11.7 C
New York
Monday, March 4, 2024

Nvidia Blackwell rumors: RTX 5090 to supply 1.7x efficiency, multi-chiplet design for compute GPUs

Rumor mill: With the RTX 4000 collection now just about behind us, leaks and rumors are rising about Nvidia’s subsequent technology of shopper graphics playing cards, the RTX 5000 line. The newest of those affords some efficiency indicators, together with the RTX 5090 displaying a 1.7x general uplift in comparison with its predecessor. One other leak factors to Nvidia lastly utilizing a multi-chiplet design within the firm’s high-performance compute GPUs.

Beginning with the buyer collection, a leaker on the Chiphell discussion board, Panzerlied, has posted what are claimed to be stats for the RTX 5090: a 50% improve in scale (which presumably refers to cores), a 52% improve in reminiscence bandwidth, 78% improve in L2 cache, 15% improve in frequency, and 1.7x efficiency uplift.

Making use of these figures to the RTX 4090 would counsel that the successor will pack round 24,000 CUDA cores, a 2.9 GHz enhance clock, and 128MB of L2 cache.

It is also recommended that the RTX 5090’s reminiscence will use GDDR7 and be boosted to 32 Gbps. The AD102 GPU successor is rumored to incorporate a 512-bit reminiscence bus, although it won’t be used within the RTX 5090. As VideoCardz notes, the cardboard might include configurations comparable to 512-bit/24 Gbps or 448-bit/28 Gbps.

Whereas everybody has their very own idea in regards to the subsequent technology of Nvidia playing cards, it is price mentioning that Panzerlied has made appropriate claims prior to now. Furthermore, these most up-to-date rumors had been “confirmed” by prolific {hardware} leaker Kopite7kimi.

If the figures are true or near the reality, one has to marvel what kind of price ticket Nvidia will slap on the RTX 5090. Staff Inexperienced was closely criticized over its Lovelace pricing, nevertheless it’s onerous to think about this RTX 5090 being cheaper than, and even the identical value as, the $1,600 RTX 4090.

Earlier RTX 5000-series rumors additionally pointed to important efficiency will increase in comparison with Lovelace. No phrase but on a launch date, although many say they’ll land subsequent 12 months.

In a associated story, Kopite7kimi additionally made some claims about Nvidia’s next-gen merchandise. He says that the Blackwell structure will probably be used throughout each shopper and datacenter GPUs, versus the present Ada Lovelace/Hopper break up. Furthermore, Nvidia will apparently be following Intel and AMD in utilizing a multi-chiplet design for the primary time in its datacenter class of GPU.

“After the dramas of GA100 and GH100, it appears that evidently GB100 is lastly going to make use of MCM,” kopite7kimi wrote. “Perhaps GB100=2*GB102.”

Even when Nvidia does go down the multi-chiplet design route for its compute GPUs, the corporate continues to be anticipated to stay with a monolithic design for shopper merchandise.

Related Articles


Please enter your comment!
Please enter your name here

Latest Articles