CHOO CHOOOOO!!!!1! Navi Hype Train be rollin'

[Vya Domus]

0.8Kb is still a lot of information if it can be kept full, but alas when it CAN'T usually you have two choices as then you have a shader using power, making heat, and not doing work. One is to improve cache hit rate but that takes a lot of tuning and tweaking, or you can just add more cache to increase the chances the data will be loaded, but that takes more power to run and makes more heat.

That's just not how this works, firstly a shader that doesn't do work doesn't use power (or very little), because of something called power-gating. Not that it wold matter because this rarely happens, GPUs are designed to maximize utilization without the need of big caches/registers and complex caching algorithms. No one does that because those things would take so much more die space that you would have to decrease the overall number of ALUs and that would nullify whatever advantage that was supposed to bring.

A 'shader' is a small program written in GLSL which performs graphics processing, and a 'kernel' is a small program written in OpenCL and doing GPGPU processing. These processes don't need that many registers, they need to load data from system or graphics memory. This operation comes with significant latency. AMD and Nvidia chose similar approaches to hide this unavoidable latency: the grouping of multiple threads. AMD calls such a group a wavefront, Nvidia calls it a warp. A group of threads is the most basic unit of scheduling of GPUs implementing this approach to hide latency, is minimum size of the data processed in SIMD fashion, the smallest executable unit of code, the way to processes a single instruction over all of the threads in it at the same time.

Secondly, the concept of improving the hit rate on a GPU cache doesn't even make sense because there is nothing you can do. You already know that the same sequence of instructions will run thousands of times across multiple CUs, therefor you can schedule the execution in such a way that you can always have the data ready if you have enough memory bandwidth. And that's what everyone does, including Nvidia.

Here's another hint : AMD calls their execution units Stream processors and Nvidia names their cores Streaming Multiprocessors. Still don't believe me ?

GP100 : 3840 shaders , 4MB L2 cache
Vega 64 : 4096 shaders, 4MB L2 cache

Turns out they aren't that different are they ? GCN's problem isn't the cache size or hit rate or anything like that, it's something else. It's the fact that they have a lot more complex logic on chip whereas Nvidia offloads most of it to software, I am not going to go into details but that's what is using a lot of power and what makes gaming performance unimpressive. I'll just name one thing, AMD has logic that allows for scalar instructions to be executed within each CU, Nvidia doesn't have such thing, this is mostly a worthless addition that creates even more scheduling overhead as far as graphics workloads go but great for compute.

 

[TheoneandonlyMrK]

I have talked about this in another thread and basically this isn't true, not just in AMD's case but for GPU designs in general. Caches are not a critical factor for achieving high performances/utilization unlike it is the case with CPUs, you don't even have to believe me you only need look at similar sized dies for GPUs and CPUs and see how much the cache/core/shader ratio differs.

For instance on GP104 theoretically there is a grand total of 0.8 Kb of L2 cache that you can expect per shader, this is an abysmally small amount yet these GPUs operate just fine. Hit misses are mostly irrelevant because their latency is hidden by the fact that there are already instructions scheduled, therefor there is no need for quick frequent access of memory which would require large fast caches with very good hit ratios.

Instead what you actually need is a lot of memory bandwidth, AMD designs their GPUs just fine from this point of view, there is literally no other way of doing it. The reason GCN based cards have had more memory bandwidth and cache than their Nvidia equivalent is because they incidentally also have more ALUs typically. There is no mystery to any of this, it's all quite simple. I don't know why people have the impression that these guys could make such huge glaring oversights in their designs, they aren't idiots, they know what they are doing very well.

It's the massive oversights people have with their perspective, not asic designers, if they make mistakes ,heads roll and millions get flushed down the loo.

All for people to vastly miss-understand that it was'nt,, surprisingly ,,designed and created JUST for them to play Crisis.
The fact that for three generations Nvidia have stepped outside that box and given people hardware that's really only good within its shelf life at the SPECIFIC task of 32bit Gaming , proven out by the tanking performance of 680's and 780's in todays games verses AMD's more complete and adaptable HAwaii's ok/meh performance.

Never mentioned and Oft forgotten AMD GPUS are not just sold by AMD as AMD GPU's ,whereas Geforce GPU's for gaming do not do other tasks well and are not as usefull outside there core use and are only sold by Nvidia.

 

[Steevo]

So you say a 680 is better than a 7970 then prove it's depending on use case and give proof that said Nvidia Gpu didn't age well.
Depending on use case the 7970 was always better depending on perspective ,i use compute.
But anyway would it not be better to actually discuss the Op then regurgitate the same arguable points about dead tech.
If Amd do Raytracing on navi 10 ill be surprised tbh.
Navi 20 i expect to have a go, we'll see how that goes in time.

Oh and he's right GPU are designed for streams of data not OoO data streams so the cache isn't used for possible hit's only expected and is quite small in footprint terms compared to Cpu caches.
That's why Gpu memory bandwidth matters more to Gpus then system memory bandwidth matters to CPU's , they can't store many instructions and dont have teired caches like CPU's do to buffer poor memory bandwidth.

https://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_680/27.html

I assume W1zz isnt biased but the numbers are here in our own reviews, along with transistor counts, power consumption, performance....

 

[Fouquin]

It is a bit of both. ISA literally means Instruction Set Architecture. There are some things on different levels this does set in stone but many others can be improved on. Whether some things that need improvements are fixed or not is not easy to know.

Indeed it is both, that's why I said it was both in my post. Someone else gave a great analogy to put perspective on why sticking with GCN is not actually a bad thing, and they said, "Netburst was also x86." Now think about that. Netburst and Skylake share the x86 ISA and are so vastly different in performance that they are realistically not even comparable to each other. GCN does have some limitations (to be fair, so does x86) but it was designed to scale, and AMD could definitely be on track to release a completely rearranged micro architecture with Navi.

 

[TheoneandonlyMrK]

https://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_680/27.html

I assume W1zz isnt biased but the numbers are here in our own reviews, along with transistor counts, power consumption, performance....

2012 called, it wants its review back. Read my post again ,i said in today's game's.

 

[biffzinker]

Indeed it is both, that's why I said it was both in my post. Someone else gave a great analogy to put perspective on why sticking with GCN is not actually a bad thing, and they said, "Netburst was also x86." Now think about that. Netburst and Skylake share the x86 ISA and are so vastly different in performance that they are realistically not even comparable to each other. GCN does have some limitations (to be fair, so does x86) but it was designed to scale, and AMD could definitely be on track to release a completely rearranged micro architecture with Navi.

That only works for the x86 ISA because of the front-end decoding logic to the internal RISC ISA. I'm not seeing AMD wanting to add in decode just so they can switch the execution back-end to another micro architecture.

 

[RealNeil]

Just got an MSI GTX-1660Ti Gaming-X for the wife's PC. Guy sold it to me cheap.

I still plan to get an AMD CPU/Board combination when they're released in a few months.
And maybe, if NAVI is all that, I'll sell my two Vega-64 cards and buy a pair of NAVI cards. (only if I'm going to see real improvement over the two Vega-64 cards)

 

[Vya Domus]

Indeed it is both, that's why I said it was both in my post. Someone else gave a great analogy to put perspective on why sticking with GCN is not actually a bad thing, and they said, "Netburst was also x86." Now think about that. Netburst and Skylake share the x86 ISA and are so vastly different in performance that they are realistically not even comparable to each other. GCN does have some limitations (to be fair, so does x86) but it was designed to scale, and AMD could definitely be on track to release a completely rearranged micro architecture with Navi.

ISAs in GPUs are irrelevant for the most part and AMD/Nvidia changes them all the time anyway.

 

[Steevo]

2012 called, it wants its review back. Read my post again ,i said in today's game's.

Well I have a 7970 and know others that have 680s and clock for clock we are still close. So please don't bring up fine wine, or Nvidia gimping, we have already beat that horse.

 

[Fouquin]

I'm not seeing AMD wanting to add in decode

GCN CUs have front-end decoding logic to SIMDs.

ISAs in GPUs are irrelevant for the most part and AMD/Nvidia changes them all the time anyway.

And both have virtual ISAs that are hardware agnostic (PTX and HSAIL). It does not matter what changes are made down at the hardware, the programming model doesn't change. GCN ISA was designed with exactly that in mind, having a configurable architecture that requires no changes in programming.

That is the whole reason I believe they could make more significant changes. Vega already displays some very substantial changes to the hardware, but the arrangement of that hardware remained the same. I believe that they do have the ability to restructure the hardware with Navi.

 

[Assimilator]

Since we're all making arbitrary predictions, here's mine: AMD will dump the graphics portion of its company, in exchange for lots of money and a 5-year exclusivity and revenue-sharing agreement with whoever buys said graphics IP. The end result is that AMD will continue to sell and support prosumer GPUs under its brand, but the third party (let's call them "ATI", for no good reason) will sell consumer GPUs under its brand, and revenue from APUs that go into consoles will be split between them.

This allows AMD to concentrate on what they're doing well at now - CPUs - while also giving them a nice cash injection, a guarantee that they have a source of GPUs to put in their chips, and keeps their corporate graphics customers happy. ATI gets investor cash which means they can finally focus on just delivering GPUs.

 

[TheoneandonlyMrK]

Well I have a 7970 and know others that have 680s and clock for clock we are still close. So please don't bring up fine wine, or Nvidia gimping, we have already beat that horse.

You were in here beating the dead horse and still are ,you started on Gpu cache in efficiency and now you're bringing up fine wine , again re read my first post.
@Assimilator so when every other Soc maker want's arm or x86 And a Gpu core in house Amd will step the other way after converting to an Soc and design house hmnn.

 

[eidairaman1]

Just chill, doesn't matter. This is about Navi despite it being at the sub atomic elements of hydrogen chloride...

End the pissing match.

I have my preference as the next guy does.

 

[Deleted member 50521]

Since we're all making arbitrary predictions, here's mine: AMD will dump the graphics portion of its company, in exchange for lots of money and a 5-year exclusivity and revenue-sharing agreement with whoever buys said graphics IP. The end result is that AMD will continue to sell and support prosumer GPUs under its brand, but the third party (let's call them "ATI", for no good reason) will sell consumer GPUs under its brand, and revenue from APUs that go into consoles will be split between them.

This allows AMD to concentrate on what they're doing well at now - CPUs - while also giving them a nice cash injection, a guarantee that they have a source of GPUs to put in their chips, and keeps their corporate graphics customers happy. ATI gets investor cash which means they can finally focus on just delivering GPUs.

Sell it to, say, Intel?

I like this hypothesis

 

[moproblems99]

Since we're all making arbitrary predictions, here's mine: AMD will dump the graphics portion of its company, in exchange for lots of money and a 5-year exclusivity and revenue-sharing agreement with whoever buys said graphics IP. The end result is that AMD will continue to sell and support prosumer GPUs under its brand, but the third party (let's call them "ATI", for no good reason) will sell consumer GPUs under its brand, and revenue from APUs that go into consoles will be split between them.

This allows AMD to concentrate on what they're doing well at now - CPUs - while also giving them a nice cash injection, a guarantee that they have a source of GPUs to put in their chips, and keeps their corporate graphics customers happy. ATI gets investor cash which means they can finally focus on just delivering GPUs.

The whole reason they bought it was APUs and they are (potentially) near parity with Intel on CPUs and have (temporary, at least) lead in GPUs. I can't see them selling Ati off now that the plan is coming to Fruition.

 

[Deleted member 50521]

The whole reason they bought it was APUs and they are (potentially) near parity with Intel on CPUs and have (temporary, at least) lead in GPUs. I can't see them selling Ati off now that the plan is coming to Fruition.

They(AMD) intended to build Fusion (CPU+GPU), which failed flat on their face. That ship has sailed and there is no going back from that. AMD is better off without (consumer) grade GPU

Big APU is never gonna work. CPU and GPU are simply too different to incorporate into one single die. Not good for effiecnecy, and difficult to scale up/down. APU was, is and will always remain in the minimum low performance segment of the market.

 

[biffzinker]

Big APU is never gonna work. CPU and GPU are simply too different to incorporate into one single die.

I could work out if not for the memory bandwidth deficiency.

 

[Deleted member 50521]

I could work out if not for the memory bandwidth deficiency.

yeah that too as well. All the data congests down over current DDR4 transfer capacity. Or make it even more expensive and difficult to scale: put a bunch of HBM2 on there.

Modularity with good efficiency > All-in-One

 

[eidairaman1]

The whole reason they bought it was APUs and they are (potentially) near parity with Intel on CPUs and have (temporary, at least) lead in GPUs. I can't see them selling Ati off now that the plan is coming to Fruition.

Chipsets was another reason

 

[64K]

Since we're all making arbitrary predictions, here's mine: AMD will dump the graphics portion of its company, in exchange for lots of money and a 5-year exclusivity and revenue-sharing agreement with whoever buys said graphics IP. The end result is that AMD will continue to sell and support prosumer GPUs under its brand, but the third party (let's call them "ATI", for no good reason) will sell consumer GPUs under its brand, and revenue from APUs that go into consoles will be split between them.

This allows AMD to concentrate on what they're doing well at now - CPUs - while also giving them a nice cash injection, a guarantee that they have a source of GPUs to put in their chips, and keeps their corporate graphics customers happy. ATI gets investor cash which means they can finally focus on just delivering GPUs.

When they split off their graphics division and made it a separate company (RTG) I wondered the same thing. Add to that that I have seen business articles that they make more revenue from CPUs than they do GPUs and the money from selling RTG would probably allow them to get completely out of debt.

But who will buy RTG? Some said Intel would but they don't need to. They have already lured some of the top management away and who knows how many of AMD's engineers. They are acquiring the best people from AMD the cheap way. Some said Samsung might but that turned out to be just a rumor. Whoever would consider buying RTG would have to be prepared to take on Nvidia and Intel next year as well. That seems pretty daunting to me.

 

[Vya Domus]

RTG isn't going anywhere, not for the following 5 years at the very least. The new consoles are knocking on the door, that'll be a nice stream of cash going in for the foreseeable future, after that who knows what will be. You are all forgetting that Navi in PCs is literally a byproduct of that. And make no mistake, if the time comes and for some reason they decide to sell their GPU division, they are not going to give it away for cheap. GPU manufactures have slowly disappeared over the years, this has become a pretty exclusive industry.

Regardless, I still don't think it'll ever happen, unless some catastrophic event forces them to do that but they seem to have come out of many crises in the past just fine.

 

[64K]

RTG isn't going anywhere, not for the following 5 years at the very least. The new consoles are knocking on the door, that'll be a nice stream of cash going in for the foreseeable future, after that who knows what will be. You are all forgetting that Navi in PCs is literally a byproduct of that. And make no mistake, if the time comes and for some reason they decide to sell their GPU division, they are not going to give it away for cheap. GPU manufactures have slowly disappeared over the years, this has become a pretty exclusive industry.

Regardless, I still don't think it'll ever happen, unless some catastrophic event forces them to do that but they seem to have come out of many crises in the past just fine.

I don't think RTG is going to be sold either. I don't see any company that would buy it considering the competition right now and in the future.

You're right AMD has faced several crisis. Just a few years ago there were several financial analyst sites that were saying AMD would most likely have to file bankruptcy but look at them now. Back to being profitable and paying off their debt as well. Lisa Su doesn't get enough credit for turning AMD around and making a profit now. She did it by focusing R&D on Ryzen and letting the GPU side somewhat stagnate. Back when AMD was focusing on their GPU business and letting their CPU side somewhat stagnate they were going into red ink by hundreds of millions of dollars each year.

 

[Vya Domus]

And who is to say that this "focus primarily on one division at a time" is not an actual strategy that they want to follow.

 

[Vario]

Well I have a 7970 and know others that have 680s and clock for clock we are still close. So please don't bring up fine wine, or Nvidia gimping, we have already beat that horse.

Both aged fine. Was using a 770 2GB until 2018 and it was fine. 1440P with medium settings and 60FPS stable smooth. Had a 7970 in 2013 and it was indistinguishable from the 770 at the time. Vram limit is a valid argument but I personally don't care much about graphic detail so turning down textures and filters to achieve a max under the 2GB threshold was not a big deal for me. I agree with steevo and let the 2012 graphics card battle die its been 7 years now!

 

[Vayra86]

2012 called, it wants its review back. Read my post again ,i said in today's game's.

What about them? There is an odd case of having better performance due to 1GB of extra VRAM but that os really all she wrote and has little to do with the argumentation which is that, once again we are hearing a repackaged AMD fine wine here. We know this doesnt really exist. In the larger scheme of things a 680, and a 7970 are equally obsolete and relegated to budget/low end performance level.

It needs no discussion that GCN is lacking the efficiency it needs to compete. That will only get worse as the die space and power budget is limited. The best architecture is the one that can keep clear of these limitations. The moment you touch them on the current node is a sign you are getting behind the curve, and AMD ignored those signs since 2013. Nvidia offers us an architectural update every time they risk having to move up from their cut down big die. The only time we got the full Titan in a consumer chip was with the 780ti, and only because Hawaii existed.