AMD releases the S9300 dual Fiji Firepro

[Starelementpoke]

Just now, don_svetlio said:

If you buy me a Fury X you can have my anal virginity. Not even joking

If you give me the money for it, no problem.

[Citadelen]

1 minute ago, don_svetlio said:

If you buy me a Fury X you can have my anal virginity. Not even joking

Then what do we get for a Radeon Pro Duo?  

[don_svetlio]

2 minutes ago, Citadelen said:

Then what do we get for a Radeon Pro Duo?  

Do you need a Harem? Can arrage that for 1 night.

 

3 minutes ago, Starelementpoke said:

If you give me the money for it, no problem.

Actually, your 7950, is it a rear exhaust? If so I would be up for a swap since my airflow is terrible and it would help me.

[Citadelen]

3 minutes ago, don_svetlio said:

Do you need a Harem? Can arrage that for 1 night.

 

[don_svetlio]

Just now, Citadelen said:

 

I'm not joking man, though some of the participants would be male.

[Citadelen]

5 minutes ago, don_svetlio said:

I'm not joking man, though some of the participants would be male.

That's fine with me.  

[Prysin]

25 minutes ago, don_svetlio said:

Actually - WHY NOT PUT THIS IN LAPTOPS?

You will probably find it in some MacPro product.

[patrickjp93]

It's not running passively and not at full tilt. Servers have 10,000RPM+ fans blowing from one end to the other. It's the equivalent of having a high-speed blower on the card itself.

 

This really doesn't interest me because of the 4&4GB memory limits. If you're solving ODEs on a GPU, then sure, it's powerful (if their OpenCL drivers ever improve...), but even then, something as compute-heavy as ODEs, if in a big, interconnected system, need to have the data for the other moving parts available. Even an 8GB combined buffer is too meager for that. You'd be streaming in pieces of data from other processors more often than the GPU would be doing computations. At best, this would be for small-scale (think a few tens of thousands of interconnected pendulums or particles) simulations on a workstation. You'd be foolish to use this in a big cluster setting.

[Prysin]

1 minute ago, patrickjp93 said:

It's not running passively and not at full tilt. Servers have 10,000RPM+ fans blowing from one end to the other. It's the equivalent of having a high-speed blower on the card itself.

 

This really doesn't interest me because of the 4&4GB memory limits. If you're solving ODEs on a GPU, then sure, it's powerful (if their OpenCL drivers ever improve...), but even then, something as compute-heavy as ODEs, if in a big, interconnected system, need to have the data for the other moving parts available. Even an 8GB combined buffer is too meager for that. You'd be streaming in pieces of data from other processors more often than the GPU would be doing computations. At best, this would be for small-scale (think a few tens of thousands of interconnected pendulums or particles) simulations on a workstation. You'd be foolish to use this in a big cluster setting.

Quote

Popular GPU neutral network implementations don’t use FP32 math, rather they use even lower precision FP16 math. And though the S9300 X2’s FP16 throughput is merely equal to its FP32 throughput, internally Fiji supports natively storing FP16 data types, which will significantly reduce register pressure on the card, and register pressure is almost always a concern for HPC kernel development.

would you assume it stores in a L3 cache? or the L2?

Because if Fiji has a larger L2 or L3, it would lessen the impact of memory usage for certain types of data.

[DevilishBooster]

1 hour ago, Prysin said:

just put a conveyor belt at @zMeuls feet when he reads this post

I'm just sitting here with my popcorn waiting for him to chime in....

[patrickjp93]

3 minutes ago, Prysin said:

would you assume it stores in a L3 cache? or the L2?

Because if Fiji has a larger L2 or L3, it would lessen the impact of memory usage for certain types of data.

But that still limits the size and scalability of the distributed neural network anyway. Like I said, it's really a compute workstation card or a small-scale supercomputer card. That said, kudos to AMD for being bold, but the product's only half-baked from my perspective.

[Prysin]

49 minutes ago, patrickjp93 said:

But that still limits the size and scalability of the distributed neural network anyway. Like I said, it's really a compute workstation card or a small-scale supercomputer card. That said, kudos to AMD for being bold, but the product's only half-baked from my perspective.

true, however this would still scale well enough to use in 40ft shipping container sized "pods" used for special geoscience stuff within the offshore workspace.

[DXMember]

is $5999 really the pricepoint? if not then you should more clearly separate your personal input for the news itself (:

 

but this is different than Radeon Pro Duo, Raja demoed at CAPSAICIN , right? Because that thing was running on water and was clearly marked as Radeon card

[Prysin]

1 minute ago, DXMember said:

is $5999 really the pricepoint? if not then you should more clearly separate your personal input for the news itself (:

 

but this is different than Radeon Pro Duo, Raja demoed at CAPSAICIN , right? Because that thing was running on water and was clearly marked as Radeon card

the MSRP is $5999 according to Anandtech

[DXMember]

1 minute ago, Prysin said:

the MSRP is $5999 according to Anandtech

oh, okay, because the way it was put in there was confusing whether troll or not

but it's still a different release from Radeon Pro Duo, right? 

[Prysin]

1 minute ago, DXMember said:

oh, okay, because the way it was put in there was confusing whether troll or not

but it's still a different release from Radeon Pro Duo, right? 

for the 3rd time in not even 2 full pages.

 

THIS IS A WORKSTATION CARD. NOT A "PRO-SUMER" OR "ENTHUSIAST" CARD....

THIS IS NOT THE RADEON PRO DUO

Do people even read the first post anymore?

 

Here, Visual comparison:

AMD RADEON PRO DUO

 

AMD RADEON FIREPRO S9300 X2

 

[CtW]

2 hours ago, don_svetlio said:

Actually - WHY NOT PUT THIS IN LAPTOPS?

yeah no.

[don_svetlio]

1 minute ago, CtW said:

yeah no.

Why not? The 175W 980 was used in laptops
This is 150W and performs about 30% better

[DXMember]

16 minutes ago, Prysin said:

for the 3rd time in not even 2 full pages.

 

THIS IS A WORKSTATION CARD. NOT A "PRO-SUMER" OR "ENTHUSIAST" CARD....

THIS IS NOT THE RADEON PRO DUO

Do people even read the first post anymore?

 

Here, Visual comparison:

AMD RADEON PRO DUO

 

AMD RADEON FIREPRO S9300 X2

 

Yeah, i know that.

but this is different from Radeon Pro Duo, right? Because at CAPSAICIN Raja Koduri showed off another dual Fiji card saying that it is already being used for VR movie experience and it was water cooled like in the first picture you linked, but the one on in the news is named differently, so are they the same cards?

[DevilishBooster]

18 minutes ago, Prysin said:

for the 3rd time in not even 2 full pages.

 

THIS IS A WORKSTATION CARD. NOT A "PRO-SUMER" OR "ENTHUSIAST" CARD....

THIS IS NOT THE RADEON PRO DUO

Do people even read the first post anymore?

You really expect people to read the thread? You have high standards, Fam....

/s

[CtW]

3 minutes ago, don_svetlio said:

Why not? The 175W 980 was used in laptops
This is 150W and performs about 30% better

The laptop 980s are underclocked and binned for efficiency which brings the TDP closer to 100w, which is still a lot for a laptop as far as heat output is concerned. Given that this is a server card designed EXCLUSIVELY for compute performance, and not for general use, you'd need to use a separate onboard GPU for video output to the laptop's monitor. The only reason someone would put this on a laptop would be if they wanted to do compute on the go, which is far to specific of a use case to justify putting in the R&D that would go into making this fit in a laptop. Would it be cool? Yes. Is it a good idea? No.

[don_svetlio]

1 minute ago, CtW said:

The laptop 980s are underclocked and binned for efficiency which brings the TDP closer to 100w, which is still a lot for a laptop as far as heat output is concerned. Given that this is a server card designed EXCLUSIVELY for compute performance, and not for general use, you'd need to use a separate onboard GPU for video output to the laptop's monitor. The only reason someone would put this on a laptop would be if they wanted to do compute on the go, which is far to specific of a use case to justify putting in the R&D that would go into making this fit in a laptop. Would it be cool? Yes. Is it a good idea? No.

The laptop 980s run at 1100Mhz aka stock aka 175W with 2 fans and 6 heatpipes and 2 copper heatsinks

This is a Fury X core running at 1GHz and being 150W Trust me. It WILL work better

[Prysin]

1 minute ago, DXMember said:

Yeah, i know that.

but this is different from Radeon Pro Duo, right? Because at CAPSAICIN Raja Koduri showed off another dual Fiji card saying that it is already being used for VR movie experience and it was water cooled like in the first picture you linked, but the one on in the news is named differently, so are they the same cards?

Different cards for different purposes.

 

Its like.

Why is a R9 390X sold at $429 and branded for gamers while the Firepro W9100 is sold for $1500 and branded for workstations?

Both use Hawaii GPUs...

Sure the Firepro W9100 has 16GB of ECC GDDR5, however ECC GDDR5 4GB-8GB costs like 30-40 bucks per chip at most. So adding it up it should only cost like 750-900 USD or so i guess.

 

The reason workstation cards cost more is:

Every driver is optimized to use every drop of power from the GPU in each application

Every driver is validated to be bug free and error free in every application

Every driver is validated for 24/7 operation at 100% load without causing issues

Every single piece of the GPU, PCB and PCB Components are rated for and guaranteed to operate at 100% load 24/7

You have 24/7 technical and or on-site support availible to you depending on service agreement

The GPUs used are the best, the highest binned most cherry picked GPUs there are.

Workstation card has full FP64 (Double Precision) enabled. Desktop cards have theirs halved or worse.

 

AMD Firepro is a "Workstation/Enterprise" brand of AMD. Similar to how Quadro and Tesla is "Workstation/Enterprise" brands of Nvidia.

These cards are meant strictly for productivity.

 

 

The Radeon Pro Duo is marketed towards game devs who want to make VR games, and PLAY VR games. Workstation cards are clocked lower then gaming cards, their drivers are missing pretty much every single game optimization there is. Workstation cards CAN play games, just way way way way worse then the "gaming" card can.

The Radeon Pro Duo has SOME workstation optimizations, however it is aimed towards GAMING with productivity sitting in the passenger seat to the side.

[Prysin]

4 minutes ago, don_svetlio said:

The laptop 980s run at 1100Mhz aka stock aka 175W with 2 fans and 6 heatpipes and 2 copper heatsinks

This is a Fury X core running at 1GHz and being 150W Trust me. It WILL work better

This is a Fury X running at 850MHz...

[don_svetlio]

1 minute ago, Prysin said:

This is a Fury X running at 850MHz...

Isn't it at 1GHz?