Radeon Instinct MI200 Accelerator Gets 128GB HBM2e
AMD's professional Radeon Instinct MI200 for the data center accelerator won't be shy of memory as it appears to be receiving 128GB of HMB2e. That is four times as much as the prevailing MI100. The accelerator is also the first GPU from AMD that is made up of multiple chiplets, much like Ryzen processors.
The Supercomputing Center Australia Pawsey works on a new supercomputer known as Setonix which will use 'MI-next' GPUs with 128GB, says HPC Wire. But further information about AMD has yet to be released, as AMD has not revealed its Instinct MI200 Card.
Last week, Twitter user Locuza shared an AMD-based CDNA2 schematic for the GPU. This would be a multichip design or an MCM. As with the Ryzen processors, several chips are coupled with chips holding core components. A total of eight stacks of 16GB HBM2e would be used, and the MI200 would be based on two chiplets, as described in the diagram. It is stated that the accelerator contains 128 compute units that may hold 16,384 stream processors, although it is not yet known if they are all operational.
The MI200 is the first CDNA2-based accelerator for 2022. The GPU is named Aldebaran. Accelerators from AMD's Instinct are utilized in data centers and supercomputers.
AMD Instinct Accelerators | ||
---|---|---|
Accelerator Name | AMD Instinct MI100 | AMD Instinct MI200* |
Architecture | 7nm CDNA1 (GFX908) | cDNA2 (GFX90A) |
GPU | Arcturus | Aldebaran (Multi-Chip-Module) |
GPU cores | 7680 | - |
GPU speed | ~1500MHz | Nnb |
FP16 Compute | 185TFLOPs | - |
FP32 Compute | 23.1TFLOPs | - |
FP64 Compute | 11.5TFLOPs | - |
Vram | 32GB HBM2 | 128GB HBM2E |
Memory Speed | 1200MHz | - |
Memory bus | 4096 bit bus | - |
Memory bandwidth | 1.23TB/s | - |
form factor | Dual Slot, Full Length | - |
Cooling | passive | - |
TPD | 300W | - |
*Specifications not yet officially confirmed.
AMD Donates $15 million worth EPYC CPUs and AMD Radeon Instinct GPUs to research COVID-19 - 04/17/2020 06:03 PM
AMD is announcing today a COVID-19 HPC fund to provide research institutions with computing resources to accelerate medical research on COVID-19 and other diseases. The fund will include an initial do...
AMD brings Radeon Image Sharpening (RIS) visuals to select Radeon RX 400/500 products - 09/16/2019 08:49 AM
With last weeks driver AMD brought Radeon Image Sharpening (RIS) feature to the Radeon RX 570, RX 580 and RX 590, as well as RX 470, and RX 480 graphics cards....
AMD EPYC CPUs, AMD Radeon Instinct GPUs to power Cray Supercomputer - 05/07/2019 01:47 PM
AMD and Cray unveil what is expected to be the world’s fastest supercomputer – defining a new standard for high-performance computing and pushing the technology boundaries of compu...
AMD announces Radeon Instinct MI60 and Instinct MI50 at 7nm - 11/06/2018 08:40 PM
AMD) today announced the AMD Radeon Instinct™ MI60 and MI50 accelerators, the world’s first 7nm datacenter GPUs, designed to deliver the compute performance required for next-gener...
AMD Next Horizon Event - Epyc 2 64 Cores - 7nm "ZEN 2" and Radeon Instinct MI60 - 11/06/2018 07:37 PM
AMD is running its 'Next Horizon' event with announcements and more details about the company's pending 7nm products are shared. The event is intended for the server and data-center centric announ...
Senior Member
Posts: 13801
Joined: 2004-05-16
https://www.nvidia.com/en-us/data-center/a100/
A100 80GB - but presumably they'll be launching a new architecture next year as well.
Senior Member
Posts: 2505
Joined: 2017-08-18
the future is now.
and the latency issue is dead
Senior Member
Posts: 1307
Joined: 2011-01-11
Like WHOA.....!!
Just for the heck of it I went ahead and checked my memory bandwidth of my 3090 ATM. And it is at 1.07TB/s with GPU-Z saying its bandwidth is 1066GB/s. And that is with a +1350 to the memory just to get it there!
These new GPU "Chiplet/Module" whatever they consist of inside are crazy!!
Just wonder what is going to trickle down to us paltry consumers...? And at what cost to boot when new things do come??
Senior Member
Posts: 2505
Joined: 2017-08-18
Like WHOA.....!!
Just for the heck of it I went ahead and checked my memory bandwidth of my 3090 ATM. And it is at 1.07TB/s with GPU-Z saying its bandwidth is 1066GB/s. And that is with a +1350 to the memory just to get it there!
These new GPU "Chiplet/Module" whatever they consist of inside are crazy!!
Just wonder what is going to trickle down to us paltry consumers...? And at what cost to boot when new things do come??
question 1 is fairly easy, RDNA 3.
question two remains uncertain. what i'm pretty damn sure of is there will be a rtx xx90 destroying sku, maybe two. they'll offer the same mcm design just adjusting for line-up like ryzen 7/9 +/- VRAM and cores.
cost wise there will be economies of scale twice over; first for the mcm taking up less silicon (= more chips per wafer), second for having essentially the same "gpu" across most of the product stack.
practically there is another benefit in a more rapid response to market conditions...they will not have to worry if one sku is under/over performing while another is in high demand.
best of all for consumers there will be more availability as there would be many times the production of mcm's vs. discrete gpu per wafer (up to 4-5x) and the "infinity fabric" has continued to drop in cost of production.
Senior Member
Posts: 6658
Joined: 2012-11-10
Having not kept up with server GPUs, what's the closest competitor from Nvidia?