AMD Radeon Pro SSG - a BETA product for 10000$

[thekeemo]

Just now, Electronics Wizardy said:

There is no space for the cards to intake air form. The xeon phis are for workstations. Servers have high power fans, so a cpu or gpu fan isn't needed. There are xeon phi's that are made for servers that have no fans, like this one https://www.google.com/search?q=xeon+phi&source=lnms&tbm=isch&sa=X&ved=0ahUKEwje7rzQo5DOAhVEHGMKHdYUAuEQ_AUICygE&biw=900&bih=1537#imgrc=pKIIoryQhAeVlM%3A

I cant speak for any of this. I dont use high end gear.

[MageTank]

1 hour ago, Electronics Wizardy said:

Yep, unless its what i think it is a just using software.

 

So ssd to pcie to cpu to pcie to cpu.

 

Its goin to be slow.

 

Might as well use the m.2 slot on you motherboard.

It was brought to my attention that the Radeon Pro Duo used a PLX bridge on the board itself. It would be completely possible for them to rework a PLX bridge to make this specific "SSG" card bypass the CPU entirely. 

 

 

Notice the PLX chip on this? That is a 48 lane switch. 16 in, 16 out, 16 mirror. Reworking this for the x4 PCIE M.2 drives on that card would be fairly easy, and would allow them to bypass the CPU entirely. They would pretty much have to, in order for that "went from 17 FPS to 90+FPS" claim to be true. 

[RagnarokDel]

Super Sayan God?

[Shahnewaz]

Quote

AMD Radeon Pro SSG - a BETA product for 10000$

1 hour ago, zMeul said:

so .. what the fuck is this?

one one hand, having extra storage directly available to the GPU as a n^th cache level seems OK, until you start thinking about it - it's still slower than RAM, still slower than VRAM

to have a PCIe SSD attached to the GPU you need a controller and I seriously doubt AMD put one in the GPU, so it's external with some sort of BUS to it; and it's also available as local storage

 

we'll have to see actual real world scenarios of this card

Typical bullshit, sensationalist zMeul reporting.

1 hour ago, thekeemo said:

Why not normal ram? or laptop ram.

A terabyte of DRAM? That would be insanely expensive.

[Fetzie]

1 hour ago, Shahnewaz said:

Typical bullshit, sensationalist zMeul reporting.

A terabyte of DRAM? That would be insanely expensive.

And take something like 4000 square centimetres of pcb. The pcb would be about a metre long and half a metre wide, covered in memory chips.

[Prysin]

3 hours ago, vinyldash303 said:

That's kinda neat. 

 

3 hours ago, thekeemo said:

Why not normal ram? or laptop ram.

 

3 hours ago, Hunter259 said:

Laptop ram would be much faster and better for this task.

 

3 hours ago, Electronics Wizardy said:

Yep, unless its what i think it is a just using software.

 

So ssd to pcie to cpu to pcie to cpu.

 

Its goin to be slow.

 

Might as well use the m.2 slot on you motherboard.

 

3 hours ago, zMeul said:

I'm guessing density and price/GB

ding ding ding ding.

 

its all about space and price. The largest GDDR5 DRAM dies atm is 2GB/stack. They cost about 25-32$ each and uses about 8-15w depending on manufacturer and product node. To get 1TB you would need 500 of them. That would require a PCB the size of a ATX motherboard. with millions of times more complex trace patterns. The power delivery for such setup would require 1200w just for the GDDR5 DRAM dies when factoring in leakage.

There is also the issue of trace length. At X distance to GPU DIE, GDDR5 latency will be so high that the raw speed wont matter. Thus you will NEVER achieve a balanced memory setup with GDDR5.

 

The Firepro W9150 with 32GB of GDDR5 is already at the limit of what you can physically do with a normal PCI Express board. To be able to make a PCB able to handle more memory you would have to make a custom board of its own form factor with custom PCB, power delivery, custom memory controllers etc.

 

 

At 3000MB/s read/write for the highest end NVMe drives, and assume they use one or two PLX chips to create 2x PCIe Gen3 x4 connections, one would be able to move between 3-6GB/s of data. Depending on whether they add a Marvell or Intel RAID controller.

 

PLX chips cost around 95-140 USD a piece depending on "bulk"prices.

not sure about RAID controllers, looking at RAID AIBs i would assume 50-75 USD for a RAID controller.

 

It is not a cheap board to produce going by form factor and product features.

[3DOSH]

Use case: "Hugely complex and highly detailed CAD models that took the better part of an hour to load up and decompress will still take the better part of an hour to load up and decompress on an SSG GPU based system. Why bother? Because it takes the better part of an hour to load and decompress the first time, then it can stay resident on the GPU’s flash storage. The second time it should take seconds. "

http://semiaccurate.com/2016/07/25/amd-puts-massive-ssds-gpus-calls-ssg/

 

 

[Sakkura]

2 hours ago, zMeul said:

what SSD is the one on the right that does 4590MB/s?!

 

a 950PRO does around 700MB/s in random 4k q32 reads and peaks at about 2.5GB/s in sequential

unless they put 2 in RAID 0 and compared to a single one - that's just misrepresenting the product .. basically lying trough their teeth

That's how fast it's rendering, genius...

[ManIkWeet]

Guys guys this is obviously AMD's attempt to dominate the market by being able to install Linux on your GPU!

 

Think about the possibilities!

 

You don't have to 'load textures' when playing a game anymore, it's already available on the GPU!

[gilang01]

15 minutes ago, ManIkWeet said:

Guys guys this is obviously AMD's attempt to dominate the market by being able to install Linux on your GPU!

 

Think about the possibilities!

 

You don't have to 'load textures' when playing a game anymore, it's already available on the GPU!

the idea of loading textures to the GPU is actually kinda neat.

[zMeul]

54 minutes ago, Sakkura said:

That's how fast it's rendering, genius...

gee ... and it's scrubing from!? the fucking SSD

genius

 

Quote

as AMD demonstrated, scrubbing through an 8K video file.

 

[Notional]

3 hours ago, zMeul said:

this needs to be tested by someone else

AMD has a history of bending the truth

 

and here is what I'm talking about:

 

what SSD is the one on the right that does 4590MB/s?!

 

a 950PRO does around 700MB/s in random 4k q32 reads and peaks at about 2.5GB/s in sequential

unless they put 2 in RAID 0 and compared to a single one - that's just misrepresenting the product .. basically lying trough their teeth

If we talk general server grade PCI SSD's, then this will do 6000 MBps http://www.samsung.com/semiconductor/products/flash-storage/enterprise-ssd/MZPLK6T4HCJL?ia=832

But since is a demo of M.2's, these should easily hit that in RAID0 http://www.tomshardware.com/reviews/samsung-sm961-ssd,4608.html

 

Though 1 might be enough if combined with the onboard VRAM.

Either way, this product is tailored to a very specific need, and I'm sure it will be the best on the market for such a need. Especially if the entire thing bypasses the CPU+sysram. Interesting.

 

The only thing that sucks, is that all these new Pro cards are only based on Polaris. They are weak compared to P6000 cards from Nvidia.

4 hours ago, Electronics Wizardy said:

The cooler sucks for servers. blowers don't work for 1u servers.

Who says this is for servers and not workstations? This card might actually improve workflows at LTT, if they work on large videos and lots of material.

[HalGameGuru]

If I was gonna have multiple M.2 SSDs I know I'd wanna RAID them...

 

I have no need for this personally, but I am really interested as to the valid use cases and what kind of real world improvement it can offer.

[Sakkura]

16 minutes ago, zMeul said:

gee ... and it's scrubing from!? the fucking SSD

genius

 

 

 

It's rendering. Stop being obtuse just in order to find some way to shit on AMD.

 

Quote

 

The Radeon Pro SSG appears to be capable of faster-than-real-time rendering of 8K raw footage. The card was rendering 4.5GB/s at that pace.

 

[zMeul]

Just now, Sakkura said:

It's rendering. Stop being obtuse just in order to find some way to shit on AMD.

and it's rendering from what source!? thin air?

it's rendering form the fucking SSDs

 

in the left shot it's rendering from some PCIe installed SSD while on the right it's doing it from two RAID0 SSDs mounted on the Radeon Pro board

there is no single SSD capable of those speeds

[Sakkura]

Just now, zMeul said:

and it's rendering from what source!? thin air?

it's rendering form the fucking SSDs

 

in the left shot it's rendering from some PCIe installed SSD while on the right it's doing it from two RAID0 SSDs mounted on the Radeon Pro board

there is no single SSD capable of those speeds

Are you incapable of rational thought? Source and output don't have to be the same size.

[zMeul]

Just now, Sakkura said:

Are you incapable of rational thought? Source and output don't have to be the same size.

are you dense!?

 

they scrub trough the same source, in one instance it's doing it at 17FPS and in the other it's doing it at 90FPS - this was AMD's comparison their Radeon Pro SSG product is spectacular

the comparison is a fucking lie and a misrepresentation of the truth and this is why I said it needs to be tested by a 3rd party - AMD has a history of bending the truth in their favor

[Sakkura]

29 minutes ago, zMeul said:

are you dense!?

 

they scrub trough the same source, in one instance it's doing it at 17FPS and in the other it's doing it at 90FPS - this was AMD's comparison their Radeon Pro SSG product is spectacular

the comparison is a fucking lie and a misrepresentation of the truth and this is why I said it needs to be tested by a 3rd party - AMD has a history of bending the truth in their favor

They are rendering, not scrubbing. Can you not read?

[zMeul]

1 hour ago, Sakkura said:

They are rendering, not scrubbing. Can you not read?

quote, source Anand as I posted in the OP:

Quote

As AMD explains it, the purpose of going this route is to offer another solution to the workset size limitations of current professional graphics cards. Even AMD’s largest card currently tops out at 32GB, and while this is a fair amount, there are workloads that can use more. This is particular the case for workloads with massive datasets (oil & gas), or as AMD demonstrated, scrubbing through an 8K video file.

 

scrubbing or rendering, it makes no fucking difference as the source is the same

 

one is doing it at 846MB/s and the other is doing it at 4590MB/s - the 2nd one is not possible to achieve without 2 NVMe SSDs in RAID0

so, why didn't they used 2 NVMe SSDs in RAID0 in the 1st setup?

 

---

 

and here's why their solution is bullshit

a 16x PCIe gen 3 connection has 15,754 GB/s per direction bandwidth - thus, 1x is 984,6 MB/s

and AMD's demonstration uses 4590 MB/s - that's slightly more than a 4x gen 3 PCIe in one direction (3,938GB/s) - totally doable if we'd have SSDs that fast, as PCIe NVMe storage uses 4x PCIe connection

 

so yes, AMD lied and misrepresented their product once more

[HKZeroFive]

Why does the Radeon Pro SSG look like it's based on Fiji instead of Polaris?

 

[Sakkura]

54 minutes ago, zMeul said:

quote, source Anand as I posted in the OP:

 

scrubbing or rendering, it makes no fucking difference as the source is the same

 

one is doing it at 846MB/s and the other is doing it at 4590MB/s - the 2nd one is not possible to achieve without 2 NVMe SSDs in RAID0

so, why didn't they used 2 NVMe SSDs in RAID0 in the 1st setup?

And they are not specifically linking the 4.5 GB/s with scrubbing, unlike the source I linked, proving your claim false.

 

And you still aren't understanding the concept of source and output being different sizes.

[Misanthrope]

It's probably going to be an ssd as a level 2 cache and not accessing it directly. Sort of how an hybrid hdd that has a tiny ssd on it works.

 

Might be slightly better than going through the CPU to get to a traditional m.2 ssd but it's certainly not possible for it to just read the ssd directly they aren't fast enough yet and if they were it would cost a hell of a lot more than 10k for the SSD alone.

[Mr.Meerkat]

3 hours ago, zMeul said:

there is no single SSD capable of those speeds

Ding ding ding, looks like time you read the thread again  

(aka, look below as I quoted it for you)

3 hours ago, Notional said:

If we talk general server grade PCI SSD's, then this will do 6000 MBps http://www.samsung.com/semiconductor/products/flash-storage/enterprise-ssd/MZPLK6T4HCJL?ia=832

 

 

20 minutes ago, Misanthrope said:

It's probably going to be an ssd as a level 2 cache and not accessing it directly. Sort of how an hybrid hdd that has a tiny ssd on it works.

Which makes sense as a NVMe SSD won't have the reads and write speeds of normal vRAM while also, you won't be able to write as much to the SSDs compared to usual vRAM  

16 minutes ago, Misanthrope said:

Might be slightly better than going through the CPU to get to a traditional m.2 ssd but it's certainly not possible for it to just read the ssd directly they aren't fast enough yet and if they were it would cost a hell of a lot more than 10k for the SSD alone.

If they used PLX chips and whatever to bolt the m.2 SSDs straight on then it means that the NVMe SSDs aren't actually taking up the CPU lanes which can help when you're running like 4 of them in a single workstation while transferring lots of data between the NVMe SSD(s) as storage and the NVMe SSD(s) on the GPU(s) .

[Tomsen]

@zMeul

The issue isn't the bandwidth of reading from the source, but rather the latency and the complexity of having the CPU micromanaging it.

Sure it will not help one bit of the initial data transfer, but past that (every read thereafter), it will have a great benefit in certain data-heavy applications.

 

We are still waiting for your explaination in how they were tampering with the polaris 11 numbers.

[MageTank]

11 hours ago, Prysin said:

 

 

 

 

ding ding ding ding.

 

its all about space and price. The largest GDDR5 DRAM dies atm is 2GB/stack. They cost about 25-32$ each and uses about 8-15w depending on manufacturer and product node. To get 1TB you would need 500 of them. That would require a PCB the size of a ATX motherboard. with millions of times more complex trace patterns. The power delivery for such setup would require 1200w just for the GDDR5 DRAM dies when factoring in leakage.

There is also the issue of trace length. At X distance to GPU DIE, GDDR5 latency will be so high that the raw speed wont matter. Thus you will NEVER achieve a balanced memory setup with GDDR5.

 

The Firepro W9150 with 32GB of GDDR5 is already at the limit of what you can physically do with a normal PCI Express board. To be able to make a PCB able to handle more memory you would have to make a custom board of its own form factor with custom PCB, power delivery, custom memory controllers etc.

 

 

At 3000MB/s read/write for the highest end NVMe drives, and assume they use one or two PLX chips to create 2x PCIe Gen3 x4 connections, one would be able to move between 3-6GB/s of data. Depending on whether they add a Marvell or Intel RAID controller.

 

PLX chips cost around 95-140 USD a piece depending on "bulk"prices.

not sure about RAID controllers, looking at RAID AIBs i would assume 50-75 USD for a RAID controller.

 

It is not a cheap board to produce going by form factor and product features.

They wouldn't need two PLX chips. The chip they used on the Radeon Pro Duo is already a 48 lane chip, which would be plenty enough for something like this card. Such a PLX chip can easily be reworked to provide ample lanes for the GPU and two SSDs. Remember, its 16 in, 16 out, 16 mirror. M.2 SSD's that run off of PCIE x4 are 32Gbps. Plenty of extra lanes to handle that off of a single PEX8747 chip. 

 

Technically, if they used an on-die controller, they could bypass the 100ns PEX8747 switch latency (peak) as the card could address the slots directly. This would make it the single most complicated GPU in history, lol.