Nvidia Maxwell Quadro Upset

[patrickjp93]

http://wccftech.com/dual-nvidia-quadro-m6000-graphics-cards-pictured-gm200-professionals-heading-launch-gtc-titan/

 

Apparently Nvidia is in fact using Maxwell to make Quadros and Teslas using GM200. For those of you who wonder why this is news, GM204 was X-Rayed and analysed by independent experts and shown to (supposedly)have no 64-bit SPs, no native double precision, and thus be limited to 1/32 performance vs. 1/2 for native support vs. SP compute performance. GM200 adds 50% more SPs over GM204 and has an additional 2 memory controllers over their GM204 chip in the GTX 980. My one question is this: how the F*CK do you double the SP bit-width, increase the number by 50%, get more memory controllers, and only increase the die size by 52% when on the same manufacturing process (28nm)? Did Nvidia come up with its own HDL, did the independent analysts fall down on the job, or is Nvidia going to try this with 1/32 performance? I'm just...WHAT???!!!

 

The M6000 Quadro SP compute performance is rated at 6.07TFlops (Knight's Landing Xeon Phi is rated for 6TFlops SP as well, due to launch in Q1 2016 with 8GB of HBM) with 12GB of 6600MHz GDDR5 on a 384-bit bus toting 3072 SPs clocked at 988MHz with just a 200W TDP (K6000 has 225W TDP). The unveiling is supposedly GTC 2015.

 

I'm just...DAMNIT WHO LIED/HELD BACK?!

[TheSLSAMG]

I knew something would come of the pic of Deadmau5's system.

[Enderman]

Do you realize that die size doesnt mean anything?

They are increasing efficiency, making things smaller, and cramming more transistors into a smaller space

 

This isnt AMD...

[Firearm2112]

I swear I have seen this pic on Deadmau5's instagram

[patrickjp93]

Do you realize that die size doesnt mean anything?

They are increasing efficiency, making things smaller, and cramming more transistors into a smaller space

 

This isnt AMD...

 

1) lie, because it's still the 28nm process

2) Hence my question about either lying/incorrect "experts" or Nvidia creating an HDL concept

3) Nvidia's Titan X is still set to lose to the 390X according to current benchmarks, so it's not like AMD sucks at making GPUs.

[theloveablemoose]

1) lie, because it's still the 28nm process

2) Hence my question about either lying/incorrect "experts" or Nvidia creating an HDL concept

3) Nvidia's Titan X is still set to lose to the 390X according to current benchmarks, so it's not like AMD sucks at making GPUs.

Well the supposed 390x benchmarks are still rumored no?

[ObeseWalrus]

I swear I have seen this pic on Deadmau5's instagram

you have

[LeChef]

Y U NO UNDERSTAND "RUMOURS"

[patrickjp93]

Y U NO UNDERSTAND "RUMOURS"

Hassan is one of the only reporters on WCCF I trust. He has a sterling record despite that of the site.

[C0LL0SS0S]

1) lie, because it's still the 28nm process

2) Hence my question about either lying/incorrect "experts" or Nvidia creating an HDL concept

3) Nvidia's Titan X is still set to lose to the 390X according to current benchmarks, so it's not like AMD sucks at making GPUs.

AAAAActually AMD sucks, because NVIDIA leads with low TDP GPUs and still has the best GPU most of the time, if NVIDIA would stop its progress and make GPUs that uses as much power as AMD, NVIDIA would rule again so stop comparing GPUs if they are not in the same league

 

[patrickjp93]

AAAAActually AMD sucks, because NVIDIA leads with low TDP GPUs and still has the best GPU most of the time, if NVIDIA would stop its progress and make GPUs that uses as much power as AMD, NVIDIA would rule again so stop comparing GPUs if they are not in the same league

 

Independent testing shows the GTX 980 doesn't use much less power than the R9 290X, and every chip maker has a different way to calculate TDP. Intel's equation is different from AMD's which is different from AMD's. THERE IS NO INDUSTRY STANDARD!

[Bsmith]

dont take wccf and other rumors to serious, combining these will release BS for sure

[patrickjp93]

dont take wccf and other rumors to serious, combining these will release BS for sure

 

WCCF's had a pretty strong record since September. Nothing they've posted been's wrong by any significant amount.

[BonSie]

I'm a noob in such things but, doesn't it have 33% (2000/3000) more SPs and 4 extra memory controllers?

[sgloux3470]

AAAAActually AMD sucks, because NVIDIA leads with low TDP GPUs and still has the best GPU most of the time, if NVIDIA would stop its progress and make GPUs that uses as much power as AMD, NVIDIA would rule again so stop comparing GPUs if they are not in the same league

 

 

The 780Ti and 290X were neck and neck in power consumption, with an overclocked 780Ti using far more power than an overclocked 290X.

 

Maxwell is the first time Nvidia can be said to be more efficient than AMD, and AMD hasn't even released new cards to actually directly compete yet.  And it's not like Nvidia didn't run hot as hell in the past, although I realize most nvidia fanboys may be a bit young to remember that.

 

Stop drinking the fanboy kool-aid.  Both companies put out great GPU's.  Competition is the reason Maxwell is cheaper than the 7xx series, by the by.

 

[Enderman]

1) lie, because it's still the 28nm process

2) Hence my question about either lying/incorrect "experts" or Nvidia creating an HDL concept

3) Nvidia's Titan X is still set to lose to the 390X according to current benchmarks, so it's not like AMD sucks at making GPUs.

1) you dont need a smaller process to make a smaller die

 

2) those benchmarks are rumors from sites that are known for posting fake info in order to get audience

there is absolutely nothing backing up those benchmarks

 

3) i never said AMD sucked at making GPUs, all I said that nvidia doesnt just stuff more transistors into a die, they actually work on optimising the power efficiency and space

have you seen how big AMD GPU dies are compared to nvidia? lol

[patrickjp93]

I'm a noob in such things but, doesn't it have 33% (2000/3000) more SPs and 4 extra memory controllers?

 

Memory controllers are always 64-bit so as to attach 2 VRAM chips to each.

 

Also, 3072/2048 = 1.5, or 50% more. You could say the 980 has 33% fewer SPs. Ratios are relative.

 

[NumLock21]

That red rocket x is a amd based gpu.

[patrickjp93]

1) you dont need a smaller process to make a smaller die

 

2) those benchmarks are rumors from sites that are known for posting fake info in order to get audience

there is absolutely nothing backing up those benchmarks

 

3) i never said AMD sucked at making GPUs, all I said that nvidia doesnt just stuff more transistors into a die, they actually work on optimising the power efficiency and space

have you seen how big AMD GPU dies are compared to nvidia? lol

1) To use the same number of transistors or more? Yes you do. HDL actually shrinks the number of transistors used per component by eliminating redundancy (a safety feature actually in case a signal doesn't make it through, extremely necessary these days with so many quantum effects to deal with) and reduce clock rates (lowering the electrical noise on the wires to improve signaling) while adding more components (like the security chip in Carrizo). Every atomic part of a CPU core uses a defined number of transistors, but the manner in which they are used/organized will determine heat density. In order to pack them closer together and avoid heat problems you must either drop electrical power (reducing heat generated per transistor) or rotate the components to line up the hot and cold regions of each to decrease thermal density or do both.

Thus, like I said, either Nvidia came up with an HDL scheme which only applies to GM 200 (really unlikely that they did that, because it would increase chip yields if you could shrink GM 204 by 30% with it) or some "experts" have some explaining to do as to how they missed 64-bit SPs on the GM 204 chip.

 

2) Chiphell? Hardly. Chiphell gets extremely accurate when you get close to launch. It's only when you're multiple months away that they're generally way off. We're within a month of the Titan X. That's fairly accurate I guarantee you.

 

3) If you think AMD doesn't you have no clue how electrical design processes work. There are less than 100 people in the world that actually have a full understanding of these large chips. If you think they don't work on optimizing for power, space, and heat, you're blind. AMD sucks at making coolers though, no doubt.

 

AMD's dies are smaller than Nvidia's, by far.

[harrynowl]

have you seen how big AMD GPU dies are compared to nvidia? lol

Seems that you haven't.

2007/6:

8800GTX G80: 484mm^2

HD3870 X2 R680: 192mm^2 x2

2008:

GTX280 GT200: 576mm^2

HD4890 RV790: 282mm^2

2010:

GTX480 GF100: 529mm^2

HD5870 Cypress XT: 334mm^2

2011:

GTX580 GF110: 520mm^2

HD6970 Cayman XT: 389mm^2

2012:

GTX680 GK104: 294mm^2

HD7970 Tahiti XT: 365mm^2

2013:

GTX780 GK110: 550mm^2

R9 290X Hawaii XT: 428mm^2

2014:

GTX980 GM204: 398mm^2

R9 290X Hawaii XT: 428mm^2

Generally when AMD are bigger it isn't that much bigger, when nvidia are bigger it's quite a bit bigger. Where are you getting your statistics from?

[Deletive]

1) you dont need a smaller process to make a smaller die

 

2) those benchmarks are rumors from sites that are known for posting fake info in order to get audience

there is absolutely nothing backing up those benchmarks

 

3) i never said AMD sucked at making GPUs, all I said that nvidia doesnt just stuff more transistors into a die, they actually work on optimising the power efficiency and space

have you seen how big AMD GPU dies are compared to nvidia? lol

 

Hawaii is ~438mm Gk110 was 561mm. Gm200 is probably 600mm Which is huge Hawaii will probably be around that size if not smaller.

AAAAActually AMD sucks, because NVIDIA leads with low TDP GPUs and still has the best GPU most of the time, if NVIDIA would stop its progress and make GPUs that uses as much power as AMD, NVIDIA would rule again so stop comparing GPUs if they are not in the same league

Nvidia leads now. AMD still hasn't released anything to compete with Maxwell. GCN 1.2 could just as easily compete from what we're seeing from carrizo

[BonSie]

I'm guessing it won't have FP64 units, meaning AMD will have a fieldday in the double precision market with its new GPUs.

I think they even said that themself.

[patrickjp93]

I'm guessing it won't have FP64 units, meaning AMD will have a fieldday in the double precision market with its new GPUs.

I think they even said that themself.

 

This is still very confusing. Quadros kinda need DP for the rendering workloads, and Teslas need it for scientific computing. To release Quadros but not Teslas is bordering on the insane (and the Kepler 2.0 revision isn't showing to be more than 7% better than the first generation per clock). That leaves Nvidia wide open for the Knight's Landing Xeon Phi to eat up much of the HPC marketshare, at least if AMD dawdles on releasing Fiji-based FirePro cards. A $1200 new Xeon Phi with 6/3 TFlops rating at a 93+% RWP/Theoretical ratio is much more enticing than an equally rated but ~60% effective ratio Tesla, especially since the Xeon Phi is 250-300W TDP and the only Tesla which could match for RWP right now is a 400W monster dual-card that costs almost $7000.

 

I reiterate: insanity given Pascal isn't until Q2/Q3 2016 and Nvidia never leads with professional or HPC cards and Nvidia has nothing else to play until that time.

[The Great Sayain Man]

Who is that Deadmau5 to have all the juciy stuff before others do? 

[cesrai]

I learned a lot from this thread, thanks for the helpful discussion everybody.