NVIDIA Ampere GA102-300-A1 GPU Die Pictured

[btarunr]

Here's the first picture of an NVIDIA "Ampere" GA102 GPU die. This is the largest client-segment implementation of the "Ampere" architecture by NVIDIA, targeting the gaming (GeForce) and professional-visualization (Quadro) market segments. The "Ampere" architecture itself debuted earlier this year with the A100 Tensor Core scalar processor that's winning hearts and minds in the HPC community faster than ice cream on a dog day afternoon. There's no indication of die-size, but considering how tiny the 10.3 billion-transistor AMD "Navi 10" die is, the GA102 could come with a massive transistor count if its die is as big as that of the TU102. The GPU in the picture is also a qualification sample, and was probably pictured off a prototype graphics card. Powering the GeForce RTX 3090, the GA102-300 is expected to feature a CUDA core count of 5,248. According to VideoCardz, there's a higher trim of this silicon, the GA102-400, which could make it to NVIDIA's next halo product under the TITAN brand.



View at TechPowerUp Main Site

 

[DeathtoGnomes]

Weeeeee, Pictures!

 

[xkm1948]

Now somebody needs to fire up Adobe Illustrator and do some pixel measurement based on the surface mount components around the die versus the TU102 core. TU104 is 754 mm2. We should be able to have an estimate of die size

Use the top right TU102 logo. A102 has something similar on the bottom.

 

[Chris34]

Ampere process is 10nm or 7nm? Because if it's 7nm that's one huge piece of silicon.

 

[TheLostSwede]

Ampere process is 10nm or 7nm? Because if it's 7nm that's one huge piece of silicon.

8nm from Samsung according to some rumours.

 

[Sykobee]

People believe it's Samsung 8nm, which is a 10nm variant.

 

[Chris34]

Thanks I lost track to what process they were using. I guess we'll know for sure when the NDA is over.

 

[xkm1948]

TW == Taiwan I believe

 

[M2B]

TW == Taiwan I believe

Interesting...

 

[Rob94hawk]

!!

 

[ppn]

Well Samsung 8nm UHD being 61Mtr,mm2, the HP variant must be around 50Mtr,mm2, or double transistor count for a 630mm2 die, ~~31500 Mtr.
In case of later 7nm TSMC or refresh it could get down to 471.

 

[RedelZaVedno]

View attachment 166647!!

I can just see Leather Jacket Lover on the stage screaming 300% ray tracing performance increase for only 40% more $$$ while concealing abysmal +35% standard rasterization performance increase

 

[Deleted member 185088]

I can just see Leather Jacket Lover on the stage screaming 300% ray tracing performance increase for only 40% more $$$ while concealing abysmal +35% standard rasterization performance increase
View attachment 166649

They already messed up with the pathetic Turing that for more money couldn't beat the 1080ti, and AMD already beat their equivalent offering (albeit with worse drivers), I believe they will offer a performance bump, and hopefully AMD too so we'll get affordable GPUs.

 

[RedelZaVedno]

They already messed up with the pathetic Turing that for more money couldn't beat the 1080ti, and AMD already beat their equivalent offering (albeit with worse drivers), I believe they will offer a performance bump, and hopefully AMD too so we'll get affordable GPUs.

I really hope so, but I remain very skeptical until I see the benchmarks. Samsung's 8nm (10 nm in real life) node doesn't give me confidence they'll be able to hit high boost frequency clocks. We know RDNA2 on 7nm EUV process can hit 2.23Ghz (SP5), but leaked numbers from Ampere all end at 1700-1900Mhz. That would likely mean 17-20 TFlops which is not all that impressive given the fact that 2080TI could reach around 13.5Tflops. I'm really afraid we're gonna get standard generational +30-40% rasterization performance increase for A LOT more money.

 

[Ja.KooLit]

Pictures means nothing. Benchmark means something.........

 

[medi01]

No estimates of the size? Disappointing thread.

 

[Legacy-ZA]

No estimates of the size? Disappointing thread.

Yes, clickbait at it's finest.

 

[Dante Uchiha]

Well Samsung 8nm UHD being 61Mtr,mm2, the HP variant must be around 50Mtr,mm2, or double transistor count for a 630mm2 die, ~~31500 Mtr.
In case of later 7nm TSMC or refresh it could get down to 471.

It doesn't work that way...
The density varies according to the chip, TSMC does marketing with the best possible scenario.

 

[Vya Domus]

No estimates of the size? Disappointing thread.

Big, the size is big, that's all we need to know really. If it's 630 mm^2 or 650 mm^2 or whatever doesn't mean much, what's important is that it's yields are inevitably bad and it's going to be ridiculously expensive.

 

[Tomgang]

Now I just wunder one thing. Rumors says big ass chip, 350 watt tdp stock, pictures of 3 slot stock cooler = is this gtx 480 ano 2020

Cause it all put together, sure as he'll sujest we have one big hot power hungry chip and if prices are just as high as the tdp is high, the chip is big and hot. We'll prepare to sell a kidney, an arm and a leg for any chance of owning one of these modern days GTX 480.

I really hope I am wrong about the above. But rumors and pictures shown so far. Just screams GTX 480.

 

[Tomorrow]

Now I just wunder one thing. Rumors says big ass chip, 350 watt tdp stock, pictures of 3 slot stock cooler = is this gtx 480 ano 2020

Cause it all put together, sure as he'll sujest we have one big hot power hungry chip and if prices are just as high as the tdp is high, the chip is big and hot. We'll prepare to sell a kidney, an arm and a leg for any chance of owning one of these modern days GTX 480.

I really hope I am wrong about the above. But rumors and pictures shown so far. Just screams GTX 480.

Better hope Samsungs process does not have high defect rate. 480's died left and right during their launch. Even Turing was not immune to failiures on the biggest die.

 

[Tomgang]

Better hope Samsungs process does not have high defect rate. 480's died left and right during their launch. Even Turing was not immune to failiures on the biggest die.

I dont remember gtx 480 as to be a gpu dying left and right. But I sure as he'll remember gtx 480 to be a very hot running gpu. Well hot running whas a general problem for Fermi.

 

[birdie]

I can just see Leather Jacket Lover on the stage screaming 300% ray tracing performance increase for only 40% more $$$ while concealing abysmal +35% standard rasterization performance increase
Crap-picture

DO NOT BUY ANYTHING FROM NVIDIA. NO ONE FORCES YOU. Vote with your wallet if you think NVIDIA is overpriced hot garbage with minimum performance improvements.

See, it's simple. Also it will save the rest of us from tens of thousands of (to put it mildly) inane comments on dozens of bulletin boards and comments sections of various tech websites.

Better hope Samsungs process does not have high defect rate. 480's died left and right during their launch. Even Turing was not immune to failiures on the biggest die.

I would like to see hard numbers and stats cause otherwise I will have to call you a liar. I perfectly remember the launch of the GTX 480 and I don't recall anything like what you're saying.

Meanwhile AMD continues to sell half-assed RX5000 cards which have the highest fault rate (was in the news recently) and a huge number of complaints in terms of stability and quality. I've recently sold my RX 5600 XT and bought the GTX 1660 Ti (yeah, on average 8-20% slower depending on the resolution but I don't care) because I waited for five months for drivers which allow me to actually use the GPU instead of seeing black screens, BSODs and game crashes all the time.

 

[Hotobu]

(Tangential thought)

My guess is after Nvidia releases all of the skus in the original product stack they'll release TI/Super variants on TSMC 7nm.

 

[Tomorrow]

I would like to see hard numbers and stats cause otherwise I will have to call you a liar. I perfectly remember the launch of the GTX 480 and I don't recall anything like what you're saying.

There are none. There never are because manufacturers do not release overall sales numbers and users dont have tens or hundreds of samples to test.
The same way there are no stats about how many 2080Ti's died at launch.

Truth be told i go by what others have said:

(10:15 mark).
TSMC's 40nm process was pretty bad back then.