AMD has officially announced its next-generation MI200 HPC GPU codenamed Aldebaran that uses a 6nm CDNA 2 architecture to deliver insane compute performance.
AMD Unveils Instinct MI200, Powering The Next-Gen Compute Powerhouse With First 6nm MCM GPU Technology & Over 95 TFLOPs FP32 Performance
AMD is officially the first to MCM technology and they are doing so with a grand product which is their Instinct MI200 codenamed Aldebaran. The AMD Aldebaran GPU will come in various forms & sizes but it’s all based on the brand new CDNA 2 architecture which is the most refined variation of Vega. Some of the main features before we go into detail are listed below:
- AMD CDNA 2 architecture – 2nd Gen Matrix Cores accelerating FP64 and FP32 matrix operations, delivering up to 4X the peak theoretical FP64 performance vs. AMD previous-gen GPUs.
- Leadership Packaging Technology – Industry-first multi-die GPU design with 2.5D Elevated Fanout Bridge (EFB) technology delivers 1.8X more cores and 2.7X higher memory bandwidth vs. AMD previous-gen GPUs, offering the industry’s best aggregate peak theoretical memory bandwidth at 3.2 terabytes per second.
- 3rd Gen AMD Infinity Fabric technology – Up to 8 Infinity Fabric links connect the AMD Instinct MI200 with 3rd Gen EPYC CPUs and other GPUs in the node to enable unified CPU/GPU memory coherency and maximize system throughput, allowing for an easier on-ramp for CPU codes to tap the power of accelerators.
AMD Instinct MI200 GPU Die Shot:
Inside the AMD Instinct MI200 is an Aldebaran GPU featuring two dies, a secondary and a primary. It has two dies with each consisting of 8 shader engines for a total of 16 SE’s. Each Shader Engine packs 16 CUs with full-rate FP64, packed FP32 & a 2nd Generation Matrix Engine for FP16 & BF16 operations.
Each die, as such, is composed of 128 compute units or 8192 stream processors. This rounds up to a total of 220 compute units or 14,080 stream processors for the entire chip. The Aldebaran GPU is also powered by a new XGMI interconnect. Each chiplet features a VCN 2.6 engine and the main IO controller.
Built on AMD CDNA 2 architecture, AMD Instinct MI200 series accelerators deliver leading application performance for a broad set of HPC workloads. The AMD Instinct MI250X accelerator provides up to 4.9X better performance than competitive accelerators for double precision (FP64) HPC applications and surpasses 380 teraflops of peak theoretical half-precision (FP16) for AI workloads to enable disruptive approaches in further accelerating data-driven research.
In terms of performance, AMD is touting various record wins in the HPC segment over NVIDIA’s A100 solution with up to 3x performance improvements in AMG.
As for DRAM, AMD has gone with an 8-channel interface consisting of 1024-bit interfaces for an 8192-bit wide bus interface. Each interface can support 2GB HBM2e DRAM modules. This should give us up to 16 GB of HBM2e memory capacity per stack and since there are eight stacks in total, the total amount of capacity would be a whopping 128 GB. That’s 48 GB more than the A100 which houses 80 GB HBM2e memory. The memory will clock in at an insane speed of 3.2 Gbps for a full-on bandwidth of 3.2 TB/s. This is a whole 1.2 TB/s more bandwidth than the A100 80 GB which has 2 TB/s.
The AMD Instinct MI200 will be powering three top-tier supercomputers which include the United States’ exascale Frontier system; the European Union’s pre-exascale LUMI system; and Australia’s petascale Setonix system. The competition includes the A100 80 GB which offers 19.5 TFLOPs of FP64, 156 TFLOPs of FP32 and 312 TFLOPs of FP16 compute power. But we are likely to hear about NVIDIA’s own Hopper MCM GPU next year so there’s going to be a heated competition between the two GPU juggernauts in 2022.
AMD Radeon Instinct Accelerators 2020
| Accelerator Name | AMD Instinct MI300 | AMD Instinct MI250X | AMD Instinct MI250 | AMD Instinct MI100 | AMD Radeon Instinct MI60 | AMD Radeon Instinct MI50 | AMD Radeon Instinct MI25 | AMD Radeon Instinct MI8 | AMD Radeon Instinct MI6 |
|---|---|---|---|---|---|---|---|---|---|
| GPU Architecture | TBA (CDNA 3) | Aldebaran (CDNA 2) | Aldebaran (CDNA 2) | Arcturus (CDNA 1) | Vega 20 | Vega 20 | Vega 10 | Fiji XT | Polaris 10 |
| GPU Process Node | Advanced Process Node | Advanced Process Node | Advanced Process Node | 7nm FinFET | 7nm FinFET | 7nm FinFET | 14nm FinFET | 28nm | 14nm FinFET |
| GPU Dies | 4 (MCM)? | 2 (MCM) | 2 (MCM) | 1 (Monolithic) | 1 (Monolithic) | 1 (Monolithic) | 1 (Monolithic) | 1 (Monolithic) | 1 (Monolithic) |
| GPU Cores | 28,160? | 14,080 | 14,080? | 7680 | 4096 | 3840 | 4096 | 4096 | 2304 |
| GPU Clock Speed | TBA | 1700 MHz | ~1700 MHz | ~1500 MHz | 1800 MHz | 1725 MHz | 1500 MHz | 1000 MHz | 1237 MHz |
| FP16 Compute | TBA | 383 TOPs | TBA | 185 TFLOPs | 29.5 TFLOPs | 26.5 TFLOPs | 24.6 TFLOPs | 8.2 TFLOPs | 5.7 TFLOPs |
| FP32 Compute | TBA | 95.8 TFLOPs | TBA | 23.1 TFLOPs | 14.7 TFLOPs | 13.3 TFLOPs | 12.3 TFLOPs | 8.2 TFLOPs | 5.7 TFLOPs |
| FP64 Compute | TBA | 47.9 TFLOPs | TBA | 11.5 TFLOPs | 7.4 TFLOPs | 6.6 TFLOPs | 768 GFLOPs | 512 GFLOPs | 384 GFLOPs |
| VRAM | TBA | 128 GB HBM2e | 128 GB HBM2e | 32 GB HBM2 | 32 GB HBM2 | 16 GB HBM2 | 16 GB HBM2 | 4 GB HBM1 | 16 GB GDDR5 |
| Memory Clock | TBA | TBA | TBA | 1200 MHz | 1000 MHz | 1000 MHz | 945 MHz | 500 MHz | 1750 MHz |
| Memory Bus | TBA | 8192-bit | 8192-bit | 4096-bit bus | 4096-bit bus | 4096-bit bus | 2048-bit bus | 4096-bit bus | 256-bit bus |
| Memory Bandwidth | TBA | ~2 TB/s? | ~2 TB/s? | 1.23 TB/s | 1 TB/s | 1 TB/s | 484 GB/s | 512 GB/s | 224 GB/s |
| Form Factor | TBA | Dual Slot, Full Length / OAM | Dual Slot, Full Length / OAM | Dual Slot, Full Length | Dual Slot, Full Length | Dual Slot, Full Length | Dual Slot, Full Length | Dual Slot, Half Length | Single Slot, Full Length |
| Cooling | TBA | Passive Cooling | Passive Cooling | Passive Cooling | Passive Cooling | Passive Cooling | Passive Cooling | Passive Cooling | Passive Cooling |
| TDP | TBA | 500W | TBA | 300W | 300W | 300W | 300W | 175W | 150W |
The Aldebaran MI200 GPU will come in three configurations, the OAM only MI250 and MI250X & the dual-slot PCIe MI210. AMD has only shared full specifications and performance figures for its MI250 class HPC GPUs. The MI250X features the full 14,080 configurations and delivers 47.9, 95.7, 383 TFLOPs of FP64/FP32/FP16 while the MI250 features 13,312 cores with 45.3,90.5,362.1 TFLOPs of FP64/FP32/FP16 performance. The memory configuration remains the same between the two GPU configurations.
AMD Instinct MI200 GPU Package:
