Nvidia dusts off dual GPU cards

Stefan Mileschin · 23rd March 2023, 07:56

At least for data centres

Nvidia has announced a new dual-GPU product, the H100 NVL but sadly not for SLI or multi-GPU gaming.

In fact based on what Nvidia says the H100 NVL (H100 NVLink)will be a rubbish card for gaming and targets the growing AI market.

From the information and images Nvidia has released, the H100 NVL will have three NVLink connectors on the top, with the two adjacent cards slotting into separate PCIe slots. This will accommodate servers that don't support Nvidia's SXM option, with a focus on inference performance rather than training.

The NVLink connections should help provide the missing bandwidth that NVSwitch gives on the SXM solutions, and there are some other notable differences as well.

Previous H100 solutions such as the SXM and PCIe came with 80GB of HBM3 memory, but the package contained six stacks, each with 16GB of memory. What we do know is that the H100 NVL will come with 94GB per GPU, and 188GB HBM3 total.

Total performance is double that of the H100 SXM: 134 teraflops of FP64, 1,979 teraflops of TF32, and 7,916 teraflops FP8.

https://fudzilla.com/news/ai/56577-n...dual-gpu-cards

23rd March 2023, 07:56	#1
Stefan Mileschin [M] Reviewer Join Date: May 2010 Location: Romania Posts: 153,575	Nvidia dusts off dual GPU cards At least for data centres Nvidia has announced a new dual-GPU product, the H100 NVL but sadly not for SLI or multi-GPU gaming. In fact based on what Nvidia says the H100 NVL (H100 NVLink)will be a rubbish card for gaming and targets the growing AI market. From the information and images Nvidia has released, the H100 NVL will have three NVLink connectors on the top, with the two adjacent cards slotting into separate PCIe slots. This will accommodate servers that don't support Nvidia's SXM option, with a focus on inference performance rather than training. The NVLink connections should help provide the missing bandwidth that NVSwitch gives on the SXM solutions, and there are some other notable differences as well. Previous H100 solutions such as the SXM and PCIe came with 80GB of HBM3 memory, but the package contained six stacks, each with 16GB of memory. What we do know is that the H100 NVL will come with 94GB per GPU, and 188GB HBM3 total. Total performance is double that of the H100 SXM: 134 teraflops of FP64, 1,979 teraflops of TF32, and 7,916 teraflops FP8. https://fudzilla.com/news/ai/56577-n...dual-gpu-cards

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Nvidia dusts off old Turing stock	Stefan Mileschin	WebNews	0	15th February 2021 08:00
WannaCrypt dusts off Linux desktop debate	Stefan Mileschin	WebNews	0	20th May 2017 15:57
Testing GTX 1080 SLI Performance with Dual Palit GeForce Cards	Stefan Mileschin	WebNews	0	6th July 2016 14:38
Patent troll dusts off SCO attack method	Stefan Mileschin	WebNews	0	3rd December 2015 14:06
Radeon 7990 Dual GPU Cards?	Stefan Mileschin	WebNews	0	24th August 2012 17:59
Head of Nvidia Criticizes Dual-GPU Approach as Dual-GPU GeForce 9800 GX2 Looms	jmke	WebNews	0	14th February 2008 15:17
Alienware Laptop Adds Dual Graphics Cards	Sidney	WebNews	0	27th June 2007 04:48
Geforce 7900 GTX Dual Core cards emerge	jmke	WebNews	0	2nd March 2006 18:25
Dual DVI cards in .BE?	Era²	General Madness - System Building Advice	10	28th June 2005 23:57
Gigabyte to introduce new dual-GPU graphics cards in February	jmke	WebNews	0	3rd February 2005 17:57