Angstronomics draws the line on RDNA 3!
We detail key specifications of Navi31, Navi32 and Navi33 that were finalized in 2020 and have not changed since!
We also highlight some architectural changes including OREO😋
WGP, Cache and die sizes inside!
https://t.co/n2qB7KiiBg
Apple WWDC 2023
Mac Pro: M2 Ultra Mac Studio with a big switch to enable 6x internal PCIe4 slots for $3000 more
Apple Vision Pro: M2 ON YOUR FACE! (and R1 sensor co-processor) for $3500 in 2024
2 hour battery life with a battery dongle
The long-delayed Intel-powered Aurora supercomputer fails to make the June 2023 TOP500 list!
"Promising early performance results on the Aurora hardware will be shown" at #ISC23 later today.
61st #Top500 revealed at #ISC23#Frontier lives to be dethroned another day!
No change in rankings of the Top 10 #supercomputers, although Frontier, LUMI & Leonardo improved their HPL score
https://t.co/NvaPmAfbJl
#HPC#AI @ISChpc @top500supercomp
Looks like @ServeTheHome is about to check out the 25.6T Broadcom Tomahawk 4 switch! Here's a clearer image of the LGA 8371 socket of the BCM56990 found on ebay:
Bryan just took this photo of me. This is probably the biggest chip I have ever installed with over 8000 pads. If you are thinking it is an Ampere Altra Max package, it is much larger.
NVIDIA's Orin is bigger than M1 Max!
TechanaLye tears apart an AGX Orin and finds a large ~455 mm² die with 17B transistors on Samsung 8nm.
M1 Max on TSMC5 is smaller but packs 3.3x the transistors.
M2 packs more transistors in 1/3rd the area!
Our die annotation is here!
And of course RTX 4000 SFF has 4x mDP1.4 vs 0 on L4.
Doing a rough estimate using RTX 4080 Laptop efficiency curves would put the RTX 4000 SFF at around 10k in TimeSpy Graphics.
A similarly priced RTX 4080 desktop card pulling 320W is almost 3x faster.
#GTC23 RTX 4000 SFF vs NVIDIA L4
48SM vs 58SM AD104
20GB 320GB/s vs 24GB 300GB/s
1.56GHz vs 2.04GHz
19.2TF vs 30.3TF
70W vs 72W
Dual slot blower vs Single slot passive (system airflow)
$1250 vs $????
#GTC23 And here is Grace Hopper!
Doing a quick compare vs GH100's 814 mm², it looks like Grace on TSMC 4N is about 730 mm²!
Pretty large for a 72-core CPU.
More details on Grace Superchip efficiency.
"NVIDIA Grace is sampling now with production in the second half of the year." - Mar 2023
"Grace CPU Superchip and Grace Hopper Superchip are expected to be available in the first half of 2023." - Aug 2022
https://t.co/Cnu36ujWrn
#GTC23 DGX Cloud
Each instance of DGX Cloud features eight NVIDIA H100 or A100 80GB Tensor Core GPUs for a total of 640GB of GPU memory per node.
- Scalable to multiple nodes
DGX Cloud instances start at $36,999 per instance per month