• Some users have recently had their accounts hijacked. It seems that the now defunct EVGA forums might have compromised your password there and seems many are using the same PW here. We would suggest you UPDATE YOUR PASSWORD and TURN ON 2FA for your account here to further secure it. None of the compromised accounts had 2FA turned on.
    Once you have enabled 2FA, your account will be updated soon to show a badge, letting other members know that you use 2FA to protect your account. This should be beneficial for everyone that uses FSFT.

AMD's SSG technology - are we interested?

Its a 9999$ product and all it essentially does is to add an M.2 slot on the card. Nothing revolutionary over that. It seems the entire reason based out in the memory limit they have with 32GB and then add an SSD slow storage to supplement it. I have trouble seeing the customers for it.

AMD Radeon Pro SSG，加入 M.2 SSD 插槽的 Polaris 10 專業繪圖卡

But we will know more next week.
 
Last edited:
it basically takes needless overhead out of processes that already exist. while not revolutionary it does help the industries it's targeted at quite a bit. even though the article tries to claim it's useful for consumers i really don't see it yet, maybe in 5-6 years, but it'll have a ton of use for the animated movie sector.
 
Semiaccurate article.
AMD puts massive SSDs on GPUs and calls it SSG - SemiAccurate

This is the most revolutionary idea in years to me. Take a SSD and give it direct access to the GPU. Don't need the CPU to do a thing after your project or game textures are loaded to the GPU's storage. No more streaming textures. No more pop in. Imagine if the little iGPU on an Intel chip is steadily loading the next area to the GPU's memory cache so that you can have massive worlds with 8K textures and never see a load screen? Imagine if the first RAGE game had this?

I could see a VR game where you create your own vehicle and then hop into a massive persistent world with thousands of others. Why would games need decal cleanup? Imagine Skyrim with the bones of an ice tiger that you killed at the beginning of the game lying on the ground forever? Come back years later in game time to a flashback and collect those bones to use as a memorial to the life you used to live. Since the GPU would have such massive storage, you could load the village with various other player's decals and memorials loaded into memory long before you even get close to entering the area where you would be able to see them in game. No pop in; just smooth gaming nirvana.

This would work really well in a game world like Star Citizen or Eve Online. Imagine an armada of 10,000 custom ships attacks a base. I'm talking about ships with hand drawn decals on them; not variants of 20 decals you can choose at character creation. Someone with this technology could load all of the textures into local GPU memory so they would never lag due to seeing a new enemy. Instead of seeing a standard enemy and then their textures loading 45 seconds later for you to realize that you were fighting the Red Baron in his custom Fokker, you get this information stored before the battle commenced.

What if there is a sleuth game and you can unlock clues to what another player did earlier in the game? Say you are clairvoyant and can see into the past and the future. Your client loads the criminal's replay into local GPU memory so that when you're dusting down the area for clues you suddenly get images and flashbacks to when the other player committed a crime 50 years prior in his game world. The criminal is trying to commit the crime and leave the least amount of clues to initiate a replay. You have saved multiple tracks that he took into memory so that there is no loading. While you're searching for clues, the other player enters your space, but in his own timeline 50 years prior. In real time you can see him for a momentary flash of time; you collapse clutching your head. The criminal player suddenly finds himself in your timeline and already has your replays loaded into GPU memory. He then has to pretend to blend into the crowd as a 70 year old man, but at the same time alter your clues to stop you from discovering who done it.

I could see lots of uses for this technology. I think it's great!
 
This would be amazing for GPU processing of video.

Wonder what it can do with mGPU, after practically making the pci bus devoid of major texture streaming?

This could be damn huge. Cards of 'xgb' will be an outdated thing in short order if it takes hold. Having a smaller HBM local memory bus also wouldn't be so much of an issue...


edit: Now we know what 'next gen memory' is for on Navi.... it's this!
 
Last edited:
N4CR said:
This would be amazing for GPU processing of video.

Wonder what it can do with mGPU, after practically making the pci bus devoid of major texture streaming?

This could be damn huge. Cards of 'xgb' will be an outdated thing in short order if it takes hold. Having a smaller HBM local memory bus also wouldn't be so much of an issue...


edit: Now we know what 'next gen memory' is for on Navi.... it's this!
Click to expand...

i think its nexgen? NexGen :: Tech Specs
 
cageymaru said:
Semiaccurate article.
AMD puts massive SSDs on GPUs and calls it SSG - SemiAccurate

This is the most revolutionary idea in years to me. Take a SSD and give it direct access to the GPU. Don't need the CPU to do a thing after your project or game textures are loaded to the GPU's storage. No more streaming textures. No more pop in. Imagine if the little iGPU on an Intel chip is steadily loading the next area to the GPU's memory cache so that you can have massive worlds with 8K textures and never see a load screen? Imagine if the first RAGE game had this?

I could see a VR game where you create your own vehicle and then hop into a massive persistent world with thousands of others. Why would games need decal cleanup? Imagine Skyrim with the bones of an ice tiger that you killed at the beginning of the game lying on the ground forever? Come back years later in game time to a flashback and collect those bones to use as a memorial to the life you used to live. Since the GPU would have such massive storage, you could load the village with various other player's decals and memorials loaded into memory long before you even get close to entering the area where you would be able to see them in game. No pop in; just smooth gaming nirvana.

This would work really well in a game world like Star Citizen or Eve Online. Imagine an armada of 10,000 custom ships attacks a base. I'm talking about ships with hand drawn decals on them; not variants of 20 decals you can choose at character creation. Someone with this technology could load all of the textures into local GPU memory so they would never lag due to seeing a new enemy. Instead of seeing a standard enemy and then their textures loading 45 seconds later for you to realize that you were fighting the Red Baron in his custom Fokker, you get this information stored before the battle commenced.

What if there is a sleuth game and you can unlock clues to what another player did earlier in the game? Say you are clairvoyant and can see into the past and the future. Your client loads the criminal's replay into local GPU memory so that when you're dusting down the area for clues you suddenly get images and flashbacks to when the other player committed a crime 50 years prior in his game world. The criminal is trying to commit the crime and leave the least amount of clues to initiate a replay. You have saved multiple tracks that he took into memory so that there is no loading. While you're searching for clues, the other player enters your space, but in his own timeline 50 years prior. In real time you can see him for a momentary flash of time; you collapse clutching your head. The criminal player suddenly finds himself in your timeline and already has your replays loaded into GPU memory. He then has to pretend to blend into the crowd as a 70 year old man, but at the same time alter your clues to stop you from discovering who done it.

I could see lots of uses for this technology. I think it's great!
Click to expand...
Last time i checked SSDs are slow, in comparison to actual RAM, so if anything it just shortens the possible lag from loading that on CPU-side, not replaces VRAM, and only on data sets so large, nobody worth a damn will make a game with them. I can see how it can be fairly useful in raw video processing situations, especially since apparently dev board is Fiji-based (cue the 4gb VRAM jokes), though.

I like your fantasy though, even if it's a bit optimistic on what this can or cannot do.
Sonicks said:
The author there seems overly excited over his own speculation.

I doubt this will work at all this way. This seems more beneficial for video playback or rendering of high resolution video. I do not see now this can magically improve games the way he talks about it.
Click to expand...
Cut Charlie some slack, nV was wrecking his "predictions" so hard lately, he has to be excited about something for once.
 
lolfail9001 said:
Last time i checked SSDs are slow, in comparison to actual RAM, so if anything it just shortens the possible lag from loading that on CPU-side, not replaces VRAM, and only on data sets so large, nobody worth a damn will make a game with them. I can see how it can be fairly useful in raw video processing situations, especially since apparently dev board is Fiji-based (cue the 4gb VRAM jokes), though. I like your fantasy though, even if it's a bit optimistic on what this can or cannot do.
Click to expand...

Actually it depends on caching algorithms in the end how the SSD function is not based on what you get in retail since there is no "bus" as what we consumers get there is no need for certain protocols which would kill bandwidth you could just throw in the maximum throughput of the hardware without other limitations that you have on "bus" based solutions

This has nothing to do with fantasy it is not even science fiction it is a smart approach to get by current day limitations...
 
Pieter3dnow said:
Actually it depends on caching algorithms in the end how the SSD function is not based on what you get in retail since there is no "bus" as what we consumers get there is no need for certain protocols which would kill bandwidth you could just throw in the maximum throughput of the hardware without other limitations that you have on "bus" based solutions

This has nothing to do with fantasy it is not even science fiction it is a smart approach to get by current day limitations...
Click to expand...
I am fairly positive that what AMD proposes is literally a direct GPU-to-SSD PCI-E bus (maybe even utilizing PLX switch between PCI-E slot and M.2 slots on card), and that bandwidth is still bad in comparison to even PCI-E x16 bus to system memory (heck, take any [h]igh end PC and compare PCI-E x4 SSD and ramdisk in bandwidth and latency).

Obviously, it's a fairly smart approach to surpass capacity limitations of modern GPUs, but whether it's good enough outside of these freak cases like raw video processing, remains to be seen.

Also, did i mention that with this SSG release what Nexgen means on AMD's roadmap is perfectly clear?
 
Big data and ultra high res developers should be super excited about this. Normal consumers. . . not so much. Although, I could see something like this (local ultra fast storage on graphics board) come to consumers, it probably won't happen for years and will be in a very different solution than this.
 
lolfail9001 said:
I am fairly positive that what AMD proposes is literally a direct GPU-to-SSD PCI-E bus (maybe even utilizing PLX switch between PCI-E slot and M.2 slots on card), and that bandwidth is still bad in comparison to even PCI-E x16 bus to system memory (heck, take any [h]igh end PC and compare PCI-E x4 SSD and ramdisk in bandwidth and latency).

Obviously, it's a fairly smart approach to surpass capacity limitations of modern GPUs, but whether it's good enough outside of these freak cases like raw video processing, remains to be seen.

Also, did i mention that with this SSG release what Nexgen means on AMD's roadmap is perfectly clear?
Click to expand...

If what you're saying is true then people with 128GB, 256GB, 512GB+, etc of system memory would have this technology already. Guess what? They are the 17 fps crowd in the demonstration where the GPU is completely choked out by waiting on calls from the rest of the components in the system. I bet that crowd also has entire 1TB or bigger SSD scratch disks dedicated to Adobe or CAD. Still only 17 fps tops. AMD SSG technology surpasses all of that with 93 fps. So it is obvious that it is much faster for the GPU to talk to it's own SSD using SSG. It is obvious that this is different from just adding a SSD to the system as some may think; not you necessarily. GPUs have been loading textures off SSDs for years and never got this type of performance. Install a game onto your C: which is most likely a SSD nowadays to test this out.

I wasn't being futuristic about the decals. When I played APB back in the day we all had custom hand created decals. Our gear was full of custom hand drawn decals. When we fought a person, all we would see would be Generic Human A until we had died 4 or 5 times in team gang warfare. Then we would be like, "Oh sh*t! We're fighting the biggest crew! I just got dusted by Big Todd" In my example with SSG technology, we would have known who we were fighting right at the onset because their custom decals would have been loaded to the GPU as soon as we entered the zone.
 
cageymaru said:
If what you're saying is true then people with 128GB, 256GB, 512GB+, etc of system memory would have this technology already. Guess what? They are the 17 fps crowd in the demonstration where the GPU is completely choked out by waiting on calls from the rest of the components in the system. I bet that crowd also has entire 1TB or bigger SSD scratch disks dedicated to Adobe or CAD. Still only 17 fps tops. AMD SSG technology surpasses all of that with 93 fps. So it is obvious that it is much faster for the GPU to talk to it's own SSD using SSG. It is obvious that this is different from just adding a SSD to the system as some may think; not you necessarily. GPUs have been loading textures off SSDs for years and never got this type of performance. Install a game onto your C: which is most likely a SSD nowadays to test this out.

I wasn't being futuristic about the decals. When I played APB back in the day we all had custom hand created decals. Our gear was full of custom hand drawn decals. When we fought a person, all we would see would be Generic Human A until we had died 4 or 5 times in team gang warfare. Then we would be like, "Oh sh*t! We're fighting the biggest crew! I just got dusted by Big Todd" In my example with SSG technology, we would have known who we were fighting right at the onset because their custom decals would have been loaded to the GPU as soon as we entered the zone.
Click to expand...
Okay, you are making me watch that presentation and not just read anandtech article, what is proper timestamp and source for this "17 fps tops". And yes, it is obviously faster to have 4x/8x PCI-E connect straight to SSD, but 5 times better? Demo pl0x.
 
lolfail9001 said:
Okay, you are making me watch that presentation and not just read anandtech article, what is proper timestamp and source for this "17 fps tops". And yes, it is obviously faster to have 4x/8x PCI-E connect straight to SSD, but 5 times better? Demo pl0x.
Click to expand...

Dude I watched the presentation last night with everyone else. The SSG thing was towards the end. ;)
 
cageymaru said:
Dude I watched the presentation last night with everyone else. The SSG thing was towards the end. ;)
Click to expand...
Yeah, and i have few proper questions to ask, for once:

1) What system was there playing first video? Make no mistake, it certainly was powerful, but question stands.

2) What codec was used for Fiji's video, because raw 8k@90 is 9 motherfucking gigabytes per second.

3) Finally
PCI Express 3.0's 8 GT/s bit rate effectively delivers 985 MB/s per lane
Click to expand...

So, if they were using 4x PCI-E SSD, this demo could not happen, because read speed exceeds THEORETICAL bandwidth of 4 lane PCI-E 3.0, and then following is present:

Details are a bit thin and I’ll update this later this evening, but in short the card utilizes a Polaris 10 GPU and includes 2 PCIe 3.0 M.2 slots for adding flash drives to the card. These slots are then attached to the GPU (it’s unclear if there’s a PCIe switch involved or if it’s wired directly), which the GPU can then use as an additional tier of storage. I’m told that the card can fit at least 1TB of NAND – likely limited by M.2 MLC SSD capacities – which massively increases the amount of local storage available on the card.
Click to expand...

And there does not exist an M.2 form-factor SSD that can use more than 4 lanes of PCI-E.

Is the thing great? Oh yes, it is. But i don't buy this demonstration one second.

EDIT: Okay, now my points 2 and 3 are irrelevant, i've got to an update in anand's article, it was RAID-0. Question about first system stands though.
 
Its not just about extra bandwidth and reduced latency on the disk loads (which are great). Its also about not hammering the CPU and main memory with constant paging requests.

This technology is most likely a happy coincidence from building HSA compliant GPU front ends. To coexist on the CPU the GPU also has to be able to address memory resources like a commodity CPU can. Once you've crossed that threshold its not oo much further to implementing virtual memory on-card for discrete GPUs.

I would be curious if anyone could spot extra PCIe lanes on the die shots to see if they are direct connecting them to dedicated lanes. Maybe they split the 16 PCie lanes into x8 to the mainboard and two x4s to the SSDs.
 
ecmaster76 said:
Its not just about extra bandwidth and reduced latency on the disk loads (which are great). Its also about not hammering the CPU and main memory with constant paging requests.

This technology is most likely a happy coincidence from building HSA compliant GPU front ends. To coexist on the CPU the GPU also has to be able to address memory resources like a commodity CPU can. Once you've crossed that threshold its not oo much further to implementing virtual memory on-card for discrete GPUs.

I would be curious if anyone could spot extra PCIe lanes on the die shots to see if they are direct connecting them to dedicated lanes. Maybe they split the 16 PCie lanes into x8 to the mainboard and two x4s to the SSDs.
Click to expand...
Yeah, the lift of load on CPU is the biggest one, still makes it irrelevant outside of REALLY big data sets.

According to anandtech's article, it uses PLX switch.
 
lolfail9001 said:
Yeah, and i have few proper questions to ask, for once:

1) What system was there playing first video? Make no mistake, it certainly was powerful, but question stands.

2) What codec was used for Fiji's video, because raw 8k@90 is 9 motherfucking gigabytes per second.

3) Finally


So, if they were using 4x PCI-E SSD, this demo could not happen, because read speed exceeds THEORETICAL bandwidth of 4 lane PCI-E 3.0, and then following is present:



And there does not exist an M.2 form-factor SSD that can use more than 4 lanes of PCI-E.

Is the thing great? Oh yes, it is. But i don't buy this demonstration one second.

EDIT: Okay, now my points 2 and 3 are irrelevant, i've got to an update in anand's article, it was RAID-0. Question about first system stands though.
Click to expand...

I wasn't physically there. You would have to ask AMD. I didn't go into it looking for holes in their story, so I assumed that both systems were supposed to be high end and equal with the exception of one card having SSG technology.
 
I could see it as a boon for games. My Skyrim has 25Gb in textures alone. Having them GPU side would make sense.
 
It's a cool idea at least. I'd love to see some compute based benchmarks and or an API on it.
 
SSG presentation starts at 58:32. SSG live showing is at 1hr 5 mins in.

 
Shintai said:
Its a 9999$ product and all it essentially does is to add an M.2 slot on the card. Nothing revolutionary over that. It seems the entire reason based out in the memory limit they have with 32GB and then add an SSD slow storage to supplement it. I have trouble seeing the customers for it.

AMD Radeon Pro SSG，加入 M.2 SSD 插槽的 Polaris 10 專業繪圖卡

But we will know more next week.
Click to expand...

Actually it's genius for companies that stream large amounts of data and have to recompress on the fly with large libraries. It's a win-win for companies like Amazon and Netflix. They can cut down on bandwidth and storage, and the client gets better quality video with less dropouts.
 
DigitalGriffin said:
Actually it's genius for companies that stream large amounts of data and have to recompress on the fly with large libraries. It's a win-win for companies like Amazon and Netflix. They can cut down on bandwidth and storage, and the client gets better quality video with less dropouts.
Click to expand...
For people that stream large amounts of data and stuff the bottleneck remains where it is, actually. AMD has not slapped an ethernet port on video cards yet as far as i know.
 
Que the constant downplaying of a new technology!
It never fails does it...
 
  • Like
Reactions: Remon
like this
{NG}Fidel said:
Que the constant downplaying of a new technology!
It never fails does it...
Click to expand...
Well, downplaying never fails with AMD, look at HBM and Polaris.

This tech can be pretty good in right situations, not going to deny it. These situations tho are fairly specific.
 
lolfail9001 said:
For people that stream large amounts of data and stuff the bottleneck remains where it is, actually. AMD has not slapped an ethernet port on video cards yet as far as i know.
Click to expand...

Mega data centers with large datasets don't use Ethernet ports.... The ones I've seen use customized connections which go to break out boxes. Sort of like a special hyper transport to a specialized switch.

But that's not the only application. As large amounts of new library content come online you can encode them quickly in multiple formats and then store them in secondary storage.

You're right it's highly specialized. But I can see benefits for such large customers.
 
DigitalGriffin said:
Mega data centers with large datasets don't use Ethernet ports.... The ones I've seen use customized connections which go to break out boxes. Sort of like a special hyper transport to a specialized switch.

But that's not the only application. As large amounts of new library content come online you can encode them quickly in multiple formats and then store them in secondary storage.

You're right it's highly specialized. But I can see benefits for such large customers.
Click to expand...
Yeah, correct, they don't use ethernet inside, i remember my network course now. You get the point.

But scenario you describe will be basically bus/CPU limited as is, slapping SSD on a card does little to help the scenario. But i'll agree, large customers will find large benefits.
 
Hitman928 said:
Big data and ultra high res developers should be super excited about this. Normal consumers. . . not so much. Although, I could see something like this (local ultra fast storage on graphics board) come to consumers, it probably won't happen for years and will be in a very different solution than this.
Click to expand...

For 10K for one atm, I doubt many consumers will be messing with one in the near future.

But I guess that has been mentioned.
 
There is a reason AMD did it. They aren't just stupid. They must see an advantage to this. This industry is very specific to begin with so there must be specific need for this. I think they are gearing this more towards video aren't they?
 
NKD said:
There is a reason AMD did it. They aren't just stupid. They must see an advantage to this. This industry is very specific to begin with so there must be specific need for this. I think they are gearing this more towards video aren't they?
Click to expand...
Basically AMD decided to make a large-margin product for a change, and coincidentally this is one idea that could work and is specific enough to get few large-margin contracts.
Unless of course nV somehow gets ahead of them even here.
 
This is a semi custom, it isn't a one solution fits all, at all.

Look, this is currently being sold for developers as a beta test because yeah, it works on the industry that commissioned it, but AMD is wondering if some other industry sees a benefit to it, if there are then it will be awesome for them, extra profits!

The main use = Oil & Gas simulation. Everything else? profit!
 
Revdarian said:
This is a semi custom, it isn't a one solution fits all, at all.

Look, this is currently being sold for developers as a beta test because yeah, it works on the industry that commissioned it, but AMD is wondering if some other industry sees a benefit to it, if there are then it will be awesome for them, extra profits!

The main use = Oil & Gas simulation. Everything else? profit!
Click to expand...

yea looks like they are making it for certain industries that are probably interested in this. I think its a win win for AMD. Like making a product that they know certain industry will buy because they will see benefit from it. Good profitable move. Looks like geared big towards content creation. I am sure those people are drooling over it.
 
Last edited:
NKD said:
There is a reason AMD did it. They aren't just stupid. They must see an advantage to this. This industry is very specific to begin with so there must be specific need for this. I think they are gearing this more towards video aren't they?
Click to expand...

Radeon Pro Duo ring a bell? Sold 1000 cards only for HPC. And pretty much 0 for anything else.

Fiji was a massive flop for HPC for multiple reasons, specially VRAM. This time they got the VRAM, but they have utterly poor performance and failing completely in the most important metric, perf/watt for the HPC TCO. So the next attempt is this we see now. They simply throw a lot of stuff and hope something sticks because they cant compete in traditional metrics.

R&D matters, and the RTG group got close to nothing in this area. They needed Vega if not Navi today. Not a Polaris chip with an SSD bolted on.
 
Shintai said:
Radeon Pro Duo ring a bell? Sold 1000 cards only for HPC. And pretty much 0 for anything else.

Fiji was a massive flop for HPC for multiple reasons, specially VRAM. This time they got the VRAM, but they have utterly poor performance and failing completely in the most important metric, perf/watt for the HPC TCO. So the next attempt is this we see now. They simply throw a lot of stuff and hope something sticks because they cant compete in traditional metrics.

R&D matters, and the RTG group got close to nothing in this area. They needed Vega if not Navi today. Not a Polaris chip with an SSD bolted on.
Click to expand...

You clearly have no grasp of what you are talking about. You portrait a company focused on 1 single side of the market where there are many more.
By your definition something as Jaguar core based cpu are a flop they did not sell they are "not worthy". Guess what is in every Xbox and PS these days , jep Jaguar core cpu and everyone was laughing back then and now it is a different story.

Stop being so narrow minded .....
 
Pieter3dnow said:
You clearly have no grasp of what you are talking about. You portrait a company focused on 1 single side of the market where there are many more.
By your definition something as Jaguar core based cpu are a flop they did not sell they are "not worthy". Guess what is in every Xbox and PS these days , jep Jaguar core cpu and everyone was laughing back then and now it is a different story.

Stop being so narrow minded .....
Click to expand...

We are not talking about low edge console products in this thread ;)
 
Shintai said:
We are not talking about low edge console products in this thread ;)
Click to expand...
Why not? You like skewing the arguments to fit your narrative. Case in point: This is a concept and you act as if it is the final product. This is thinking outside the box and you can't see beyond the walls within.
 
Last edited:
JustReason said:
Why not? You like skewing the arguments to fit your narrative. Case in point: This is a concept and you act as if it is the final product. This is thinking outside the box and you can see beyond the walls within.
Click to expand...
Can hardly call it thinking outside the box when i have seen a reddit thread asking how come ssds have not yet replaced ram few weeks prior to this announcement, though.
 
You must log in or register to reply here.
Back
Top