AMD CPU speculation... and expert conjecture

[juanrga]

Well, I gave up on linux drivers after I rebuilt kernel, recompiled some things, and got a hard lock when initializing xorg. Back on open source.

The new GPU leaks are interesting. Nvidia seems to be closing in on AMD's high resolution advantage, looking at how GM200 does well against Fiji/Bermuda at 1440 but at 4k it's not so hot.

Open source drivers still are lagging behind by a bit. But they are less work to deal with keeping working on Linux and the binary drivers don't have working backlight control on my laptop last I checked. And switching in Gentoo can be a pain sometimes. Maybe if I get bored over the holidays.

I use and recommend the open drivers.

The leaked benchmarks have killed the hype about AMD preparing a supercard would leave Nvidia in the dust for "10 years" (sic). At 4K the 390X is only 11% ahead of the Titan 2, which is well within the benchmark error and driver version.

This fight also reveals two philosophies at getting performance: At the one hand the brute force approach of AMD, which even using 20nm and HBM to reduce power consumption will require liquid cooling; at the other, the efficiency architectural approach of Nvidia. Lovely!

 

[colinp]

In the high end GPUs I'm not sure that HBM is the solution to the problem. I mean, are AMD GPUs really that bandwidth constrained? I'm sure HBM will be great if it ever makes it to APUs or laptop GPUs though. But fundamentally, what they need is a more efficient arch. A more efficient arch means moar cores for the same power budget.

 

[esrever]

In the high end GPUs I'm not sure that HBM is the solution to the problem. I mean, are AMD GPUs really that bandwidth constrained? I'm sure HBM will be great if it ever makes it to APUs or laptop GPUs though. But fundamentally, what they need is a more efficient arch. A more efficient arch means moar cores for the same power budget.

HBM itself is a huge power savings over high bandwidth GDDR. And the only reason we aren't bandwidth constrained is because nothing has been pushing bandwidth in gaming. But in HPC, the need for bandwidth is very significant.

Also HBM isn't that great for low end parts because it doesn't save as much power and it is going to be expensive for a while. Expect to only see it in the top end for a while.

 

[8350rocks]

So, anyone hit up the AMA with robert hallock?

 

[Reepca]

So, anyone hit up the AMA with robert hallock?

I read through it, didn't see too much that interested me personally - it's mainly just answers from the graphics department that don't tell us too much we don't already know (I learned a good deal about FreeSync though). Perhaps the most useful information I got from it is that there is an AMA regarding the CPU side of AMD in... January I think? And that even if we ask nicely, they still can't tell us anything about upcoming cards :'{

 

[esrever]

I feel like AMD is having a lot of trouble right now and they will have a rough 2015. Just my gut feeling.

 

[juanrga]

In the high end GPUs I'm not sure that HBM is the solution to the problem. I mean, are AMD GPUs really that bandwidth constrained? I'm sure HBM will be great if it ever makes it to APUs or laptop GPUs though. But fundamentally, what they need is a more efficient arch. A more efficient arch means moar cores for the same power budget.

Current cards are too slow to exploit full HBM bandwith. AMD will use HBM on the 390 series to reduce power consumption and compensate for the lack of efficient architecture. Same reason why will use 20nm instead 28nm.

New GCN architecture is planned for 2016, it would be oriented towards power efficiency, but I am a bit skeptic on what AMD can achieve. Nvidia uses VLIW, but AMD migrated from VLIW4 to RISC SIMD with GCN and, a priori, the RISC approach is less efficient.

AMD could try to increase the SIMD width from 2048bit to 4096bit to increase efficiency (and throughput) by ~2x, but this will hurt performance on divergences unless AMD find some way to solve this problem.

 

[blackkstar]

In the high end GPUs I'm not sure that HBM is the solution to the problem. I mean, are AMD GPUs really that bandwidth constrained? I'm sure HBM will be great if it ever makes it to APUs or laptop GPUs though. But fundamentally, what they need is a more efficient arch. A more efficient arch means moar cores for the same power budget.

HBM will have significantly reduced power consumption.

https://www.skhynix.com/gl/products/graphics/graphics_info.jsp

I looked online to see if I could remember how much power 4GB of GDDR5 draws, but IIRC it's a lot, like 50w or so. If that's the case, HBM would only draw ~20w for 4GB. That's 30w savings you can use to increase GPU performance. When you have a 300w limit, that's 10% more TDP you can use for GPU just by switching to HBM.

Also, HBM will eliminate the bandwidth problem almost entirely. Do you think it's a coincidence AMD is adding downsampling to their drivers? They are preparing to make downsampling from 4k to 1080p and 2k a reality. And they are doing this to leverage their architecture's large advantage over Nvidia at higher resolutions while 4k monitors still remain a luxury item.

AMD is preparing a next generation geared for 4k. However 4k monitors are still expensive. But if they render at 4k and then downsample to 1080p or so, the results look way, way better than shader based AA or even MSAA. And they'll have a massive software advantage over Nvidia. Blurry FXAA is going to look absolutely terrible next to a screenshot of a 4k downsampled screenshot. And if you don't believe me, people already downsample, you can see yourself by googling. And Nvidia will be left behind because they already have an architecture that doesn't scale as well to higher resolutions (the difference between AMD and Nvidia grows more and more in AMD's favor as resolution increases).

HBM on 300 series is going to be a very big deal and it's going to give AMD a huge image quality advantage over Nvidia thanks to downsample. And it's going to free up a bunch of TDP and power consumption budget for AMD to shift from powering GDDR to increasing clocks, die size, etc.

I don't think you should be worried about HBM bandwidth going to waste at 1080p. I think you should be more worried about the performance difference between an AMD GPU with HBM and an Nvidia one at 4k downsampled as well as power consumption.

 

[juanrga]

HBM will have significantly reduced power consumption.

https://www.skhynix.com/gl/products/graphics/graphics_info.jsp

I looked online to see if I could remember how much power 4GB of GDDR5 draws, but IIRC it's a lot, like 50w or so. If that's the case, HBM would only draw ~20w for 4GB. That's 30w savings you can use to increase GPU performance. When you have a 300w limit, that's 10% more TDP you can use for GPU just by switching to HBM.

The power consumption depends of the clocks. According to AMD labs, the 4GB of GDDR5 on the 6990 consume about 25% of the total power of the card. HBM reduces consumption by about one half. This means that using HBM improves efficiency by about 12.5% (i.e. saves 47W), which means 12.5% more performance on same power budget: max 375W.

Also, HBM will eliminate the bandwidth problem almost entirely.

Nope. We will continue relying on registers and caches.

 

[juanrga]

AMD is preparing a next generation geared for 4k. However 4k monitors are still expensive. But if they render at 4k and then downsample to 1080p or so, the results look way, way better than shader based AA or even MSAA. And they'll have a massive software advantage over Nvidia. Blurry FXAA is going to look absolutely terrible next to a screenshot of a 4k downsampled screenshot. And if you don't believe me, people already downsample, you can see yourself by googling. And Nvidia will be left behind because they already have an architecture that doesn't scale as well to higher resolutions (the difference between AMD and Nvidia grows more and more in AMD's favor as resolution increases).

Cof cof!

Leaks show a Titan 2 only 10% behind the 390X at 4k, despite Nvidia is using 28nm and GDDR5. This clearly shows that the architecture scales up well. Once Nvidia jumps to 16FF and HBM Nvidia will be ahead again... and then AMD will jump to 14FF and so on and so on.

 

[jdwii]

I've been concerned about Amd architecture for their GPUs for sometime and even stated its not as efficient as even Kepler and with maxwell well let's just ouch to amd. Now we are seeing hbm memory and 20nm and its not even 30% better. Nvidia as of today doesn't have a ti product such as a 980ti so we will see those two fight it off.

Remember not final drivers

 

[esrever]

I've been concerned about Amd architecture for their GPUs for sometime and even stated its not as efficient as even Kepler and with maxwell well let's just ouch to amd. Now we are seeing hbm memory and 20nm and its not even 30% better. Nvidia as of today doesn't have a ti product such as a 980ti so we will see those two fight it off.

Remember not final drivers

GCN is more efficient than kelper, 7870 was the most efficient GPU until the 750ti. The high end GCN cards have redundancies for compute and bigger buses that make them less efficient.

 

[jdwii]

That is actually true it did have decent efficiency at lower levels but the design can't scale up as efficiently as nvidia. When i talk about their cards I generally talk about their higher end cards

 

[esrever]

Its not about scailing as it is about AMD trying to get into professional graphics market. All their top end cards have magnitude more compute performance than nvidia's cards. Nvidia has been cutting compute from all their cards, even the titan had less compute than the 290x. It makes AMD's top end cards look inefficient but if you look, they are some of the most efficient HPC accelleraters in the market.

AMD needs a top to bottom release and introduce a mid range card without compute and they will easily be able to compete with nvidia in power consumption.

 

[jdwii]

Amd seems like they are always trying to find themselves. If Nvidia really thought compute was more important for sales they would probably make a product like that. No one buys a 290x for compute that is their firepro. Not to mention shouldn't Amd focus on markets were there isn't severe competition that's what I thought the k12 was all about.

Performance per watt is the MOST important metric in the work station and server market along with price and performance. I know at my school we run amd 6670's in all the PCs except for are work stations which use Nvidia cards I'm actually not sure what model

 

[esrever]

The chips are the same for both companies, firepro and radeons use the same chips and so do geforce and quadros. Nvidia already hold massive market share in the professional markets and its where the money is. AMD is trying to gain marketshare there because the margins are better than consumer cards. There is barely any competition in servers in terms of performance or performance/watt for AMD, they are beating nvidia by miles.

And performance per watt is what AMD is currently the crown of in the workstation and server market. But the most important thing is actually mindshare. Companies and people who buy those cards are conservative and will tend to keep buying from the same vendors if it works for them.

 

[jdwii]

In terms of raw compute amd is the winner but that doesn't matter to us gamers. I find it very important however. I honstly never knew Amd was gaining anything with Nvidia having such a big name with those companies. That and cuda is generally favored saddly.

I'm interested in a 200 watt kick but gaming card seems like after that 200 wattmetric GPUs tend to lose to much in terms of price performance and performance per watt something that matters in SO CPC gamers.

 

[jdwii]

Soc and PC gamers sorry

 

[juanrga]

hey juan would not using it as a L3 help ?
https://semiaccurate.com/forums/showthread.php?t=8125

For dGPUs HBM will be used as replacement for GDDR5, i.e. as VRAM.

For CPUs/APUs? it depends. For laptop/desktop, I expect HBM will be used as LLC (Last Level Cache) and complemented with DDR4 as system memory. For some semicustom/embedded designs HBM could be used as system memory (think for example of a PS5 console with GDDR5 replaced by HBM).

As I mention in the second page of the thread that you give, Intel will use stacked DRAM plus DDR4 for the next Xeon Phi CPU. However, Intel will allow the stacked DRAM to be configured as LLC, as addressable memory, or as a mixture.

In any case, HBM doesn't solve the bandwidth problem and we will continue relying on registers and several levels of cache.

 

[juanrga]

Its not about scailing as it is about AMD trying to get into professional graphics market. All their top end cards have magnitude more compute performance than nvidia's cards. Nvidia has been cutting compute from all their cards, even the titan had less compute than the 290x. It makes AMD's top end cards look inefficient but if you look, they are some of the most efficient HPC accelleraters in the market.

AMD gamer cards have more compute than Nvidia gamer cards, but this is not so for professional/compute cards

AMD S9150: 2.53 TFLOPS
Nvidia K80 2.91TFLOPS

AMD needs a top to bottom release and introduce a mid range card without compute and they will easily be able to compete with nvidia in power consumption.

I.e you are basically suggesting AMD to undo what they did when migrated from VLIW4 to GCN. But they will not do it, because compute is a basic ingredient for HSA and for their APU plans. Thus future graphic architectures will be more and more focused to compute.

 

[cdrkf]

Its not about scailing as it is about AMD trying to get into professional graphics market. All their top end cards have magnitude more compute performance than nvidia's cards. Nvidia has been cutting compute from all their cards, even the titan had less compute than the 290x. It makes AMD's top end cards look inefficient but if you look, they are some of the most efficient HPC accelleraters in the market.

AMD gamer cards have more compute than Nvidia gamer cards, but this is not so for professional/compute cards

AMD S9150: 2.53 TFLOPS
Nvidia K80 2.91TFLOPS

AMD needs a top to bottom release and introduce a mid range card without compute and they will easily be able to compete with nvidia in power consumption.

I.e you are basically suggesting AMD to undo what they did when migrated from VLIW4 to GCN. But they will not do it, because compute is a basic ingredient for HSA and for their APU plans. Thus future graphic architectures will be more and more focused to compute.

The 'problem' AMD are facing in terms of efficiency for their gaming cards is down to the fact that they are trying to save money by utilising a 'one size fits all' aproach.

They have designed *one GPU* in each segment and are using it for both gaming *and* professional applications. Nvidia have developed 2 gpu's in key areas (one for gaming, the other for pro) to get the best of both worlds. Quadro cards have considerably more compute than Geforce for this reason. It is true though, for all those bashing on GCN, it does outshine Nvidia in the workstation / pro markets at pretty much every price point. Also, from personal experience in video studios AMD's solution to multi screen support is considerably more advanced and flexible than the best Nvidia can offer. The crazy thing is that the current situation of GNC vs Maxwell (high power, high compute vs Lower Power, no computer) is pretty much the reversal of the situation of Caymen vs Fermi. At the time nVidia were bigging up the importance of their cards having compute capability to accelerate physics and such in games. Now all of a sudden that no longer matters :S

Their issue is CUDA, driver support and critically opinion. Everyone *knows* Nvidia is somehow better (even if they can't show it in benchmarks, tests or actual use, they still know). AMD are going to have to be better, and remain so for a few generations before that myth will change.

The same is true in the gaming GPU market. Nvidia have a couple of decent cards, but they aren't the home run people are raving about. AMD are selling better cards for less money across the board- yet Nvidia still gain market share. AMD need to work on their marketing imo as they have decent products. The sad thing about all this is, if AMD can't shift much when they are competitive and have decent product, if they do truly get behind (honestly they're really not as far as I can see- Nvidia have one card that is faster than AMD's best, that's all) then we're going to be stuck with only 1 GPU vendor. I like Nvidia kit but I also know they will inflate the price 10x if they get the opportunity.

 

[esrever]

Its not about scailing as it is about AMD trying to get into professional graphics market. All their top end cards have magnitude more compute performance than nvidia's cards. Nvidia has been cutting compute from all their cards, even the titan had less compute than the 290x. It makes AMD's top end cards look inefficient but if you look, they are some of the most efficient HPC accelleraters in the market.

AMD gamer cards have more compute than Nvidia gamer cards, but this is not so for professional/compute cards

AMD S9150: 2.53 TFLOPS
Nvidia K80 2.91TFLOPS

AMD needs a top to bottom release and introduce a mid range card without compute and they will easily be able to compete with nvidia in power consumption.

I.e you are basically suggesting AMD to undo what they did when migrated from VLIW4 to GCN. But they will not do it, because compute is a basic ingredient for HSA and for their APU plans. Thus future graphic architectures will be more and more focused to compute.

K80 is a dual GPU, also I doubt it will go much beyond the base 1.87TFLOPs when running a large workflow. It is only the most powerful card right now because AMD's top end is only based on a single hawaii gpu. The S10000 is based on tahiti.

Also your second part of your post makes no sense and does not reflect anything I said.

 

[jdwii]

^^^ You said they need to release a mid-range card without compute either juan and i and possibly others don't understand what you are saying or you mean Amd should " introduce a mid range card without compute", that would require a redesign something they won't do with their one approach for all method(which juan i must point out Nvidia does the same here as well).

Amd just needs to redesign their GCN a bit that's all it looks like the 390X will do just fine this time around heck for the last couple of generations Amd has been even with nvidia or one is slightly better or worse it just took amd a die shrink and HBM memory to do it this time around and probably more power consumption and heat.

 

[juanrga]

Its not about scailing as it is about AMD trying to get into professional graphics market. All their top end cards have magnitude more compute performance than nvidia's cards. Nvidia has been cutting compute from all their cards, even the titan had less compute than the 290x. It makes AMD's top end cards look inefficient but if you look, they are some of the most efficient HPC accelleraters in the market.

AMD gamer cards have more compute than Nvidia gamer cards, but this is not so for professional/compute cards

AMD S9150: 2.53 TFLOPS
Nvidia K80 2.91TFLOPS

K80 is a dual GPU, also I doubt it will go much beyond the base 1.87TFLOPs when running a large workflow. It is only the most powerful card right now because AMD's top end is only based on a single hawaii gpu. The S10000 is based on tahiti.

The S10000 only gives 1.48 TFLOPS. That is about one half of the above cards.

The K80 was tested by Xcelerit on Monte-Carlo and yields up to 1.90x the performance of the K40 (which is close to the theoretical 2x expected for the new card). Moreover, one week ago the K80 did set a new worldwide record on another benchmark

http://insidehpc.com/2014/12/tesla-k80-gpu-sets-new-records-stac-a2-financial-benchmark/

You saying that AMD has the fastest compute cards was just untrue.

Of course the S9150 is a very good card, but what are the merits? I know it has appeared in the top Green-500, but this is a bit misleading; first. because that list measures the whole computer, not only the cards; second, because was competing against old computers with outdated K20s. I predict that #1 will stay a pair of months when K40s and K80s will be added.

 

[Rafael Luik]

AMD will use HBM on the 390 series to reduce power consumption and compensate for the lack of efficient architecture. Same reason why will use 20nm instead 28nm.

Why isn't the move from 28nm to 20nm considered an advance in architecture to a more efficient one in itself for AMD?
And why isn't HBM considered an advance in architecture also?

You seem to cherry-pick and negate AMD's achievements calling them "compensations" just to say NVIDIA has a better architecure and approach, but the truth is AMD is also advancing in architecure just in different areas than NVIDIA. Correct me if I'm wrong...