flop on GPU partnerships...). Are there any real world examples of FP16 performance gains?

is the Vega 56 any better then a 1060 ?

We won't know for sure till reviews hit, but it should be at least 20% faster, considering the Vega 56 is positioned against the 1070.

 

[Aztechnology]

is the Vega 56 any better then a 1060 ?

Yea it's seemingly between a 1070 and 1080 so quite a bit.

 

[Nheco]

is the Vega 56 any better then a 1060 ?

A 480/580 is already better than an 1060. Common.

 

[Space_nut]

A 480/560 is already better than an 1060. Common.

How about the 580?

 

[Aztechnology]

A 480/560 is already better than an 1060. Common.

The 560 is assuredly not. 480/580 is on par with some games better or worse. But you get Freesync which is s bonus.

 

[michaelius]

A 480/560 is already better than an 1060. Common.

Joke post or mistype?

 

[Nheco]

How about the 580?

The 560 is assuredly not. 480/580 is on par with some games better or worse. But you get Freesync which is s bonus.

I mean 480/580, it was a typo.

A good comprehensive review after release of both cards: http://www.hardwarecanucks.com/foru.../73945-gtx-1060-vs-rx-480-updated-review.html

 

[franklyshankly]

I mean 480/580, it was a typo.

A good comprehensive review after release of both cards: http://www.hardwarecanucks.com/foru.../73945-gtx-1060-vs-rx-480-updated-review.html

https://www.techpowerup.com/reviews/Sapphire/RX_580_Nitro_Plus/30.html

and here's one from more recently when the 580 launched:

Pretty much what I would call even.

 

[LelouchZero]

https://www.techpowerup.com/reviews/Sapphire/RX_580_Nitro_Plus/30.html

and here's one from more recently when the 580 launched:


Pretty much what I would call even.

The RX 480 and 580 have more ram with their 8GB models so they are superior in the memory department.

 

[Rains]

Will my 650 watt psu been enought for vega

 

[Coulomb_Barrier]

For the most part they are close to each other with the 1060 still taking a slight lead in the newest games like nier, ghost recon, me:a and even the amd sponsored prey. RE 7 seems to run faster on polaris though.

On average, if you look at bigger PC games released in 2017 (can't find benches for most indie titles), the 480 is faster than the 1060. It's faster in For Honor, Res 7, Sniper Elite 4, DIRT 4, Dawn of War 3 etc etc. Not disputing it's slightly behind it in those titles you mention though.