AMD: Southern Islands (7*** series) Speculation/ Rumour Thread

[Gipsel]

The article on ComputerWorld was removed. AMD released the FirePro S10000 (not SM10000) today and it's really a dual-GPU board, which was showed by Mark Papermaster this summer.

http://www.amd.com/PublishingImages...tShots/375WPNG/52625A_FirePro_S10000_375W.png

http://www.amd.com/us/products/workstation/graphics/firepro-remote-graphics/S10000/Pages/S10000.aspx

http://www.amd.com/us/press-releases/Pages/amd-most-powerful-server-graphics-2012nov12.aspx

Strange thing is, according to SemiAccurate it uses two Tahiti LE salvage chips (28 CUs) at 825 MHz, while one could get the same FLOP/s rating also with two full Tahitis at only 725 MHz, the latter version probably consuming quite a bit less power (especially when binned for a reduced voltage). I never thought the rasterizer and ROP performance would be that important in that market.

Spoiler

By the way, AMD's data sheet for the S10000 contains an error for the stated performance/Watt. While the claimed 3.94 GFlops/W for DP matches the 375W TDP and the 1.48 TFLOP/s, the 5.77 GFlops/W in SP should probably read 15.77 GFlops/W.

 

[OpenGL guy]

Strange thing is, according to SemiAccurate it uses two Tahiti LE salvage chips (28 CUs) at 825 MHz, while one could get the same FLOP/s rating also with two full Tahitis at only 725 MHz, the latter version probably consuming quite a bit less power (especially when binned for a reduced voltage). I never thought the rasterizer and ROP performance would be that important in that market.

Triangle rate is higher with the higher clock speed.

 

[Gipsel]

Triangle rate is higher with the higher clock speed.

Of course it is. But for pure numbercrunching it is quite irrelevant and two full Tahitis @ 725 MHz (with reduced voltage) would probably easily fit within a 300W TDP. At 300W it would have almost the same SP perf/Watt as the K10 (which is a duck at DP and has no ECC on internal SRAMs) and would also sit only slightly behind the DP perf/Watt of a K20 (while being in front in SP). It would just look more competitive with 725MHz and 300W instead of using cut down Tahitis running with 825MHz and 375W. This thing isn't named W10000, but S10000 (which should imply more focus on number crunching). But obviously AMD didn't do this distinction and just opted otherwise.

 

[A1xLLcqAgt0qc2RyMz0y]

It always bugged me why AMD did not do a 7990..and let Nvidia run scot free on the high end....these cards are very easy money i would think from the high mark up....

However i remember catching a glimpse of "a" reference 7990 from AMD's own hands....and it has with 3 fans....maybe that has something to do about the missing dual Tahiti...when the 6990 came with a single fan stuck in the center...and Nvidia rushed out the 690.

That card you saw is now called the S10000 and uses 375 Watts.

AMD did not produce a 7990 because 375 watts is not viable in a desktop machine.

Nvidia produced the 690 and stayed in the Power Envelope of a desktop. The 690 won rave reviews about it's design so I have no idea where your "rushed out" comes from.

AMD wasn't able to do a dual card and stay in the Power Envelope of a desktop so they didn't release it for the desktop but for the workstation market where they could get away with the very high power usage.

 

[Kaotik]

That card you saw is now called the S10000 and uses 375 Watts.

AMD did not produce a 7990 because 375 watts is not viable in a desktop machine.

Nvidia produced the 690 and stayed in the Power Envelope of a desktop. The 690 won rave reviews about it's design so I have no idea where your "rushed out" comes from.

AMD wasn't able to do a dual card and stay in the Power Envelope of a desktop so they didn't release it for the desktop but for the workstation market where they could get away with the very high power usage.

You completely missed HD6990 and GTX590 which both went to that 375W? (or was it even over?)

 

[iMacmatician]

You completely missed HD6990 and GTX590 which both went to that 375W? (or was it even over?)

IIRC the 6990 is 375 W and the 590 is 365 W.

I'm guessing NVIDIA wanted to go back to 300 W with the 690, for one reason or another.

 

[AlexV]

Of course it is. But for pure numbercrunching it is quite irrelevant and two full Tahitis @ 725 MHz (with reduced voltage) would probably easily fit within a 300W TDP. At 300W it would have almost the same SP perf/Watt as the K10 (which is a duck at DP and has no ECC on internal SRAMs) and would also sit only slightly behind the DP perf/Watt of a K20 (while being in front in SP). It would just look more competitive with 725MHz and 300W instead of using cut down Tahitis running with 825MHz and 375W. This thing isn't named W10000, but S10000 (which should imply more focus on number crunching). But obviously AMD didn't do this distinction and just opted otherwise.

These are not Firestream (lulz, that's dead) cards though, and the markets they are intended for tend to care quite a bit about triangle processing rates (and other traditional CAD / modelling things) than they do about number-crunching. Even thoug AMD's messed up product stack makes this seem as a competitor to Tesla, it is actually a competitor to Quadro.

Also what AMD needs (actually would have needed before the train left a long time ago) to compete with Tesla is not more FLOPs, but rather a less dubious software proposition.

 

[Dave Baumann]

These are not Firestream (lulz, that's dead) cards though, and the markets they are intended for tend to care quite a bit about triangle processing rates (and other traditional CAD / modelling things) than they do about number-crunching. Even thoug AMD's messed up product stack makes this seem as a competitor to Tesla, it is actually a competitor to Quadro.

"S" pretty much stands for server, and these are targetted towards number crunching workloads not CAD workloads; Visualisation and sim are more maths problems. So these are more inline with prior "Firestream" offering than Quadro competitors.

Also what AMD needs (actually would have needed before the train left a long time ago) to compete with Tesla is not more FLOPs, but rather a less dubious software proposition.

Why? Software stack for professional solutions is very different from consumer.

 

[gongo]

That card you saw is now called the S10000 and uses 375 Watts.

AMD did not produce a 7990 because 375 watts is not viable in a desktop machine.

Nvidia produced the 690 and stayed in the Power Envelope of a desktop. The 690 won rave reviews about it's design so I have no idea where your "rushed out" comes from.

AMD wasn't able to do a dual card and stay in the Power Envelope of a desktop so they didn't release it for the desktop but for the workstation market where they could get away with the very high power usage.

Yap! That's the mythical card i saw...3 frick'in fans when their older 6990 had only one centrifugal spinner! This "7990" consists of 2 Tahiti Pro instead of the XTs....could it even fought on par with GTX 690? No wonder they canned it officially..

What caused AMD to missed their perf/mm/watt target they held on since 3870...? Tahiti isnt the largest die....leakage?

 

[Gipsel]

"S" pretty much stands for server, and these are targetted towards number crunching workloads not CAD workloads; Visualisation and sim are more maths problems. So these are more inline with prior "Firestream" offering than Quadro competitors.

So why not using fully enabled Tahitis at a lower clock/voltage/power to maximize the FLOPs/W?
At 300W (appears reasonable to me for 725MHz and a slightly reduced voltage, nV squeezed two GK104@745MHz in 225W on the K10) and 32CUs it would aim squarely at both the K10 and K20 Teslas. I would imagine the ~14% lower rasterizer and ROP performance isn't that crucial for the number crunching tasks you mentioned. And it would look much better on a spec sheet comparing it to K10 and K20.

 

[Dave Baumann]

Yap! That's the mythical card i saw...3 frick'in fans when their older 6990 had only one centrifugal spinner! This "7990" consists of 2 Tahiti Pro instead of the XTs....could it even fought on par with GTX 690? No wonder they canned it officially..

A single blower cuts through the HSF, limiting the thermal capacity; Axial fans have a lower CFM (airflow) than blower do, but can gain more thermal density from the heatsink by not chopping it into two separate blocks. These will result in a quieter solution, hence why you see the partner solutions release so far having a similar configuration.

So why not using fully enabled Tahitis at a lower clock/voltage/power to maximize the FLOPs/W?

Sorry, not my BU, so I don't know the decisions that drove this configuration.

 

[AlexV]

Why? Software stack for professional solutions is very different from consumer.

I'm curious how do you define different so as to make a dysfunctional stack and spotty support an advantage...Also, I'm curious how things in pro are peachy when NV is making piles upon piles of money, with great margins, whilst ATI is a noise factor in that particularly profitable space. If everything is always working right, how come it never shows up anywhere (but it's always coming, honest!)?

 

[Alexko]

Sorry, not my BU, so I don't know the decisions that drove this configuration.

Might it have something to do with needing the memory controller to be fast to ensure good memory performance?

 

[Dave Baumann]

I'm curious how do you define different so as to make a dysfunctional stack and spotty support an advantage...Also, I'm curious how things in pro are peachy when NV is making piles upon piles of money, with great margins, whilst ATI is a noise factor in that particularly profitable space. If everything is always working right, how come it never shows up anywhere (but it's always coming, honest!)?

Personally I don't know what you're speaking of when you say "dysfunctional and spotty"; software for WS goes through a different cycle and a different support structure as well as going through all the application certs that are key for the market.

And just because our financial reporting does not place as much emphasis on this one area (and I have to operate within those bounds) does not mean that marketshare has not been improving. Dynamics of the traditional WS market are very different to others a change in marketshare is always going to be a relatively slow thing because it is not just about hitting OEM schedules and requirements but more about appealing to the customers of the OEM's. Pointing at one thing is rarely telling the whole story.

 

[CarstenS]

So why not using fully enabled Tahitis at a lower clock/voltage/power to maximize the FLOPs/W?

Probably profit maximization? The one area where the FirePro shines the most - raw FLOPS that is - is already won by a healthy margin. And other reasons such as the ones hinted at by AlexV, wouldn't be disregarded by potential customers if only there was a 300 watt power envelope to the card.

AMD would need good Tahiti XT2s for FirePro W9000 as well as 7970 GHz and non-GHz Editions.

 

[Gipsel]

Probably profit maximization? The one area where the FirePro shines the most - raw FLOPS that is - is already won by a healthy margin. And other reasons such as the ones hinted at by AlexV, wouldn't be disregarded by potential customers if only there was a 300 watt power envelope to the card.

AMD would need good Tahiti XT2s for FirePro W9000 as well as 7970 GHz and non-GHz Editions.

And profit maximization would dictate, that you should put some good low power dies (you probably would need a relatively low amount compared to Radeon retail volumes) in a FirePro, not a HD7970 if there would be really a shortage of those (which I don't think is the situation, considering Tahiti production runs for a year already and there appears to be plenty of supply with recent price reductions and game bundle sales promotion ongoing).

 

[Dave Baumann]

Professional markets have different longevity expectations though.

 

[CarstenS]

And profit maximization would dictate, that you should put some good low power dies (you probably would need a relatively low amount compared to Radeon retail volumes) in a FirePro, not a HD7970 if there would be really a shortage of those (which I don't think is the situation, considering Tahiti production runs for a year already and there appears to be plenty of supply with recent price reductions and game bundle sales promotion ongoing).

I disagree because [Price of FirePro + Price of HD 7970] > [Price of FirePro + Price of HD 7950]

 

[Gipsel]

I disagree because [Price of FirePro + Price of HD 7970] > [Price of FirePro + Price of HD 7950]

But we have no indication of a shortage of appropriate Tahiti dies (and a bin capable of low power, [in this context <120W for the chip alone] operation @ 725MHz with let's say 1.0V [edit: or even slightly lower as the HD7950 needs 1.0V for 800 MHz] is not exactly the same bin as for 1000/1050 MHz operation at 1.17/1.22V within a 200+W power budget). But even going with your assertion that a FirePro would use the same (or "better") bin than a HD7970 or HD7950 respectively, you forget about the in my opinion most probable alternative: You sell less FirePro cards if they are not attractive. You simply loose sales. That turns your point into:
What is better? Selling a HD7970 and having some FirePros piling up in your inventory or actually selling a FirePro and a HD7950?

Might it have something to do with needing the memory controller to be fast to ensure good memory performance?

There is a P-state for the HD7970 in which the chip/memory runs at 500/1375 MHz (i.e. the memory controller runs at full speed) with a voltage of ~0.95V (may differ slightly for different dies). So this can't be the reason.

 

[Alexko]

But we have no indication of a shortage of appropriate Tahiti dies (and a bin capable of low power, [in this context <120W for the chip alone] operation @ 725MHz with let's say 1.0V [edit: or even slightly lower as the HD7950 needs 1.0V for 800 MHz] is not exactly the same bin as for 1000/1050 MHz operation at 1.17/1.22V within a 200+W power budget). But even going with your assertion that a FirePro would use the same (or "better") bin than a HD7970 or HD7950 respectively, you forget about the in my opinion most probable alternative: You sell less FirePro cards if they are not attractive. You simply loose sales. That turns your point into:
What is better? Selling a HD7970 and having some FirePros piling up in your inventory or actually selling a FirePro and a HD7950?

Or you can sell your "better" FirePros for a higher price, which boils down to the same result.

There is a P-state for the HD7970 in which the chip/memory runs at 500/1375 MHz (i.e. the memory controller runs at full speed) with a voltage of ~0.95V (may differ slightly for different dies). So this can't be the reason.

I see, thanks. I have no clue why AMD did this, then.