Built primarily to accelerate deep learning, artificial neural networks, and HPC/GPGPU applications, Instinct GPUs were originally launched in December 2016 based on the GCN third- and fourth- generation microarchitectures.
The Instinct MI50 belongs to the AMD Vega 20 GCN 5.1 architecture. It has a 16 GB HBM2 memory, a 1,746 MHz boost clock speed, 3,840 stream processors, 60 compute units, and 1 TB/s memory bandwidth. It offers full FP64 performance at 6.7 TFLOPS (1:2).
AMD's Instinct GPUs are direct competitors to NVIDIA's Tesla GPUs and are also made with the purpose of accelerating computational tasks with unrestricted computing power.
Like NVIDIAs Tesla GPUs, the Instinct series can also immensely contribute to computations where floating-point calculations require double precision for greater accuracy. Apart from pure computations, high-end image generation for applications in professional and scientific fields is also possible with ROCm.
In general, Radeon Pro and Instinct GPUs have higher double-precision power compared to the consumer-level Radeon GPUs. Instinct GPUs cost around $10,000, while the Pro series are also available for mid-to-high range budgets, comparable to the Radeon series.
Reviewing all three GPU prices to performance ratios, a notable advantage for gamers who also happen to be equally passionate developers is the price point at which the new Radeon VII will be available at the time of writing this chapter. A margin of $699 USD really makes it easier for them to get involved with GPU computing, considering its 1:8 FP64 metric at 1.6 TFLOPS!
Let's compare the specifications of the three GPU models, which also includes all three FP64 metrics:
- Memory and features comparison:
|
Radeon VII |
Radeon Vega Frontier Edition |
Radeon Instinct MI50 |
|
|
Microarchitecture |
Vega second generation |
Vega second Generation |
Vega second generation |
|
Memory size |
16 GB |
16 GB |
16 GB |
|
Memory type |
HBM2 |
HBM2 |
HBM2 |
|
Memory bus |
4,096-bit |
2,048-bit |
4,096-bit |
|
Bandwidth |
1 TB/s |
484 GB/s |
1 TB/s |
|
xGMI connect support |
No |
No |
No |
|
DirectGMA support |
No |
Yes |
No |
- Performance comparison:
|
Radeon VII |
Radeon Vega Frontier Edition |
Radeon Instinct MI50 |
|
|
Pixel rate |
112.0 GPixel/s |
102.4 GPixel/s |
111.7 GPixel/s |
|
Texture rate |
420.0 GTexel/s |
409.6 GTexel/s |
419.0 GTexel/s |
|
FP16 (half) performance |
27.7 TFLOPS |
26.2 TFLOPS |
26.8 TFLOPS |
|
FP32 (float) performance |
13.8 TFLOPS |
13.1 TFLOPS |
13.4 TFLOPS |
|
FP64 (double) performance |
3.46 TFLOPS |
0.8 TFLOPS |
6.7 TFLOPS |
- Clock speeds:
|
Radeon VII |
Radeon Vega Frontier Edition |
Radeon Instinct MI50 |
|
|
GPU clock |
1,400 MHz |
1,382 MHz |
1,200 MHz |
|
Boost clock |
1,750 MHz |
1,600 MHz |
1,746 MHz |
|
Memory clock |
1,000 MHz |
945 MHz |
1,000 MHz |
|
Effective memory clock |
2,000 MHz |
1,890 MHz |
2,000 MHz |
For a deep learning (DL) performance comparison of GPUs without tensor cores, they can be equivalently compared to their conventional FLOPS to benchmark maximum DL performance (particularly FP16/FP32 mixed precision).
We can see that apart from the 1:4 FP64 metric, the Radeon VII is just a revised version of the Instinct MI50 card! The Vega Frontier edition has a much lower FP64 value!
If we revisit the affordability level that we spoke of in the Computing on NVIDIA GPUs section for NVIDIA, AMD GPU prices with the mentioned specifications can be considered a win-win situation both for organizations (academia/industry) and individuals who are independent researchers.
Just like NVIDIA has GPUDirect and NVLink, AMD has DirectGMA and xGMI Interconnect (upcoming). GMA stands for Graphic Memory Access and xGMI stands for inter-chip Global Memory Interconnect. DirectGMA is open source via AMD's GPUOpen initiative. GCN refers to AMD's Graphic Core Next architecture.