Wednesday, December 25th 2024
NVIDIA GeForce RTX 5090 PCB Pictured, Massive GPU Die and 16-Chip Memory Configuration
NVIDIA's GeForce RTX 5090 graphics card printed circuit board has allegedly been shown in the flesh, showing the memory layout and some interesting engineering choices. The custom PCB variant (non-Founders Edition) houses more than 40 capacitors, which is perhaps not standard on the FE reference board, and 16 GDDR7 memory modules. The leaked PCB, which extends beyond standard dimensions and traditional display connector configurations, is reportedly based on NVIDIA's PG145 reference design. The memory modules are distributed systematically: five on the left, two below, five on the right, and four above the GPU die. The interface is PCIe 5.0 x16.
As NVIDIA has reportedly designated 32 GB GDDR7 memory capacity for these cards, this roughly translates into 16 x 2 GB GDDR7 memory modules. At the heart of the card lies what sources claim to be the GB202 GPU, measuring 24×31 mm within a 63×56 mm package. Power delivery uses a 16-pin 12V-6x2 power connector, as expected. The entire PCB features only a single power connector, so the 16-pin 12V-2x6, but with an updated PCIe 6.0 CEM specification, is the logical choice.
Sources:
Chiphell, @9550pro, via VideoCardz
As NVIDIA has reportedly designated 32 GB GDDR7 memory capacity for these cards, this roughly translates into 16 x 2 GB GDDR7 memory modules. At the heart of the card lies what sources claim to be the GB202 GPU, measuring 24×31 mm within a 63×56 mm package. Power delivery uses a 16-pin 12V-6x2 power connector, as expected. The entire PCB features only a single power connector, so the 16-pin 12V-2x6, but with an updated PCIe 6.0 CEM specification, is the logical choice.
86 Comments on NVIDIA GeForce RTX 5090 PCB Pictured, Massive GPU Die and 16-Chip Memory Configuration
Oof.
For the 5090, with its 512-bit bus, given that each controller is 32-bit, you need a minimum of 16 memory chips, period. Currently on GDDR7 the smallest memory ICs available are 16Gb ones, there are no 8Gb ones.
So yeah, to use less chips you'd need a smaller memory bus, which would decrease the performance.
Good response though. Obviously you know what you're talking about about.
You regular x86 desktop has 64-bit (or 2x32-bit that kinda work as a single unit) controllers, so for your regular consumer CPU with 128-bit (the so called "dual-channel"), you need a minimum of 2 sticks.
Apple's Mx lineup uses 16-bit controllers, and so on and so on.Yeah, but you ended up talking about a different thing. The comment you were replying to was correct still: decreasing the amount of memory modules for the 5090 would imply in a smaller bus, and thus lower performance.
Huge déjà vu of all the arguments from 4 years ago about how the 3090 was not a Titan.
The only "locked" feature on Ampere and Ada was tensor fp16 with fp32 acc, as I had said before, which is not really a "professional" stuff given how other stuff is "unlocked", but it is a market segmentation tactic nonetheless.
Regular fp16 performance has the same rates in both the GeForce and in their professional lineup, that's the point I wanted to correct in your original statement.
If you find these to be obscenely overpriced and useless, that is OK. Others will have a different opinion, and that is also OK.
All other chips, be it geforce, quadro or tesla, lack proper FP64. You'll only be seeing those in the x100-based chips.
Regular FP16 has not been capped between products. Tensor FP16 with FP32 acc has, which is clearly a market segmentation tactic, as I had said before.
I do agree with your points.I guess there's a point in that the titan used to have no artificial limitations whatsoever, while the 3090/4090 did, even if those differences were mostly irrelevant, but were still differences nonetheless.
Still, complaining about this is kinda moot IMO, the GPU delivers the performance of the highest-end available product, has a price tag to match, and fits perfectly for the use of prosumers, like a halo product should be.