NVIDIA's next-gen GeForce RTX 50 series graphics cards are expected to utilize the new Blackwell GPU architecture. Recently, there are lots of rumors popping up regarding the chip so we decided to round them up and see what Blackwell is shaping up to be.
NVIDIA GeForce RTX 50 "Blackwell" GPU Rumored To
The latest rumors come from TechTubers, RedGamingTech & Moore's Law is Dead, who published their recent videos talking about NVIDIA's next-gen architecture. Since Blackwell is still a few years away from launch, this information isn't final and rumors are meant to be taken with a grain of salt though still interesting because more renowned leakers have made similar statements in the past.
Based on previous information, we already know that Blackwell is the codename for NVIDIA's next-gen GPU architecture that is meant to succeed Hopper, and unlike Hopper, Blackwell comes with a consumer-esque GB102 GPU in addition to the GB100 GPU which was leaked a while back.
There have also been reports that Blackwell GPU architecture will be introduced in 2024 which means that GTC 2024 might be the first server-level introduction before the architecture ships out to GeForce cards.
NVIDIA Blackwell GPUs To Retain Monolithic Die
So starting off with the details, the NVIDIA GeForce RTX 50 series that utilizes the Blackwell GPUs is going to retain a monolithic die approach. Moore's Law is Dead states that NVIDIA isn't worried by AMD's chiplet strategy given what they were able to release during the first consumer-chipset gen in the form of Navi 31 "RDNA 3". The green is expected to keep on utilizing a monolithic die which still retains more advantages with the biggest one being having the product manufactured on time.
ADA-next seems to be another monolithic chip.
— kopite7kimi (@kopite7kimi) July 19, 2022
Maybe Blackwell.
— kopite7kimi (@kopite7kimi) July 19, 2022
Last year, Kopite7kimi also revealed that Ada-Next, Blackwell, might be another monolithic chip. It looks like NVIDIA has so far taken the rear seat in chiplet designs as both its consumer and server offerings still use monolithic GPUs. AMD may go the full chiplet route in RDNA 4 generation but the chipmaker is making chiplet GPUs for its server offerings.
NVIDIA To Utilize TSMC 3nm Process Node To Make Blackwell GPUs?
Update: Kopite7limi states that NVIDIA is not going to use a 3nm process node but something else.
I'm afraid it won't use 3nm.
— kopite7kimi (@kopite7kimi) April 10, 2023
The second rumor is that NVIDIA will be utilizing TSMC's 3nm process node to manufacture the Blackwell GPUs. So far, all recent GPU launches by NVIDIA have been a node shrink starting with the Pascal GPUs that were made on TSMC's 16nm, Turing on TSMC's 12nm, Ampere on Samsung 8nm, and the current Ada GPUs utilize the TSMC 4N (5nm Optimized) node.
TSMC commenced volume production of its 3nm node in Q4 2022 and the node is stated to be better than expected, as per rumors. However, just as impressive as the node might sound, it's going to be super expensive too. According to insiders, TSMC 3nm wafers cost 25% more than TSMC's 5nm wafers which might result in increased prices for next-gen GPUs such as the NVIDIA GeForce RTX 50 "Blackwell" series. NVIDIA's CEO also reportedly visited Taiwan to talk with TSMC's CEO about securing 3nm wafers early on for their next-gen GPU lineup.
NVIDIA GeForce RTX 50 "Blackwell" Graphics Cards Early "Rumored" Specs
Next up, we have a ton of information from RedGamingTech who have received it from their own sources. The details start off by stating that Blackwell GPUs will feature massive overhaul to CUDA architecture though this isn't a ground-up design to say. The Blackwell SM units are expected to get a "New Structure" and there will be further optimizations and additions made to the denoising and Ray/Path Tracing hardware units.
Considering NVIDIA's recent push for Path Tracing in modern AAA titles, it looks like we will get even stronger on-chip hardware to process them. Maybe RT units will be replaced with more enhanced PT units, anything's possible, right?
The Blackwell GPUs are said to feature support for GDDR7 memory and while GDDR6X still has some room to spare but if GDDR7 provides better efficiency then it would definitely be a better choice. Now Samsung is working on its brand new GDDR6W DRAM which should provide even better performance & higher capacities though it's been a while since NVIDIA went back to the Sammy team on the consumer-end chips since leveraging Micron's technology. NVIDIA is using Samsung HBM solutions for its HPC and AI chips.
With that said, GDDR7 doesn't sound that far-fetched at all considering Cadence has already introduced the first solutions for verification testing of the new memory standard. Even Samsung announced that it will offer up to 36 Gbps transfer speeds with PAM3 signaling on GDDR7 dies & Micron is also expected to unveil its solution sooner or later. We are still a full year away from '24 so an intermediate refresh with GDDR6W and GDDR6X may still be possible (*cough* RTX 4090 Ti *cough*).
It's too early to discuss Blackwell now. I won't be surprised if it has a 512-bit memory interface.
— kopite7kimi (@kopite7kimi) April 7, 2023
Kopite7kimi also raises the possibility that Blackwell might feature a 512-bit SKU though NVIDIA has gone that route in a long time. The last 512-bit card from the vendor was the GTX 285 all the way back in 2009. But just in case there's a monster Blackwell GPU designed by NVIDIA, then we can get over 2 TB/s of bandwidth using 36 Gbps dies:
- 128-bit @ 36 Gbps: 576 GB/s
- 192-bit @ 36 Gbps: 846 GB/s
- 256-bit @ 36 Gbps: 1152 GB/s
- 320-bit @ 36 Gbps: 1440 GB/s
- 384-bit @ 36 Gbps: 1728 GB/s
- 512-bit @ 36 Gbps: 2304 GB/s
The NVIDIA Blackwell GeForce RTX 50 series is also rumored to utilize a PCIe Gen 5 interface and offer clock speeds of 3 GHz+. The GB102 GPU should be the flagship of the gaming lineup and is said to rock 144 SMs, and 96 MB of L2 cache. That's the same amount of SMs and cache on the fully-enabled Ada AD102 GPU but it looks like the underlying changes to the architecture will be providing a significant boost in performance.
NVIDIA Blackwell GPUs To Offer Another 2x Increase Over Ada?
Finally, the performance is said to be another 2x leap over Ada GPUs. The Ada GPUs themselves offered a massive jump over the previous gen such as the RTX 4090's cut-down AD102 core but there were significant uplifts in all segments (Rasterization / Ray Tracing / DLSS). Further updates can push Blackwell GPUs to new heights and we can wait to see what NVIDIA has in store for the gaming community.
