No proof for density improvements 4N vs 4NP. Yes, B100 has around 123MTr/mm2 vs 98MTr/mm2 GH100 but AD102 has 125MTr/mm2. Could be just B100 being more dense packed and GB202 will be at best 130MTr/mm2. So GB202 will need to be 800mm2+~ for 192SM 128MB L2 512bit.
Turing vs Pascal all over again.
If top dog from RDNA4 lineup can't beat 7900XTX then 5090 will get GB203 with 256bit bus and 32Gb/s 3GB Micron GDDR7.
My wild prediction
GB202 216SM 384bit 1536GB/s
Quadro Blackwell 72GB 212SM
Titan Blackwell 36GB 192SM
GB203 128SM 256bit 1024GB/s (5090 24GB)
Bonus
Who is Nvidia fan? :D I just posted the hybrid raster + RT from 3Dmark showing that if Navi31 will have 3X the perf of 6900XT in RT + Raster then RTX 4090 will cooked.
2X pure raster over 6900XT = Nvidia cooked in raster
3X in RT + Raster = Nvidia cooked everywhere.
Fastest 6900XT vs Fastest RTX 4090
That's raster + RT benchmark
Navi31 with 12288 shaders and 3ghz+ clocks should be 2x faster in raster and 2,5-3X in hybrid RT+Raster imho.
This is why you get 192bit bus instead of 256bit and 300mm2~ TSMC 5nm instead of 400mm2 Samsung 8nm.
Going by die size and bus width the true succesor to 500$ RTX 3070 will be RTX 4080 700-800$ MSRP
This site uses cookies to help personalise content, tailor your experience and to keep you logged in if you register.
By continuing to use this site, you are consenting to our use of cookies.