NVIDIA RTX "Blackwell" GPU with 96 GB GDDR7 Memory on 512-Bit Bus Appears

[AleksandarK]

Recent shipping manifests suggest that NVIDIA works on a graphics card with 96 GB of GDDR7 memory. Documents reveal a product utilizing a 512-bit memory bus and a clamshell (memory on both PCB sides) design that combines two 3 GB modules per memory controller. This setup effectively doubles the memory capacity of existing workstation-oriented cards. The product is believed to use the GB202 chip, the only Blackwell desktop GPU with a 512-bit interface. The documents refer to a board labeled PG153, a designation not seen in any of NVIDIA's existing consumer GPUs. This finding points toward a professional or workstation model rather than a gaming product. There is a possibility that it could be part of the RTX 6000 Blackwell or RTX 8000 Blackwell series.

NVIDIA's current top workstation card, the RTX 6000 "Ada," features 48 GB of memory. A move to 96 GB would be a substantial jump, enabling more complex workloads for content creation, data analysis, and AI. This GPU could carry a significantly higher power target than current workstation models. However, professional GPUs often maintain lower clock speeds to keep power consumption within limits that accommodate more stable operation in professional environments. There is no confirmed information regarding the card's official name or final specifications, such as core count or actual clock frequencies. NVIDIA's workstation GPUs have historically provided a higher core count than their gaming counterparts. If the rumored 96 GB GPU follows this pattern, it may surpass even the potential GeForce RTX 5090, which comes with 32 GB of GDDR7. NVIDIA is expected to hold its annual GPU Technology Conference in March. This event is viewed as a likely venue for official announcements. Until then, these details remain unverified.



View at TechPowerUp Main Site | Source

 

[rattlehead99]

2x 3GB chips per 32 bit controller? Are they clam shelled and stack on both sides of the PCB running at full bandwidth, or are they running in 16 bit mode and bandwidth his halved?

 

[Legacy-ZA]

Yes, it's about time that they got rid of their 2 dimensional thought processes, there is a whole other side of a PCB and with the correct design, you can use the other side too, they are making the products more thick anyways, as long as the pathways don't overheat or interfere with others, why not?

For example, why not go for a 3D hexagon shape with a central node/pathway and an enclosure with water tubes/nitrogen for cooling?

So backward.

 

[rattlehead99]

Yes, it's about time that they got rid of their 2 dimensional thought processes, there is a whole other side of a PCB and with the correct design, you can use the other side too, they are making the products more thick anyways, as long as the pathways don't overheat or interfere with others, why not?

For example, why not go for a 3D hexagon shape with a central node/pathway and an enclosure with water tubes/nitrogen for cooling?

So backward.

It's more expensive, complex, and requires cooling on the back side of the PCB if the memory is clocked high. Probably other reasons too.

 

[Hereticbar]

Yes, it's about time that they got rid of their 2 dimensional thought processes, there is a whole other side of a PCB and with the correct design, you can use the other side too, they are making the products more thick anyways, as long as the pathways don't overheat or interfere with others, why not?

For example, why not go for a 3D hexagon shape with a central node/pathway and an enclosure with water tubes/nitrogen for cooling?

So backward.

The problem with that is it gets really hot, I mean really hot. My RTX 3090 uses the clam shell design with memory chips on both the front and back, at full load the memory temperature can be 10-20 degrees higher than the CPU hotspot.

 

[londiste]

The problem with that is it gets really hot, I mean really hot. My RTX 3090 uses the clam shell design with memory chips on both the front and back, at full load the memory temperature can be 10-20 degrees higher than the CPU hotspot.

There are two types or people - ones who advocate for clamshell memory designs on consumer cards with cutting edge memory technologies and ones who have/had a 3090

 

[Dirt Chip]

Yes, we must have that 10K$ consumer GPU this generation. Must.

 

[HBSound]

I am a fan of the pro GPUs, as I have never had any issues. The jump to 96GB makes sense, especially since the 5090 has 32GB right now, and the next generation of could have 48 GB. At some point, the GPU for gamers and professionals can't sit in the same seat. Nvidia must set them apart for the nature of the business.

But this joker is going to 96GB, so it will not be a card I will use or purchase. For just talking, I would get a couple of 5090s if I needed them for workload for pennies on the dollar in a few years.

 

[freeagent]

There might be 1 person on this forum that I could think of that might buy this card..

@venturi

 

[wNotyarD]

Yes, we must have that 10K$ consumer GPU this generation. Must.

But this isn't a consumer card, it's a workstation model. A Quadro, if Nvidia kept the brand alive.

 

[tpuuser256]

Workstation hardwareoften have better specs. They have more cores but are clocked lower due to the egregious power consumption, but they could very well beat the top consumer hardware with proper overclock and cooling.
It's wayyy too expensive though

 

[InVasMani]

Short riser cables with a vertical oriented GPU on a larger PCB should probably a be thing at some point. They could be using a larger PCB and thus get more out of a GPU and cooling of them. They could even do a bit of a CF/SLI approach around that with multiple short length riser cables connecting to one mammoth GPU PCB motherboard of sorts. It would all be feeding a single GPU PCB so maybe it wouldn't have the complications of CF/SLI does as well, but not entirely sure on the technical aspects of that. Really with a few minor tweaks to the PCIE slots cases could be readily enhanced to better support that type of design too. Anyway that's a lot of capacity for AI and other large data simulation stuff these GPU's will be doing.

That said I guess their goal is as much about more density within a system than it is one more powerful GPU so probably plays a big role in part. It does seem like GPU PCB's could play a more prominent part within overall case space and how much it occupies in the not so far off future. It would be good to see some more forward thinking design approaches and innovation is my thoughts on it personally.

 

[hsew]

Yes, we must have that 10K$ consumer GPU this generation. Must.

That would be the 64GB version

 

[Gmr_Chick]

But this isn't a consumer card, it's a workstation model. A Quadro, if Nvidia kept the brand alive.

The 4090 and 5090 aren't what I'd call "consumer GPUs" either. They are both workstation-class cards masquerading as "gaming" GPUs. But here we are. *shrug*

 

[venturi]

There might be 1 person on this forum that I could think of that might buy this card..

@venturi

actually. ... I'd buy 2.

But not for gaming but for CNN, RCNN

and obviously dependent on pain.. I mean price point