AMD has announced the official launch of its flagship AI GPU accelerator, the MI300X, which offers up to 60% better performance than NVIDIA’s H100.
AMD Finally Has The GPU To Tackle NVIDIA In The AI Segment, MI300X Up To 60% Faster Than H100
The AMD Instinct MI300 class of AI accelerators will be another chiplet powerhouse, making use of advanced packaging technologies from TSMC. Today, AMD not only announced the launch of these chips but shared the first performance benchmarks of the MI300X which look great. AMD first used the general specs as a comparison and their CDNA 3 accelerator offers (versus NVIDIA H100):
- 2.4X Higher Memory Capacity
- 1.6X Higher Memory Bandwidth
- 1.3X FP8 TFLOPS
- 1.3X FP16 TFLOPS
- Up To 20% Faster Vs H100 (Llama 2 70B) In 1v1 Comparison
- Up To 20% Faster Vs H100 (FlashAttention 2) in 1v1 Comparison
- Up To 40% Faster Vs H100 (Llama 2 70B) in 8v8 Server
- Up To 60% Faster Vs H100 (Bloom 176B) In 8v8 Server
In general LLM Kernel TFLOPs, the MI300X offers up to 20% higher performance in FlashAttention-2 and Llama 2 70B. Looking from a platform perspective which compares an 8x MI300X solution to an 8X H100 solution, we see a much bigger 40% gain in Llama 2 70B & a 60% gain in Bloom 176B.
AMD mentions that in training performance, the MI300X is on par with the competition (H100) and offers competitive price/perf while shining in inferencing workloads.
The driving force behind the latest MI300 accelerators is ROCm 6.0. The software stack has been updated to the latest version with powerful new features which include support for various AI workloads such as Generative AI and Large language models.
The new software stack supports the latest compute formats such as FP16, Bf16, and FP8 (including Sparsity). The optimizations combine to offer up to 2.6x speedup in vLLM through optimized inference libraries, 1.4x speedup in HIP Graph through optimized runtime, and 1.3x Flash Attention speedup through optimized Kernels. ROCm 6 is expected later this month alongside the MI300 AI accelerators. It will be interesting to see how ROCm 6 compares to the latest version of NVIDIA’s CUDA stack which is its real competition.
AMD Instinct MI300X – Challenging NVIDIA’s AI Supremacy With CDNA 3 & Huge Memory
The AMD Instinct MI300X is the chip that will be highlighted the most since it is targeted at NVIDIA’s Hopper and Intel’s Gaudi accelerators within the AI segment. This chip has been designed solely on the CDNA 3 architecture and there is a lot of stuff going on. The chip is going to host a mix of 5nm and 6nm IPs, all combining to deliver up to 153 Billion transistors (MI300X).
Starting with the design, the main interposer is laid out with a passive die which houses the interconnect layer using a 4th Gen Infinity Fabric solution. The interposer includes a total of 28 dies which include eight HBM3 packages, 16 dummy dies between the HBM packages, & four active dies and each of these active dies gets two compute dies.
Each GCD based on the CDNA 3 GPU architecture features a total of 40 compute units which equals 2560 cores. There are eight compute dies (GCDs) in total so that gives us a total of 320 Compute & 20,480 core units. For yields, AMD will be scaling back a small portion of these cores and we will be seeing a total of 304 Compute units (38 CUs per GPU chiplet) enabled for a total of 19,456 stream processors.
Memory is another area where you will see a huge upgrade with the MI300X boasting 50% more HBM3 capacity than its predecessor, the MI250X (128 GB). To achieve a memory pool of 192 GB, AMD is equipping the MI300X with 8 HBM3 stacks and each stack is 12-Hi while incorporating 16 Gb ICs which give us 2 GB capacity per IC or 24 GB per stack.
The memory will offer up to 5.3 TB/s of bandwidth and 896 GB/s of Infinity Fabric Bandwidth. For comparison, NVIDIA’s upcoming H200 AI accelerator offers 141 GB capacities while Gaudi 3 from Intel will be offering 144 GB capacities. Large memory pools matter a lot in LLMs which are mostly memory-bound and AMD can show its AI prowess by leading in the memory department. For comparisons:
- Instinct MI300X – 192 GB HBM3
- Gaudi 3 – 144 GB HBM3
- H200 – 141 GB HBM3e
- MI300A – 128 GB HBM3
- MI250X – 128 GB HBM2e
- H100 – 96 GB HBM3
- Gaudi 2 – 96 GB HBM2e
In terms of power consumption, the AMD Instinct MI300X is rated at 750W which is a 50% increase over the 500W of the Instinct MI250X and 50W more than the NVIDIA H200.
One configuration showcased is the G593-ZX1/ZX2 series of servers from Gigabyte which offer up to 8 MI300X GPU accelerators and two AMD EPYC 9004 CPUs. These systems will be equipped with up to eight 3000W power supplies, totaling 18000W of power. AMD also showcased its own Instinct MI300X platform which includes 8 of these AI accelerator chips, offering some solid numbers over the NVIDIA HGX H100 platform. Some numbers shared by AMD include:
- 2.4X Higher HBM3 Memory (1.5 TB vs 640 GB)
- 1.3X More Compute FLOPS (10.4 PF vs 7.9 PF)
- Similar Bi-Directional Bandwidth (896 GB/s vs 900 GB/s)
- Similar Single-Node Ring Bandwidth (448 GB/s vs 450 GB/s)
- Similar Networking Capabilities (400 GbE vs 400 GbE)
- Similar PCIe Protocol (PCIe Gen 5 128 GB/s)
For now, AMD should know that their competitors are also going full steam ahead on the AI craze with NVIDIA already teasing some huge figures for its 2024 Hopper H200 GPUs & Blackwell B100 GPUs and Intel prepping up its Guadi 3 and Falcon Shores GPUs for launch in the coming years too. Companies such as Oracle, Dell, META, and OpenAI have announced support for AMD’s Instinct MI300 AI chips in their ecosystem.
One thing is for sure at the current moment, AI customers will gobble up almost anything they can get and everyone is going to take advantage of that. But AMD has a very formidable solution that is not just aiming to be an alternative to NVIDIA but a leader in the AI segment.
AMD Radeon Instinct Accelerators
| Accelerator Name | AMD Instinct MI400 | AMD Instinct MI300X | AMD Instinct MI300A | AMD Instinct MI250X | AMD Instinct MI250 | AMD Instinct MI210 | AMD Instinct MI100 | AMD Radeon Instinct MI60 | AMD Radeon Instinct MI50 | AMD Radeon Instinct MI25 | AMD Radeon Instinct MI8 | AMD Radeon Instinct MI6 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| CPU Architecture | Zen 5 (Exascale APU) | N/A | Zen 4 (Exascale APU) | N/A | N/A | N/A | N/A | N/A | N/A | N/A | N/A | N/A |
| GPU Architecture | CDNA 4 | Aqua Vanjaram (CDNA 3) | Aqua Vanjaram (CDNA 3) | Aldebaran (CDNA 2) | Aldebaran (CDNA 2) | Aldebaran (CDNA 2) | Arcturus (CDNA 1) | Vega 20 | Vega 20 | Vega 10 | Fiji XT | Polaris 10 |
| GPU Process Node | 4nm | 5nm+6nm | 5nm+6nm | 6nm | 6nm | 6nm | 7nm FinFET | 7nm FinFET | 7nm FinFET | 14nm FinFET | 28nm | 14nm FinFET |
| GPU Chiplets | TBD | 8 (MCM) | 8 (MCM) | 2 (MCM) 1 (Per Die) |
2 (MCM) 1 (Per Die) |
2 (MCM) 1 (Per Die) |
1 (Monolithic) | 1 (Monolithic) | 1 (Monolithic) | 1 (Monolithic) | 1 (Monolithic) | 1 (Monolithic) |
| GPU Cores | TBD | 19,456 | 14,592 | 14,080 | 13,312 | 6656 | 7680 | 4096 | 3840 | 4096 | 4096 | 2304 |
| GPU Clock Speed | TBD | 2100 MHz | 2100 MHz | 1700 MHz | 1700 MHz | 1700 MHz | 1500 MHz | 1800 MHz | 1725 MHz | 1500 MHz | 1000 MHz | 1237 MHz |
| INT8 Compute | TBD | 2614 TOPS | 1961 TOPS | 383 TOPs | 362 TOPS | 181 TOPS | 92.3 TOPS | N/A | N/A | N/A | N/A | N/A |
| FP16 Compute | TBD | 1.3 PFLOPs | 980.6 TFLOPs | 383 TFLOPs | 362 TFLOPs | 181 TFLOPs | 185 TFLOPs | 29.5 TFLOPs | 26.5 TFLOPs | 24.6 TFLOPs | 8.2 TFLOPs | 5.7 TFLOPs |
| FP32 Compute | TBD | 163.4 TFLOPs | 122.6 TFLOPs | 95.7 TFLOPs | 90.5 TFLOPs | 45.3 TFLOPs | 23.1 TFLOPs | 14.7 TFLOPs | 13.3 TFLOPs | 12.3 TFLOPs | 8.2 TFLOPs | 5.7 TFLOPs |
| FP64 Compute | TBD | 81.7 TFLOPs | 61.3 TFLOPs | 47.9 TFLOPs | 45.3 TFLOPs | 22.6 TFLOPs | 11.5 TFLOPs | 7.4 TFLOPs | 6.6 TFLOPs | 768 GFLOPs | 512 GFLOPs | 384 GFLOPs |
| VRAM | TBD | 192 GB HBM3 | 128 GB HBM3 | 128 GB HBM2e | 128 GB HBM2e | 64 GB HBM2e | 32 GB HBM2 | 32 GB HBM2 | 16 GB HBM2 | 16 GB HBM2 | 4 GB HBM1 | 16 GB GDDR5 |
| Infinity Cache | TBD | 256 MB | 256 MB | N/A | N/A | N/A | N/A | N/A | N/A | N/A | N/A | N/A |
| Memory Clock | TBD | 5.2 Gbps | 5.2 Gbps | 3.2 Gbps | 3.2 Gbps | 3.2 Gbps | 1200 MHz | 1000 MHz | 1000 MHz | 945 MHz | 500 MHz | 1750 MHz |
| Memory Bus | TBD | 8192-bit | 8192-bit | 8192-bit | 8192-bit | 4096-bit | 4096-bit bus | 4096-bit bus | 4096-bit bus | 2048-bit bus | 4096-bit bus | 256-bit bus |
| Memory Bandwidth | TBD | 5.3 TB/s | 5.3 TB/s | 3.2 TB/s | 3.2 TB/s | 1.6 TB/s | 1.23 TB/s | 1 TB/s | 1 TB/s | 484 GB/s | 512 GB/s | 224 GB/s |
| Form Factor | TBD | OAM | APU SH5 Socket | OAM | OAM | Dual Slot Card | 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 | TBD | Passive Cooling | Passive Cooling | Passive Cooling | Passive Cooling | Passive Cooling | Passive Cooling | Passive Cooling | Passive Cooling | Passive Cooling | Passive Cooling | Passive Cooling |
| TDP (Max) | TBD | 750W | 760W | 560W | 500W | 300W | 300W | 300W | 300W | 300W | 175W | 150W |