NVIDIA and AMD take different paths with their graphics accelerators

NVIDIA and AMD have already shown their cards. Both have presented their new graphic accelerators for the professional sector, and with this we have been able to confirm that they have decided to follow different paths. It had been rumored that Hopper, NVIDIA’s new graphics architecture for the professional sector, was going to use an MCM-type design, but in the end it didn’t.

The NVIDIA H100 uses a design of monolithic core, and the H100 GPU it mounts has 16,896 active shaders out of a total of 18,432 shaders. It is a huge chip, both in terms of size and density of transistors, and of course it represents a significant leap compared to the previous generation. For comparison purposes, just remember that the NVIDIA A100, which is the predecessor of the NVIDIA H100, only has 6,912 shaders.

That NVIDIA has been able to design a chip with 18,432 shaders, and that this has been viable at wafer level under TSMC’s 4nm node is a real feat. We have already explained to you on previous occasions what exactly this means, and why it is so complicated to transfer a design with such a high number of shaders to the wafer: because having such a high number increases the chances that one of them will end up being unusable, and in that case the chip can no longer be used for its intended purpose.

NVIDIA and AMD (2)

In contrast, AMD has chosen to use an MCM design with the Radeon Instinct MI250, a high-performance graphics accelerator that combines two GPUs to create a super GPU that adds 13,312 shaders. With these numbers on the table, we can already see that the NVIDIA solution brings together a greater number of shaders, and that it does so with a monolithic core design.

The NVIDIA H100 represents a major advance in terms of design and at the silicon level for that monolithic core concept, and has also been confirmed as a much more powerful solution. We just have to see the maximum power values ‚Äč‚Äčthat each one reaches: the NVIDIA H100 reaches 2,000 TFLOPs in FP16, while the Radeon Instinct MI250 stays in the 362.1 TFLOPs. At FP64 the NVIDIA solution has a peak of 64 TFLOPswhile the AMD option reaches 45.26 TFLOPs.

There is no doubt that NVIDIA has won the game with its Hopper architecture so far, and it is clear that the GH100 is the most advanced and most powerful GPU out there.

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *