The super-sized future is here with AMD’s Instinct MI300
Lisa Su waited until the tail end of her speech last night to pull a literal behemoth out of her bag of tricks with the introduction of the Instinct MI300, a new high performance processor that blends advanced CPU and GPU technology towards maximum benefit for AI and HPC workloads. It’s not surprising to see AMD deliver another performance best. They’ve been on quite a roll for years most recently with the addition of their performance leading 4th generation EPYC processors late last year. What caught my eye about the introduction of the MI300 was the elegance of its design and what it foretells for the entire semiconductor arena moving forward. Let’s break it down.
Chiplet architecture wins the day
The MI300 is, per AMD’s claim, the world’s first chiplet based processor combining CPU and GPU cores. This is important for high performance computing as a combination of CPU and GPU can best support the demands of HPC and AI workloads. AMD has pulled this feat off through 3D stacking of nine 5nm chiplets sitting atop four 6nm chiplets and surrounded by a breathtaking 128GB of HBM3 memory. What we are seeing is modern silicon innovation tapping different silicon architectures tightly coupled with high performance memory and I/O reducing both compute latency and platform overhead, and it’s a harbinger of where the entire silicon ecosystem is aiming for future products.
The performance claimed by AMD is eye-opening. 8X the performance and 5X the perf/efficiency of the MI250, the nifty processor that fuels the AMD based Exascale class supercomputer at Oakridge National Lab. As Dr. Su held the chip aloft you could see that as a microprocessor architect herself she well understood what she’s delivered to the market and the importance to the trajectory of silicon innovation, and with that initial chip reveal the team at AMD should be proud.
Super-sized silo designs
But some questions come to mind. When will we see others jump into this APU/XPU pool? Intel shared more details of its upcoming Falcon Shores product at SC’22 which will feature a combination of Intel Xeon CPU and Intel Data Center GPU Max Series and is hinted to hit the market in 2024, behind anticipated MI300 launch in 2H'23. NVIDIA is also hard at work innovating what must be one of the best named products in the industry with its Grace Hopper superchip, again combining ARM CPU and GPU into a single package met with a proprietary 900 GB per second coherent interface.
As you look at this landscape it doesn’t take too long to realize that while these super-sized value meal chips will disrupt the market, some customers will want more control of exactly what chiplets they use. Enter UCIe, the new industry standard that will provide a common interconnect for chiplet designs which could open the door for a mixed architecture implementation. This is the reason I called UCIe one of the biggest disruptions on the compute landscape in an earlier blog. My mind also goes to the importance of foundry services in an era of multi-vendor chiplet designs as well as the emerging RISC-V architecture open-sourcing logic in a way we’ve never seen. Watch this space for more updates on all of these topics in 2023. For now, we’ll all need to wait as the MI300 and other supersized processors are readied for market. As always, thanks for engaging - Allyson
Lisa Su waited until the tail end of her speech last night to pull a literal behemoth out of her bag of tricks with the introduction of the Instinct MI300, a new high performance processor that blends advanced CPU and GPU technology towards maximum benefit for AI and HPC workloads. It’s not surprising to see AMD deliver another performance best. They’ve been on quite a roll for years most recently with the addition of their performance leading 4th generation EPYC processors late last year. What caught my eye about the introduction of the MI300 was the elegance of its design and what it foretells for the entire semiconductor arena moving forward. Let’s break it down.
Chiplet architecture wins the day
The MI300 is, per AMD’s claim, the world’s first chiplet based processor combining CPU and GPU cores. This is important for high performance computing as a combination of CPU and GPU can best support the demands of HPC and AI workloads. AMD has pulled this feat off through 3D stacking of nine 5nm chiplets sitting atop four 6nm chiplets and surrounded by a breathtaking 128GB of HBM3 memory. What we are seeing is modern silicon innovation tapping different silicon architectures tightly coupled with high performance memory and I/O reducing both compute latency and platform overhead, and it’s a harbinger of where the entire silicon ecosystem is aiming for future products.
The performance claimed by AMD is eye-opening. 8X the performance and 5X the perf/efficiency of the MI250, the nifty processor that fuels the AMD based Exascale class supercomputer at Oakridge National Lab. As Dr. Su held the chip aloft you could see that as a microprocessor architect herself she well understood what she’s delivered to the market and the importance to the trajectory of silicon innovation, and with that initial chip reveal the team at AMD should be proud.
Super-sized silo designs
But some questions come to mind. When will we see others jump into this APU/XPU pool? Intel shared more details of its upcoming Falcon Shores product at SC’22 which will feature a combination of Intel Xeon CPU and Intel Data Center GPU Max Series and is hinted to hit the market in 2024, behind anticipated MI300 launch in 2H'23. NVIDIA is also hard at work innovating what must be one of the best named products in the industry with its Grace Hopper superchip, again combining ARM CPU and GPU into a single package met with a proprietary 900 GB per second coherent interface.
As you look at this landscape it doesn’t take too long to realize that while these super-sized value meal chips will disrupt the market, some customers will want more control of exactly what chiplets they use. Enter UCIe, the new industry standard that will provide a common interconnect for chiplet designs which could open the door for a mixed architecture implementation. This is the reason I called UCIe one of the biggest disruptions on the compute landscape in an earlier blog. My mind also goes to the importance of foundry services in an era of multi-vendor chiplet designs as well as the emerging RISC-V architecture open-sourcing logic in a way we’ve never seen. Watch this space for more updates on all of these topics in 2023. For now, we’ll all need to wait as the MI300 and other supersized processors are readied for market. As always, thanks for engaging - Allyson
