Blackwell GPU die sizes revealed; regression in some cases

[agent_x007]

Is Blackwell transistor count accurate there? Blackwell dies seem to have essentially less transistors per SM. Looking at what has been revealed that does not quite make sense. Blackwell should have slightly more transistors. Did they cut down somewhere? Cache?

It may be simply issue of "not-scaling" case (not all things scale the same), also more density = more focus heat, so it's not always good idea to go max. density all the time.

Lastly, since NV marketed "AI" as shader part (with M$ throwing it's support behind this on DirectX side), it's possible tensor cores are being shifted from seperate units to actual workload on shaders (a.k.a. Cuda Cores).
In short, each shader/cuda core may be capable of doing tensor operations (provided by Async compute ).
I'd speculate that to get full capability of "old hardware tensor" from Ada/Ampere gen, an SM level context limit is needed (analogy : CCs are pretty dumb, like car engine - by itself it just can't go anywhere, wheels/axels/steering/etc. are required).
^This is 100% guesswork at this moment, since we don't have a public whitepaper on RTX 50 series (GB202/203/205/206/207) yet.

 

[Zach_01]

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

Hey if 5070=4090 then the 5080 can smoke it single handed... lol
Joking aside this is a move from nVidia that states (what we already knew?) loud and clear that games are not the real "game".
Blackwell is the perfect (so far) architecture for professional/prosumer workloads and MultiFG is the by-product in the try to satisfy gamers too.

Its a multi-billion company that prioritizes multi-million business/contracts. Takes no brain.
Bubble bursting or not, it is what it is.

We cant escape this... There will be a time that the entire game will be created by AI, with endless possibilities.
With all the good and the bad that this will bring with it. Nothing can be all positive or all negative.
I have no doubts about if, only about when.

 

[mama]

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.

Cheer up. Wait for proper reviews.