Nvidia’s GPU Know-how Convention (GTC) is underway. Throughout CEO Jensen Huang’s keynote, particulars of Nvidia’s subsequent technology Hopper architecture have been revealed. Although it is an AI and information centre centered GPU, it provides us a couple of hints of what we are able to anticipate from Nvidia’s gaming-oriented Ada Lovelace GPU structure, which is due for launch later in 2022.
The H100 is a serious step ahead over the present flagship A100. The total GPU comprises 80 billion transistors or 26 billion extra that the A100. It’s constructed on a customized TSMC 4nm course of. It helps as much as 80GB of HBM 3 reminiscence delivering as much as 3 TB/s of bandwidth.
The H100 helps PCIe 5.0 and NVLink for connecting a number of GPUs collectively. It will possibly ship 2,000 TFLOPS of FP16 and 1,000 TFLOPS of TF32 efficiency, triple that of the A100. Hopper introduces a brand new instruction set known as DPX. It’s designed to speed up efficiency in fields as different as illness analysis, quantum simulation, graph analytics and routing optimizations.
The total H100 GPU consists of 18432 CUDA cores and 576 Tensor cores. That compares to the A100 with 8192 and 512 respectively, although for not the entire cores are unlocked, presumably to maximise yields. The core clocks are additionally not finalised. Regardless of being fabricated on such a complicated node, the SXM model of the H100 comes with a TDP of 700W. That’s proper, seven. hundred. watts.
The H100 is ready to be a monster of a card, however is it related to PC avid gamers? The reply is sort-of. H100 is all about compute efficiency and never graphics, however we are able to take some bits of knowledge and use it to foretell what the gaming model may appear to be.
The transfer to a customized TSMC 4nm node is a serious step ahead over the Samsung 8nm course of used for the RTX-30 collection. It is seemingly for use for RTX-40 collection playing cards too. Additionally noteworthy is assist for PCIe 5.0. Although by itself it is not anticipated to ship any actual efficiency profit over PCIe 4.0, it could properly do over PCIe 3.0 which remains to be extensively in use throughout many gaming techniques.
However maybe the most important nugget of all is the somewhat astonishing 700W TDP of the high-end configuration. Simply have a look at the VRM of that card! 700W for a knowledge centre product is one thing that may be managed, but when we get something like that for a flagship RTX 4090 then we’d be shocked. Sadly, rumours of steep increases in power consumption proceed to floor. Even 500W is a soar and it implies that 4 slot graphics playing cards could turn into the norm, on the prime finish of the market anyway.
Nvidia remains to be engaged on the H100. If its principal traits are shared with the RTX 40 collection, it’s honest to say that the excessive finish playing cards can be sizzling and energy hungry, however packed filled with tech and far sooner than the RTX 3090 (and the soon to be released RTX 3090 Ti). AMD will compete with its RDNA3 based cards and it’s shaping as much as be a hell of a battle, with all out efficiency clearly being a precedence for each firms on the expense of energy effectivity. We will’t wait!
