Nvidia Graphics Card With 7,936 CUDA Cores, 32GB HBM2 May Be Coming

[harlan4096]

An unidentified Nvidia graphics card has landed on Geekbench, and the submission points to a whopping CUDA core count.

An Nvidia graphics card with a huge core count was apparently benchmarked with Geekbench 5 before any official announcement from the manufacturer, as spotted by a Twitter user. We were already taken aback by two unidentified Nvidia graphics cards last week listed with 7,552 and 6,912 CUDA cores. But this latest finding is mind-blowing, with the CUDA core count allegedly reaching 7,936.

Since this is hardware that Nvidia hasn't even confirmed yet, we should take the benchmark results with a grain of salt. Interestingly enough, the Geekbench 5 submission is from the same time frame as the submission for those other two unnanounced Nvidia cards. The system used even has the same CPU, an Intel Core i7-8700K, and motherboard, an Asus Prime Z370-A.

Given the scale of the specifications, all three unnanounced graphics cards potentially hail from the Quadro or Tesla families -- more likely the latter. What we can infer based on the specs shared on Geekbench is that the cards target workstation and data center users as opposed to gamers. The graphics cards are rumoredly based on Nvidia's upcoming Ampere microarchitecture using TSMC and Samsung's 7nm facilities.

The table below compares specs shared for the recently discovered Nvidia graphics card (Nvidia GPU 3), as well as the two unannounced cards we covered last week (Nvidia GPU 1 and 2).

Specs

The Nvidia graphics card in question allegedly has 124 Stream Multiprocessors (SMs). For the sake of discussion, let's assume that Nvidia doesn't modify the number of CUDA cores per SM on its next GPU architecture. At 64 CUDA cores inside each SM, the nameless Nvidia graphics card would have a whopping 7,936 CUDA cores, putting the Tesla V100S' 5,120 CUDA cores to shame.

The graphics card finished the Geekbench 5 benchmark with a 1.10 GHz base clock, but early engineering samples aren't representative of the final specifications. A relatively lower base clock is expected, however, as data center products don't operate at high clock speeds. The Tesla V100S, for example, ticks at a 1.245 GHz base clock.

According to the Geekbench 5 report, there's 32MB of L2 cache on the graphics card. For comparison, the Tesla V100S has a 6MB L2 cache, so 32MB would be a pretty substantial upgrade.

The graphics card was also listed with 32GB of ECC memory. It's allegedly running at 1.2 GHz across a 4,096-bit memory interface, which points to the presence of HBM2.

With the coronavirus turmoil, Nvidia isn't canceling GTC 2020, its annual GPU event, but is instead holding the event online. That means we'll still see CEO Jensen Huang deliver his keynote address. And since Nvidia has gone out of its way to ensure that the GTC show goes on, we suspect it has something extremely interesting to share with the hardware world.
...

Continue Reading