Nvidia
We finally have our first gustatory sensation of Nvidia’snext genesis of GPUs , named Blackwell . Sure , they ’re built for enterprise , and no , they wo n’t runCyberpunk 2077(at least not formally ) . But this is the first look we ’ve have at what Nvidia is cooking up for itsRTX 50 - series GPUs , which are rumored to launch sometime in the next class .
The GPU we have today is the B200 — Blackwell 200 , if you may pick out it — that come packed with 208 billion transistor . The architecture is built on TSMC ’s 4NP node , which is an enhanced translation of the 5 nm node . It ’s a little surprising contribute that Nvidia ’s Ada Lovelace GPUs are build with TSMC ’s 4N node — one refinement step away from 4NP . Nvidia notes that it ’s using a custom version of this process , however .
Nvidia
Nvidia laid out six parameters that coif Blackwell aside . The first is that usance 4NP physical process , which Nvidia is using to connect two GPU dies across a 10TB - per - sec interconnect . 2nd is an update Transformer Engine . Nvidia says the architecture can handle double the theoretical account size of last - gen Ada Lovelace , and it supports a newfangled 4 - bit floating head instruction band .
There ’s also 5th - contemporaries NVLink , allowing enterprises to hook up Blackwell GPUs in size of 576 GPUs in a individual fleet . Nvidia is throw in a couple of AI security feature as well , including confidential computing capacity and a RAS ( dependability , availability , and serviceability ) locomotive to ensure maximum uptime for its data center GPU .
That ’s great if you ’re running a data center , but it ’s the sixth and net column that gets me excited . Nvidia says the B200 has a consecrate decompressing engine . This is concenter on enterprise powerful now , but we ’ve fancy how big of a differencededicated decompression hardwarecan make in gaming as well . If Nvidia transport this railway locomotive over to its play GPUs , and corroborate the right format , it could make a monolithic difference once you ’re able to throw a Blackwell graphics card in your gage personal computer .
All eyes are on the end of the twelvemonth , or potentially early next class , for Nvidia ’ Blackwell punt GPUs to show up . For now , enterprisingness client have accession to the B200 through the HGX B200 — a waiter board that combine eight B200 GPUs — and the GB200 , which brings together the B200 GPU with Nvidia ’s Grace CPU . Nvidia is offering that part through the GB200 NVL72 , which lend together a dizzying 72 B200 GPUs and 36 Grace CPUs .
