Recent reports suggest that NVIDIA plans to continue using the monolithic die approach for their upcoming Blackwell GPU architecture, featured in the GeForce RTX 50 series graphics cards. It indicates that NVIDIA is not concerned about AMD’s chipset strategy, as they successfully released Navi 31 “RDNA 3” during the first consumer-chipset generation. A monolithic die approach is expected to provide several benefits, including efficient product manufacturing.
Kopite7kimi on Twitter had previously disclosed that Ada-Next, also known as Blackwell, might adopt a monolithic chip architecture. NVIDIA seems to be sticking to monolithic GPU designs for their consumer and server products. Meanwhile, AMD may fully adopt chiplet technology for their RDNA 4 generation, although they have already implemented chiplet GPUs for their server offerings.
There are rumors that NVIDIA intends to use TSMC’s 3nm process node for its upcoming Blackwell GPUs, which would be a departure from their use of node shrinks in previous generations.
According to unconfirmed reports, TSMC began mass-producing 3nm wafers in Q4 2022, and the new node is expected to surpass initial expectations. However, the production cost for 3nm wafers is significantly higher than that of 5nm wafers, potentially up to 25% more expensive. As a result, the Blackwell series graphics cards using the 3nm process node may be priced higher. NVIDIA’s CEO is rumored to have visited TSMC’s CEO in Taiwan to secure early access to 3nm wafers for their next-gen GPUs.
ADA-next seems to be another monolithic chip.
— kopite7kimi (@kopite7kimi) July 19, 2022
RedGamingTech recently shared insider information regarding the possible specifications of NVIDIA’s rumored Blackwell GPUs. These sources suggest that the CUDA architecture will undergo significant changes, although it will not be entirely reworked. The new SM units in Blackwell GPUs are expected to have a modified structure. Various enhancements and features will be incorporated into the hardware units responsible for denoising in Ray/Path Tracing.
The Blackwell GPUs from NVIDIA, which are expected to be released in the future, may feature more advanced hardware components to handle Path Tracing effectively, based on NVIDIA’s focus on this technique in high-end games. We may see an upgrade to the existing Ray Tracing units, which could be replaced with even more advanced Path Tracing units, but the details still need to be discovered. In any case, the prospects for improved performance in this area are exciting.
Reports suggest that NVIDIA may opt for GDDR7 memory in their upcoming Blackwell GPUs instead of the current GDDR6X memory. The decision may be based on the superior efficiency of GDDR7. Furthermore, Samsung is developing GDDR6W DRAM, which is anticipated to offer even more outstanding performance and capacity than GDDR6X. While NVIDIA has previously used Micron’s technology for its consumer-end chips, it has relied on Samsung’s HBM solutions for its high-performance computing and AI chips.
Since Cadence has already introduced verification testing solutions for the new GDDR7 memory standard and Samsung has announced transfer speeds of up to 36 Gbps with PAM3 signaling on GDDR7 dies, the Blackwell GPUs may support GDDR7 memory. Micron is also expected to unveil its GDDR7 solution soon. However, it should be noted that we are still a year away from 2024, so there is a possibility that an interim refresh using GDDR6W or GDDR6X, such as the rumored RTX 4090 Ti, may occur before the launch of the Blackwell GPUs.
Kopite7kimi has speculated that NVIDIA might introduce a 512-bit SKU for the Blackwell GPUs, which would be a departure from their recent GPU designs. The last time NVIDIA used a 512-bit card was with the GTX 285 in 2009. However, if NVIDIA decides to release a monster Blackwell GPU, we could achieve over 2 TB/s of bandwidth using 36 Gbps dies. The bandwidth would range from 576 GB/s to 2304 GB/s depending on the bit size of the SKU, with the latter being achieved by a potential 512-bit SKU.