NVIDIA Allegedly Reduces RTX 4090 Production in Favor of the H100 Data Center GPU

NVIDIA has allegedly trimmed the production of its AD102 graphics die powering the Ada Lovelace flagship. The GeForce RTX 4090 has been in short supply since it launched earlier this month. Newegg has already run dry of Love(lace), and Amazon has a few units going at inflated pricing (>$2,300). The AD102 production has supposedly been reduced to accommodate the increasing demand for the H100 data center accelerator.

GPU	GeForce RTX 4090	GeForce RTX 4080 16G	~~GeForce RTX 4080 12G~~-> RTX 4070 Ti	RTX 4070	GeForce RTX 3090 Ti
GPU Name	AD102	AD103	AD104	AD104	GA102
SMs	128	76	60	46	80
FP32 Cores	16384	9728	7680	5888	10240
TMUs/ROPs	512 / 176	320 / 112	240 / 80	184/64	320 / 112
Tensor/RT Cores	512 / 128	304 / 76	240 / 60	184/46	320 / 80
Base Clock	2230 MHz	2210 MHz	2310 MHz		1365 MHz
Boost Clock	2520 MHz	2510 MHz	2610 MHz		1665 MHz
Memory Buffer	24 GB G6X	16 GB G6X	12 GB G6X	12GB G6X	24 GB G6X
Bus	384-bit	256-bit	192-bit	192-bit	384-bit
Memory Speed	21Gbps	23Gbps	21Gbps	21Gbps	19Gbps
Bandwidth	1008 GB/s	736 GB/s	504 GB/s	504GB/s	912 Gbps
L2 Cache	72MB	64MB	48MB	36MB	6MB
TBP	450W	320W	285W	250W	350W
MSRP	$1599	$1199	$899	$649?	$1199
Launch	Oct 2022	Nov 2022	5th Jan 2023		3rd June 2021

The partially enabled AD102 die powers the GeForce RTX 4090 with 16,384 FP32 cores (128 SMs) alongside 24GB of GDDR6X memory via a 384-bit bus and a 72MB LLC buffer. It has a memory bandwidth of 1,008GB/s with a TGP of 450W. The GPU is fabbed on TSMC’s N4 node.

The GH100 core utilized by the H100 accelerator is also produced using the N4 process. It integrates 80 billion transistors and CoWoS 2.5D packaging technology. It packs 18,432 FP32 cores, 60MB L2 cache, and six HBM3 stacks across a 6,144-bit bus. The SMX variant of the H100 features 15,872 cores, while the PCIe version comes with 14,952 and a TGP of 700W.

Source: Mydrivers (Via: @Harukaze)