Blackwell GPU die sizes revealed; regression in some cases

[Daven]

Videocardz has compiled some great data about the die sizes and transistor density of Blackwell GPUs using HardwareLuxx as the source. It is becoming clear that Blackwell is only about AI/DLSS/RT. So the question is, with the exception of the 5090, will Blackwell have the same or less pure rasterization performance as Ada Lovelace?

NVIDIA reveals die sizes for GB200 Blackwell GPUs: GB202 is 750mm², features 92.2B transistors - VideoCardz.com

 

[remekra]

Yet some still believe that 5080 will be faster than 4090.

 

[Daven]

Yet some still believe that 5080 will be faster than 4090.

Seems unlikely since the 5080 has an 84 SM count and the 4090 has 144. Everyone might be confused by the DLSS4 only benchmarks provided by Nvidia.

 

[TSiAhmat]

Am I crazy or does this feel like a refresh similar to raptor lake 13k -> 14k, with some extra software as a side main dish. (edit: now that I thought about it... that's the main dish)

Just a bit more juice for all the cards and if your lucky you get some cores on top...

Actually pretty disappointed. But I wasn't in the market for a Nvidia gpu anyway. (expect if they offered a crazy good card PerformanceTo€ wise)

Ohhhh well... let's hope it gets better in the next 4 years... (send help)

 

[JustBenching]

How can it have less raster performance?

 

[Daven]

How can it have less raster performance?

Some of the SKUs have less cores down the stack than the Supers in the 4000 series. Also rumors point to same or less cores in the 5060 series.

Of course you can also have less raster performance if clocks are lowered by more than cores are increased.

Nvidia added a lot more AI/RT/DLSS goodness in hardware. Since the 4 nm node is exactly the same between the 4000 and 5000 series, it looks like Nvidia is being stingy on cores and clocks to compensate.

 

[TechBuyingHavoc]

Some of the SKUs have less cores down the stack than the Supers in the 4000 series. Also rumors point to same or less cores in the 5060 series.

Of course you can also have less raster performance if clocks are lowered by more than cores are increased.

Nvidia added a lot more AI/RT/DLSS goodness in hardware. Since the 4 nm node is exactly the same between the 4000 and 5000 series, it looks like Nvidia is being stingy on cores and clocks to compensate.

Expect more of this as process nodes get slower and slower to develop and release to market. Profits must go up and if there is no new node to improve performance/power/area, the die area is going to shrink and shrink.

 

[Zyll Goliat]

Yep as many people predicted seems like 5090 is going to be around 30%-35% faster in raster then 4090 but when it comes to the 5080 Vs 4080 situation is even worst most likely only around 15% here the latest benchmark leaks from 2 games where wasn't used frame generation Resident Evil and Horizon Forbidden West

P.S. Note that in Resident Evil Ray Tracing was enabled and in Horizon Forbidden West DLSS was ON....so there is possibility that if RT and DLSS was Off performance increase could be even smaller

 

[mb194dc]

Yes, it's a margin expansion exercise pretty much. Only the 5090 looks interesting... It'll probably be a decent card still in 5 years or even 10 given how progress has slowed down. So you can kind of justify the 2k for it in that way.

Pretty depressing times in tech.

 

[N/A]

5070 Ti also worth it with 128 rope and 64 mo, if true. 4070 TiS was heavily cut to 96 rop and 48 cache. And yeah it's pretty much over now N5 to N3 30% chip density or 378~ shrinks to 290 mm sq, and N3 to N2 with diminishing returns 15%, next A16 with 7% backside power and nano sheet. But you can safely buy now and enjoy long years of 4k60 with 4x DLSS MFG 240 Hz OLeD without the risk of it depreciating and becoming e-waste by the year's end like my 2080 Ti.

 

[Daven]

Yep as many people predicted seems like 5090 is going to be around 30%-35% faster in raster then 4090 but when it comes to the 5080 Vs 4080 situation is even worst most likely only around 15% here the latest benchmark leaks from 2 games where wasn't used frame generation Resident Evil and Horizon Forbidden West

View attachment 380215

P.S. Note that in Resident Evil Ray Tracing was enabled and in Horizon Forbidden West DLSS was ON....so there is possibility that if RT and DLSS was Off performance increase could be even smaller

RT and DLSS performance most likely will be higher but pure rasterization performance increases are in doubt.

5070 Ti also worth it with 128 rope and 64 mo, if true. 4070 TiS was heavily cut to 96 rop and 48 cache. And yeah it's pretty much over now N5 to N3 30% chip density or 378~ shrinks to 290 mm sq, and N3 to N2 with diminishing returns 15%, next A16 with 7% backside power and nano sheet.

Increases in ROPs will be interesting. That might bring higher performance than my pessimism was thinking.

 

[Vayra86]

Seems unlikely since the 5080 has an 84 SM count and the 4090 has 144. Everyone might be confused by the DLSS4 only benchmarks provided by Nvidia.

Nope they said and I quote 'but the node is better, and there is always a gain every gen regardless'.

DLSS4 wasn't even in the picture back then. Gonna be a rough wake up call... yes the market is indeed fubar

 

[Dr. Dro]

Seems unlikely since the 5080 has an 84 SM count and the 4090 has 144. Everyone might be confused by the DLSS4 only benchmarks provided by Nvidia.

Only 128 out of 144 are enabled. There are no SKUs with the full AD102 chip, RTX 6000 Ada Generation (professional card) maxes out at 142 SM. The 4090 also only has 72 out of the 96 MB of L3 enabled from AD102.

 

[Daven]

Only 128 out of 144 are enabled. There are no SKUs with the full AD102 chip, RTX 6000 Ada Generation (professional card) maxes out at 142 SM. The 4090 also only has 72 out of the 96 MB of L3 enabled from AD102.

Oh you are right. The Videocardz table is showing SMs for the fully unlocked die. The sentiment is still there. It is doubtful that 128 SM 4090 will be beaten by the 84 SM 5080.

 

[Vayra86]

Yep as many people predicted seems like 5090 is going to be around 30%-35% faster in raster then 4090 but when it comes to the 5080 Vs 4080 situation is even worst most likely only around 15% here the latest benchmark leaks from 2 games where wasn't used frame generation Resident Evil and Horizon Forbidden West

View attachment 380215

P.S. Note that in Resident Evil Ray Tracing was enabled and in Horizon Forbidden West DLSS was ON....so there is possibility that if RT and DLSS was Off performance increase could be even smaller

As predicted pre release... : at best 15% from clocking... maybe 5% from optimizations and/or reduced bottlenecking in VRAM.