AMD Unveils Its 7nm Vega GPU

[Christophe Corazza]

4 minutes ago, Ildiesalot said:

Well maybe this will finally give AMD the push it needs to become competitive, and we can actually start seeing a healthier market.

 

Of course, probably not.

 

Why not? AMD is having fun at Intel's expense.

Intel being Intel, they won't let it all pass by without reacting to AMD's pleasures

[ScratchCat]

1 hour ago, Pangea2017 said:

35% faster marketing is bad. Reality will be some where between 10-20. Paired with the bad vega performance this does not look like they will be better then the 10 series of nvidia.

Just look at intel what upto x times faster means.

AMD have been rather truthful with performance gain estimates.

https://www.anandtech.com/show/12625/amd-second-generation-ryzen-7-2700x-2700-ryzen-5-2600x-2600

1 hour ago, AtlasWraith said:

I agree, but the feeling of hope for the power of the next consumer cards still remains. If they can destroy pascal with their new GPUs, then we'll finally have some damn competition in the market.

Destroying a 3 year old product is not something which can be seen as impressive. AMD will have to improve drastically on the level of Hawaii -> Polaris to keep up efficiency wise.

[Christophe Corazza]

15 minutes ago, ravenshrike said:

 

You gotta think that whoever pushed that internally in Intel is either currently getting slapped upside the head or soon will be.

 

Thanks to that guy we have now something to talk about 

[ScratchCat]

12 minutes ago, ravenshrike said:

 

You gotta think that whoever pushed that internally in Intel is either currently getting slapped upside the head or soon will be.

It did manage to prevent a noticeable amount of attention, we now have 32 core SHED CPUs and the thread has fewer replies than the 28 core thread.

13 minutes ago, Christophe Corazza said:

 

Why not? AMD is having fun at Intel's expense.

Intel being Intel, they won't let it all pass by without reacting to AMD's pleasures

Having fun does not exactly pass as an R&D budget. If AMD can get a foothold in the laptop market using Zen+ and extract revenue from there the R&D budget for Radeon Technologies can be increased. It is quite impressive AMD survives on the budget it has.

[Christophe Corazza]

5 minutes ago, ScratchCat said:

It is quite impressive AMD survives on the budget it has.

 

So lets hope AMD will keep going forward challenging Intel.

The consumer CPU market has all the benefits for AMD in doing so.

[ScratchCat]

4 minutes ago, Christophe Corazza said:

 

So lets hope AMD will keep going forward challenging Intel.

The consumer CPU market has all the benefits for AMD in doing so.

Someone really should make a graph of core count/ YoY performance gain vs time and mark the periods AMD was competitive. I would be surprised if the successes of AMD lead to sudden performance gains.

[Mira Yurizaki]

4 minutes ago, ScratchCat said:

Someone really should make a graph of core count/ YoY performance gain vs time and mark the periods AMD was competitive. I would be surprised if the successes of AMD lead to sudden performance gains.

I sort of did that in a blog post. I may need to go back to it and really examine the data though.

 

But the short of it was that I wanted to see if these so-called "incremental improvements" were always a thing and people were basically complaining the sun is hot and needs to stop (also I can't think of a good metaphor right now)

Edited June 6, 2018 by M.Yurizaki

[Christophe Corazza]

1 minute ago, ScratchCat said:

Someone really should make a graph of core count/ YoY performance gain vs time and mark the periods AMD was competitive. I would be surprised if the successes of AMD lead to sudden performance gains.

 

I'm not saying that Intel's past consumer CPUs performance increases were caused by AMD's successes trying to chase/hunt down Intel.

However, I truly hope that now, there will be more competition on the CPU market.

[ravenshrike]

10 minutes ago, ScratchCat said:

It did manage to prevent a noticeable amount of attention, we now have 32 core SHED CPUs and the thread has fewer replies than the 28 core thread.

Sure, but the 32 core TRs aren't being released for another 3 months. And when they are finally released you'll get all that hype anyway. Meanwhile the second link for 28 core 5 ghz part is the toms hardware article which has been clearly updated to expose the sham. Word of mouth about the implementation will get around as well. So when the new TRs are released, you'll have all the publicity you want, and the discussion around the 28 core Intel part will almost certainly about how the game was rigged and not about the part itself.

[ScratchCat]

2 minutes ago, M.Yurizaki said:

I sort of did that in a blog post. I may need to go back to it and really examine the data though.

 

But the short of it was that I wanted to see if these so-called "incremental improvements" were always a thing and people were basically complaining the sun is hot and needs to stop

IPC only. It is hard to tell from this graph but it seems that the gradient for 2015-2017 is slightly less than 2006-2013. It is also worth noting that this is a linear graph rather than a log graph, therefore the percentage improvement must be decreasing each year if the line on this graph is linear.

[Christophe Corazza]

3 minutes ago, ravenshrike said:

the discussion around the 28 core Intel part will almost certainly about how the game was rigged and not about the part itself.

 

pretty much spot on 

Although... Intel's own fault

[ScratchCat]

4 minutes ago, ravenshrike said:

Sure, but the 32 core TRs aren't being released for another 3 months. And when they are finally released you'll get all that hype anyway. Meanwhile the second link for 28 core 5 ghz part is the toms hardware article which has been clearly updated to expose the sham. Word of mouth about the implementation will get around as well. So when the new TRs are released, you'll have all the publicity you want, and the discussion around the 28 core Intel part will almost certainly about how the game was rigged and not about the part itself.

Generally announcements produce more hype than releases. For example Apple releases the iPhones within a week of the announcement when people still are interested while in comparison most other companies release their products weeks or months later when no one really cares and therefore loosing attention. 

 

"You can buy this amazing product now!" is more convincing than "You know that amazing product you heard of 3 months ago, you can now buy it".

[Mira Yurizaki]

13 minutes ago, ScratchCat said:

IPC only. It is hard to tell from this graph but it seems that the gradient for 2015-2017 is slightly less than 2006-2013. It is also worth noting that this is a linear graph rather than a log graph, therefore the percentage improvement must be decreasing each year if the line on this graph is linear.

The biggest question to me starts to become "why did those huge jumps happen?" Though only thing that jumps out at me is that Prescott and Bulldozer have narrower execution architectures. Similar jumps in performance happened when architectures went wide like Apple's A series and in some respect, NVIDIA's Denver with regards to ARM processors.

 

But then the burning question is "if going wide helped, how come going even wider doesn't help more?" Though I'm sure you already have an idea of why.

[ScratchCat]

41 minutes ago, M.Yurizaki said:

The biggest question to me starts to become "why did those huge jumps happen?" Though only thing that jumps out at me is that Prescott and Bulldozer have narrower execution architectures. Similar jumps in performance happened when architectures went wide like Apple's A series and in some respect, NVIDIA's Denver with regards to ARM processors.

 

But then the burning question is "if going wide helped, how come going even wider doesn't help more?" Though I'm sure you already have an idea of why.

Most of the jumps seem to align with complete redesigns which would align with the mentioned increase in width. At least in the mobile space these improvements seemed to occur when the previous architecture was reaching it's limit (A7 -> A53) or attempting to obtain a performance crown (M2 -> M3).

 

(Not sure on CPU architecture) From what I can understand wider architectures can only execute independent instructions in separate pipelines to increase performance, therefore using wider and wider architectures results in the probability that some pipelines are not filled resulting in lost performance. Is this called Instruction Level Parallelism?

[Mira Yurizaki]

9 minutes ago, ScratchCat said:

(Not sure on CPU architecture) From what I can understand wider architectures can only execute independent instructions in separate pipelines to increase performance, therefore using wider and wider architectures results in the probability that some pipelines are not filled resulting in lost performance. Is this called Instruction Level Parallelism?

Yeah pretty much. You can only make a pipe so wide before making it wider doesn't do anything.

 

Though I did see some other improvements that may have helped between Prescott and Conroe in the form of Conroe having faster SSE execution which Cinebench may be using. I'd argue we'd need tests from other workloads to get a better picture though.

[ScratchCat]

49 minutes ago, M.Yurizaki said:

Though I did see some other improvements that may have helped between Prescott and Conroe in the form of Conroe having faster SSE execution which Cinebench may be using. I'd argue we'd need tests from other workloads to get a better picture though.

It seems like Conroe vastly improved the memory subsystem, reducing latency by almost half and more than doubling the bandwidth

Quote

With the numbers available to us now, we have reason to believe that the Athlon 64 X2's latency advantage will shrink to only 15 to 20%. For comparison, the memory subsystem of the Pentium 4 was almost twice as slow as the Athlon 64 (80-90 ns versus 45-50 ns).

Fused instructions/Macro Ops accounted for 11% of the improvement according to Intel

Quote

When two x86 instructions are fused together, the 4 decoders can decode 5 instructions in one cycle. The fused instruction travels down the pipeline as a single entity, and this has other advantages: more decode bandwidth, less space taken in the Out of Order (OoO) buffers, and less scheduling overhead. If Intel's "1 out of 10" claims are accurate, macro-ops fusion alone should account for an 11% performance boost relative to architectures that lack the technology.

Quote

 However, the load and execute SSE/SSE2 operations can be fused on Core, while this is not the case on K8: packed SSE operations result in two macro-ops.

So how do Intel's Core and AMD's Hammer compare when it comes to decoding? It is hard to say at the moment without access to Intel's optimization manuals. However, we can get a pretty good idea. In almost every situation, the Core architecture has the advantage. It can decode 4 x86 instructions per cycle, and sometimes 5 thanks to x86 fusion. AMD's Hammer can do only 3.

Other improvements I can find are reduced misprediction penalty, increased fetch bandwidth, more decoders and more SSE units:

Quote

https://www.anandtech.com/show/1998/4

[TheRandomness]

10 minutes ago, Pangea2017 said:

If you need 100W more then a 1080 for simular performance it is bad perfomance and with prices higher then a 1080 there is no compelling argument for the Vega crads for the average gamer.

The RX480/RX580 had also a higher tdp then nvidia but perform simular to nvidia at the price point they where sold which made them comparable.

No, if you need 100W more than a 1080 for similar performance it's called having a bad performance per watt. The performance of Vega 64 was quite decent, but the performance of all Vega cards isn't what's bringing them down.

[bcredeur97]

nice. but non consumer part

[ScratchCat]

11 hours ago, TheRandomness said:

No, if you need 100W more than a 1080 for similar performance it's called having a bad performance per watt. The performance of Vega 64 was quite decent, but the performance of all Vega cards isn't what's bringing them down.

The performance efficiency (Performance/Theoretical compute performance) was also miserable. Vega 64 could reach 12.5 TFLOPS while a 1080 Ti could reach 11 TFLOPS while in games the 1080 Ti beat Vega. In some games like Doom Vega could show off it's prowess however that is only one game.

Even with the inclusion of HBM which was supposed to use less power than GDDR5(X) Vega uses more power than Pascal, Vega's efficiency is truly miserable.

 

[cj09beira]

34 minutes ago, ScratchCat said:

The performance efficiency (Performance/Theoretical compute performance) was also miserable. Vega 64 could reach 12.5 TFLOPS while a 1080 Ti could reach 11 TFLOPS while in games the 1080 Ti beat Vega. In some games like Doom Vega could show off it's prowess however that is only one game.

Even with the inclusion of HBM which was supposed to use less power than GDDR5(X) Vega uses more power than Pascal, Vega's efficiency is truly miserable.

 

Vega 64's power efficiency is not great while gaming not the same thing as te architecture being inefficient, because what happened there is the combination of node not clocking as high as expected and gcn not scaling well past 50Cus, for example a vega with polaris specs would destroy the 1060 and would probably be a 1070 contender, vega also has many more features that pascal doesn't which use energy to run, vega is a compute monster first and a gaming gpu second 

[laminutederire]

40 minutes ago, ScratchCat said:

The performance efficiency (Performance/Theoretical compute performance) was also miserable. Vega 64 could reach 12.5 TFLOPS while a 1080 Ti could reach 11 TFLOPS while in games the 1080 Ti beat Vega. In some games like Doom Vega could show off it's prowess however that is only one game.

Even with the inclusion of HBM which was supposed to use less power than GDDR5(X) Vega uses more power than Pascal, Vega's efficiency is truly miserable.

 

To be fair we should ask game dev to optimise their games like doom is regardless of what amd or Nvidia offers. Part of the story is that games on pc are made to leverage nvidias arch. Despite how how good another arch might be, the software will still reduce performance significantly if not done right. That's also why Nvidia is not willing to change of architecture at all.

[Taf the Ghost]

One random thought, looking at the original link:

 

The node shrink allows for double the density and 2x the power efficiency, while clocking higher. So the 480/580 replacement, Navi, is going to be something of a reworked Vega (hopefully they've worked hard on those memory bandwidth bottlenecks) and just shrunk. The Polaris replacement is going to be a pretty fast GPU.

[TechyBen]

Depends how much is 7nm... as with CPUs, apparently a LOT is not at the smallest headline measurement, and even that measurement may only be half true (measuring half of something to get 10nm, when a 12nm part is about the same size, to "fudge" the results).

 

So the proof of the pudding will be in the eating.

[ScratchCat]

45 minutes ago, cj09beira said:

Vega 64's power efficiency is not great while gaming not the same thing as te architecture being inefficient.

They correlate and for acceptable Frequency/Voltage domains can be assumed to be the same, if the architecture does more work than required it results in the same outcome - lower performance at the same power. I do not know if the issue lies with the game engines, the drivers or GCN not being able to cope with so many CUs are you said.

[Taf the Ghost]

19 minutes ago, ScratchCat said:

They correlate and for acceptable Frequency/Voltage domains can be assumed to be the same, if the architecture does more work than required it results in the same outcome - lower performance at the same power. I do not know if the issue lies with the game engines, the drivers or GCN not being able to cope with so many CUs are you said.

RX Vega has to clock outside of the silicon's efficiency range. That's what does a chunk of it in. 

 

 

Vega 56 vs 1070 Ti. 1070 Ti clocks in at 8% faster in the games tested (at 1440p), but only about 10% less power draw... at stock. OC'd, Vega kind of has the run away power issue, almost exactly like Ryzen does on the exact same node.

 

Nvidia was on a better node this generation, HBM2 prices were about 2x what AMD expected and Mining means almost no gamers are buying the parts. Some products just go sideways.

 

I'm still rather curious if Nvidia is going to be on the exact same node as AMD for their 7nm generation. (It appears Nvidia is staying on a tweaked 16nm.) Can't blame them given the yields and that AMD isn't brining out a high-end GPU until 2020 at the earliest.