Nvidia May Be Prepping a Massive GPU With 7,936 CUDA Cores, 32GB HBM2
Nvidia may be planning to bring massive new GPUs to market when it makes the jump to 7nm, possibly in a bid to protect its supercomputing market share from any products Intel might launch. A number of new Nvidia GPUs have popped up on places like Geekbench in recent weeks, all with massive core counts: 6,912, 7,552, and as of this week, 7,936.
Currently, Nvidia’s top-end GPU in terms of core count is the Tesla V100, with 5,102 cores. Even the 6,912-core card would be a 1.35x increase, while the 7,936-core GPU is a 1.55x increase. The chart below by Tom’s Hardware shows the relative configuration of each part.
Each of these chips has a lower base clock than Nvidia’s current Volta or Tesla products, but that’s not at all surprising. AMD and Nvidia both tend to target specific thermal envelopes with their products. It’s often more power-efficient to run a larger GPU at a lower clock as opposed to relying on a smaller GPU at higher clock. It’s also possible that Geekbench isn’t detecting clock properly or that these cards are low-clocked engineering samples.
If we assume that these clocks are final, it would mean the 7,936-core GPU runs at 72 percent the speed of the Nvidia Quadro RTX 8000 (4608 cores, 1395MHz base clock), but offers 1.72x the cores. RAM capacity is significantly lower (32GB versus 48GB) but we can assume this hypothetical Nvidia GPU would have more overall memory bandwidth than the current top-end Quadro.
One of the ways to evaluate the likelihood of a rumor or leak is to consider whether a product makes sense in a company’s current lineup. Here, we see evidence that Nvidia may have used the 7nm node to dramatically increase the top-end number of GPU cores it ships. That tracks with 7nm offering a substantial density improvement over 12nm FinFET. Not everyone needs 48GB of memory — 32GB of ultra-high-bandwidth HBM2 might be a better fit for some workloads. Cutting GPU clock while boosting core count is a known strategy and ultra-wide GPUs tend to have lower clocks than their narrower cousins elsewhere in the stack. Finally, we know that Intel is planning to make a serious attack on the data center market. Nvidia has good reason not to sit on its laurels in 2020, and it looks like the company designed a part loaded for bear.
Of course, take this rumor with a grain of salt. Even if it’s a real GPU, the stats recorded here could be preliminary.