Vega GPU benchmarks emerge, revealing several GPU specifications

[WYP]

SiSoft benchmarks for AMD's upcoming Vega GPU have been revealed, showcasing some impressive performance and reveals some detailed specifications.



Read more on AMD's RX Vega Leaked benchmarks.

 

[Shingara]

I need to check something with the wizkids in here who know all the black magic. According to what we have been told the 8gb's of vram is a minimum, as in thats whats on the card, we can put additional memory to it via the memory controller built into vega, is that about right.

Also i was following a discussion about infinity fabric on another forums trying to learn about it and how it actually works, and something that came up there is how ryzen uses it and they were on about how vega may connect to ryzen via infinity fabric.

So im wondering if anyone can actually explain how if it is infact going todo that when teamed with a ryzen and what are the benefits of that.

 

[WYP]

I need to check something with the wizkids in here who know all the black magic. According to what we have been told the 8gb's of vram is a minimum, as in thats whats on the card, we can put additional memory to it via the memory controller built into vega, is that about right.

Also i was following a discussion about infinity fabric on another forums trying to learn about it and how it actually works, and something that came up there is how ryzen uses it and they were on about how vega may connect to ryzen via infinity fabric.

So im wondering if anyone can actually explain how if it is infact going todo that when teamed with a ryzen and what are the benefits of that.

In simple terms, Vega is designed to be able to easily access memory outside of their frame buffer, effectively allowing the GPU to use DRAM and even system memory to create a larger frame buffer.

In a VRAM overspend scenario, Vega is designed to keep all of the most important info in the GPU's onboard HBM memory, keeping other data in external storage, effectively allowing the GPU to use larger datasets with minimal performance degradation.

We speak about this and several other Vega features in out Vega GPU architectural analysis below.

https://www.overclock3d.net/news/gpu_displays/amd_vega_gpu_architectural_analysis/1

Vega essentially has a very smart/innovative solution for using external storage as a frame buffer with a minimal performance loss. This feature is most useful for compute applications, though it will be very useful for games that require lots of VRAM.

Imagine a user of an older GPU, perhaps a 3GB or 2GB model, who frequently runs out of VRAM in some modern games but have plenty of GPU performance. This will allow those users to get past certain traditional VRAM limitations, which will be handy for gamers moving forward.

I hope this helps.

 

[Shingara]

Thanks thats very informative, so in essence and ill pick what i think is going tobe a very intensive game being star citizen, very high poly counts and large maps. Vega could be in a position where when your gaming with it vega could throw data that it doesnt need right away like the other side of a galaxy into vram that isnt on the gpu and if you zap to there then vega can flush its vram and pull the data direct to it from outside its card vram to load that faster.

That about right ?

 

[AlienALX]

In simple terms, Vega is designed to be able to easily access memory outside of their frame buffer, effectively allowing the GPU to use DRAM and even system memory to create a larger frame buffer.

In a VRAM overspend scenario, Vega is designed to keep all of the most important info in the GPU's onboard HBM memory, keeping other data in external storage, effectively allowing the GPU to use larger datasets with minimal performance degradation.

We speak about this and several other Vega features in out Vega GPU architectural analysis below.

https://www.overclock3d.net/news/gpu_displays/amd_vega_gpu_architectural_analysis/1

Vega essentially has a very smart/innovative solution for using external storage as a frame buffer with a minimal performance loss. This feature is most useful for compute applications, though it will be very useful for games that require lots of VRAM.

Imagine a user of an older GPU, perhaps a 3GB or 2GB model, who frequently runs out of VRAM in some modern games but have plenty of GPU performance. This will allow those users to get past certain traditional VRAM limitations, which will be handy for gamers moving forward.

I hope this helps.

From what I could see AMD's drivers are already doing something like that now with the Fury cards on 4gb HBM. It starts out well, until you use your paging file, then it all slows to a crawl.

Having said that it's better than a few months ago where the system would either lock up or reboot itself.

 

[WYP]

Thanks thats very informative, so in essence and ill pick what i think is going tobe a very intensive game being star citizen, very high poly counts and large maps. Vega could be in a position where when your gaming with it vega could throw data that it doesnt need right away like the other side of a galaxy into vram that isnt on the gpu and if you zap to there then vega can flush its vram and pull the data direct to it from outside its card vram to load that faster.

That about right ?

If a game does not need more than the GPUs reported frame buffer of 8GB, nothing will change from traditional memory architectures.

In a scenario where a larger frame buffer is needed the VRAM will work as a high-level cache, where all the most frequently accessed data is stored.

In many games, the most accessed data is nowhere near as large as the amount of VRAM that is typically filled in a GPU (see image above). If more than 8GB of VRAM is allocated the extra data will be moved to external storage, like DRAM or an M.2 SSD, though Vega will keep all of the most accessed data on the GPU's VRAM.

All the additional storage can then directly speak to the GPU, though at this time the exact mechanism, latency penalty or performance loss is unknown. That being said the solution will be faster than traditional methods, which usually slow games to a crawl.

Below is an image that AMD shown to show the performance change in a memory overspend scenario using this technique.

Before now, when a GPU is in a VRAM overspend the solution is not anywhere near as smart, not being as selective for what data is kept in VRAM and how external storage is used/interacts with GPUs.

This is why if you force big VRAM overspend in games that they usually lower to stupidly low framerates

Sadly AMD has not revealed exactly how this new memory architecture functions, so sadly I cannot go into any more details as I simply do not know any of the deeper details of Vega's functions.

 

[Shingara]

So ram speeds could be as important to vega as its starting to turn out it is for ryzen. Back onto the infinity fabric thing also, an ye i know thats a broad spectrum of things under one banner.

But as we stand now infinity fabric runs at something like a 1/2 ram clocks, so faster ram means faster xfer from vram to external vram and so on and so forth. But the people i was reading on anandtech forums where saying how they interprit the info vega could do infinity connection instead of pcie gen 3, now thats where i disconect.

How could you do infinity fabric over the lanes connected direct to ryzen as they are pcie gen3 lanes, or is it simply the algorythm going over the direct connection ?

 

[AlienALX]

Sadly AMD has not revealed exactly how this new memory architecture functions, so sadly I cannot go into any more details as I simply do not know any of the deeper details of Vega's functions.

I hope it works on their existing cards. That would be a right Screaming Lord Such.

 

[Shingara]

I hope it works on their existing cards. That would be a right Screaming Lord Such.

wouldnt that require the card in question to have that memory controller on it like vega is going to have.

 

[WYP]

So ram speeds could be as important to vega as its starting to turn out it is for ryzen. Back onto the infinity fabric thing also, an ye i know thats a broad spectrum of things under one banner.

But as we stand now infinity fabric runs at something like a 1/2 ram clocks, so faster ram means faster xfer from vram to external vram and so on and so forth. But the people i was reading on anandtech forums where saying how they interprit the info vega could do infinity connection instead of pcie gen 3, now thats where i disconect.

How could you do infinity fabric over the lanes connected direct to ryzen as they are pcie gen3 lanes, or is it simply the algorythm going over the direct connection ?

TBH it is impossible to know until people have product on hand. AMD could call it infinity fabric though it could easily just have been given the same name despite very different functionality.

It would be great to know more, but it will be really hard to tell what the limitations will be.

 

[Shingara]

TBH it is impossible to know until people have product on hand. AMD could call it infinity fabric though it could easily just have been given the same name despite very different functionality.

It would be great to know more, but it will be really hard to tell what the limitations will be.

Well the infinity fabric im refering to is whats in ryzen, thats what does all the sensemi and the interconnect between the 2 core groups in ryzen. I just dont see amd using the same name for 2 different things, espec when they got this tech from buying someone i believe, and its sort of there secret sauce.

Just at the mo everything is trying to run it by whacking it with a sledgehammer cos bio, drivers os are so raw.

 

[AlienALX]

wouldnt that require the card in question to have that memory controller on it like vega is going to have.

I don't know. From what I remember with Fury the memory controller was on the chip itself, which was why they stunk for overclocking. Apparently as soon as you touch the core you also mess with the HBM speed, which made the cards bomb out fast. And there was no way to separate it, which may be the same way they do Vega?

It's like Mark said though, try getting the info out of them.... I'll just bench my card etc when they release the driver tbh.

 

[Shingara]

I don't know. From what I remember with Fury the memory controller was on the chip itself, which was why they stunk for overclocking. Apparently as soon as you touch the core you also mess with the HBM speed, which made the cards bomb out fast. And there was no way to separate it, which may be the same way they do Vega?

It's like Mark said though, try getting the info out of them.... I'll just bench my card etc when they release the driver tbh.

Well im thinking it wont cos older cards cant cant farm out its data to other sources, that said i can see people getting the old furys and nanos etc just to stick in to give the vega more vram, thats if there cheap as chips on ebay.

 

[AlienALX]

Well im thinking it wont cos older cards cant cant farm out its data to other sources, that said i can see people getting the old furys and nanos etc just to stick in to give the vega more vram, thats if there cheap as chips on ebay.

It depends how AMD have laid out Fury and Vega I guess. If they both work in the same way then it shouldn't be too hard to get Fury to do the same. I mean, if they can get it to read from the paging file (which they have on Fury) then I guess they could get it to use resources from elsewhere?

We'll see I guess.

 

[Shingara]

It depends how AMD have laid out Fury and Vega I guess. If they both work in the same way then it shouldn't be too hard to get Fury to do the same. I mean, if they can get it to read from the paging file (which they have on Fury) then I guess they could get it to use resources from elsewhere?

We'll see I guess.

Well its the hbcc that controls it i think, vega has that, r9 does not.

 

[AlienALX]

Ah OK well if it's a new thing then I guess it won't work with the Fury X.

 

[Shingara]

Well it is odd that you are noticing the fury trying todo it, that could mean if the drivers are so far along that vega could be alot closer then we think. If that is the case then i do wonder where amd would announce it or how they would.

Maybe let the 1080ti settle in and retune vega to compensate. The fact we are getting leaked benchmarks right now is a good sign though.

 

[Greenback]

what would be funny is people go out buy a 1080ti and vega comes and stomps over it at £450, I know fantasy and all that

 

[AlienALX]

what would be funny is people go out buy a 1080ti and vega comes and stomps over it at £450, I know fantasy and all that

It needs to be at least as fast as a 1080. The Fury X is a good rival to the 1070 in DX12 today, because of Async. So it needs to be faster, otherwise what's the point?

I've never really thought about that until now, but yes, if it comes along with 1070 performance then it's pretty much going to be a Fury X with more VRAM.

A huge side step.

So with any luck yeah, it will offer 1080 performance for less and I will be able to jump onto that

It's 1080 performance I am looking for, I'm just not paying £500 for it.

 

[NeverBackDown]

TBH it is impossible to know until people have product on hand. AMD could call it infinity fabric though it could easily just have been given the same name despite very different functionality.

It would be great to know more, but it will be really hard to tell what the limitations will be.

Infinity fabric is the bread and butter for AMD now. They designed it in mind from the start with Vega. They are moving towards an equalibrim essentially with everything they make is "scalable" and modular to end up with ultra power APUs. This is why Vega has the ability to access dram outside of itself and be more aware. Another form of senseMI. This would be the first step for them to achieve a more coherent APU that is smart and aware of all the resources at it's disposal