GamingGPUsNews

NVIDIA RTX 4090 to Draw up to 600W Power at Max (Not 800-900W), May Leverage 4nm Node From TSMC [Report]

NVIDIA launched its next-gen data center graphics card the other day in the form of the H100 (GH100), giving us a first look at the Hopper architecture. Leveraging TSMC’s N4 4nm process node on a massive die, it doubles down on Ampere’s compute capabilities. From what we’ve heard about the GeForce RTX 4080/4090 (Ada Lovelace), it’d be fair to say that the AD102 will be a derivative of the GH100.

Data Center GPUNVIDIA Tesla P100NVIDIA Tesla V100NVIDIA A100NVIDIA H100
GPU CodenameGP100GV100GA100GH100
GPU ArchitectureNVIDIA PascalNVIDIA VoltaNVIDIA AmpereNVIDIA Hopper
SMs5680108132
TPCs28405466
FP32 Cores / SM646464128
FP32 Cores / GPU35845120691216896
FP64 Cores / SM32323232
FP64 Cores / GPU1792256034568448
INT32 Cores / SMNA646464
INT32 Cores / GPUNA512069128448
Tensor Cores / SMNA8424
Tensor Cores / GPUNA640432528
Texture Units224320432528
Memory Interface4096-bit HBM24096-bit HBM25120-bit HBM2512-bit x5
Memory Size16 GB32 GB / 16 GB40 GB128GB?
Memory Data Rate703 MHz DDR877.5 MHz DDR1215 MHz DDR1600 MHz DDR?
Memory Bandwidth720 GB/sec900 GB/sec1555 GB/sec?
L2 Cache Size4096 KB6144 KB40960 KB60MB
TDP300 Watts300 Watts400 Watts700W
TSMC Manufacturing Process16 nm FinFET+12 nm FFN7 nm N74 nm N4

Both feature a maximum of 144 SMs or 18,432 cores across 12 GPCs and 72 TPCs. The IN32: FP32 cores exist in a 1:2 ratio per SM, much like Ampere (with the FP64 cores disabled on Ada). The only tangible differences are with respect to the L2 cache and the memory controllers. Lovelace should feature GDDR6X/GDDR7 controllers while Hopper uses HBM2e. The former is expected to pack up to 96MB of L2 cache while the latter is limited to 60MB.

Other than that, the two graphics architectures have the same floorplan, and it won’t be surprising if NVIDIA uses the N4 node for the AD102 as well. Looking at NVIDIA’s spending on foundry costs over the past months, it’s looking more and more likely that Lovelace will be an N4 die. Either way, the PPA difference between N5 and N4 is minimal at best, and shouldn’t affect the final product or performance.

GPUTU102GA102AD102AD103AD104
ArchTuringAmpereAda LovelaceAda LovelaceAda Lovelace
ProcessTSMC 12nmSam 8nm LPPTSMC 5nmTSMC 5nmTSMC 5nm/ 4nm
GPC671275
TPC3642724230
SMs72841448460
Shaders4,60810,75218,43210,7527,680
TP16.137.6~90 TFLOPs?~50 TFLOPs~35 TFLOPs
Memory11GB GDDR624GB GDDR6X24GB GDDR6X16GB GDDR616GB GDDR6
L2 Cache6MB6MB96MB64MB48MB
Bus Width384-bit384-bit384-bit256-bit192-bit
TGP250W350W600W?350W?250W?
LaunchSep 2018Sep 2020Aug-Sep 2022Q4 2022Q4 2022

Then, there’s the matter of power consumption. There have been several rumors claiming that NVIDIA’s next-gen RTX 4080/4090 graphics cards will have an abysmal power consumption of up to 700-800W. As already stated in earlier posts, this is highly unlikely and the TBP of the Ada Lovelace GPUs should top out at 600W, with only extreme overclocker cards such as the RTX 4090 Kingpin coming close to this limit. Toms from MLID has gotten similar hints from his sources:

Areej

Computer hardware enthusiast, PC gamer, and almost an engineer. Former co-founder of Techquila (2017-2019), a fairly successful tech outlet. Been working on Hardware Times since 2019, an outlet dedicated to computer hardware and its applications.
Back to top button