In the production cycle of any upcoming product, many details aren’t decided until later stages. Even significant details in specifications may not have been finalized till a couple of months before the launch.
That’s what has been happening with NVIDIA’s next-gen RTX 40 series GPUs, especially the GeForce RTX 4070. NVIDIA has already changed the card’s specifications once before, offering 12 GB GDDR6 memory with a 192-bit interface. But it seems like there is another change in plans.
According to leaker Kopite7kimi, NVIDIA has two different specifications planned for RTX 4070, one with 12 GB memory and a 192-bit interface, and the other with a lower 10 GB memory and 160-bit interface.
The leaker mentions that there are currently four SKUs in contention: 340, 341, 336, and 337, with a separate board design. However 340/341 and 336/337 pairs denote the same SKU and only one out of each will likely be used for the reference design (Founders Edition), while the other will go to semi-custom designs by board partners.
The 340/341 is supposedly based on full AD104 GPU (AD104-400) with 7680 FP32 (CUDA) cores. This model supposedly features 12GB of GDDR6X memory across a 192-bit bus and 21 Gbps modules, resulting in a maximum bandwidth of 504 GB/s.
336/337 on the other hand features a cut-down AD104 GPU (possibly AD104-275) with 56 Streaming Multiprocessors, so 7168 CUDA cores. Furthermore, this card would have a limited memory size of 10 GB and as a result narrower bus width to 160-bit. This SKU will have a bandwidth of 420 GB/s.
There aren’t any changes in the 340/341 variant compared to the previously known specs, however, it seems the Total Board Power will be 285W, and 250W for the 336/337 variant. Note that TBP is not the same as TDP, and the earlier leaked TDP of 300W might still be true. Nevertheless, the GPU will still consume more power than RTX 3070.
Till now, it seems that NVIDIA has got the performance spot on, but it still needs to somehow work on the power consumption as it is just too high. Hopefully, we might see a lower TDP rating by the time the GPU launches in the fourth quarter of this year.
NVIDIA Corporation is an American multinational technology company incorporated in Delaware, based in Santa Clara, California. They design graphics processing units (GPUs) for the gaming and professional markets, as well as system on chip units (SoCs) for the mobile computing and automotive market.
Best known for the “GeForce” lines of GPUs, they are a direct competitor to AMD’s “Radeon” series. NVIDIA has also expanded its offerings with its handheld game consoles Shield Portable, Shield Tablet, and Shield Android TV and its cloud gaming service GeForce Now.