AMD Radeon Instinct MI100 is an upcoming GPU from AMD based on its new CDNA technology. In the latest news, we have a leak showing off real-world performance benchmarks of the card. These numbers leaked by AdoredTV show that this High-Performance Compute graphics card will give NVIDIA’s ampere series a good fight.
The leak stacks MI100 against NVIDIA’S Volta and Ampere architecture cards. We see that the AMD Radeon Instinct MI100 performs faster in benchmarks than NVIDIA’s recently released A100 In FP32 Compute With Better Perf/Value. The A100 is the first card from Nvidia based on the new Ampere technology.
We already know that Radeon Instinct MI100 GPU is supposed to come into the consumer market in the second half of this year. According to AdoredTV, the leaked slides are from an official presentation for AMD Radeon Instinct MI100. This could be a release presentation that they intended to show to people, or it can be an internal one.
The pictures posted by the leaker look modified, but they seem believable and give out some interesting new details. Along with the technical data, these slides also mention AMD’s launch plans and some server configurations that ADM is working on with their partners.
Instinct MI100 based 1U Server:
AMD plans to launch its 1U server machine built on a 2P design. It will be equipped with two EPYC CPUs of either Rome or Milan architecture. In addition, we have four Radeon Instinct MI100 cards, two for each CPU, connected via AMD’s second-generation infinity fabric interconnect. Together the four GPUs will deliver a constant 136 TFLOPs of FP32 output. This gives a rough idea of a single card performance of around 34 TFLOPs of FP32 compute.
The newer Radeon Instinct MI100 GPUs are power-hungry machines and will have a TDP of 300W. They Require two 8-Pin PCIe power connectors to run. Other details mention the bandwidth of the GPU. The total bandwidth of the new GPU is about 256 GB/s with the newer PCIe Gen-4 protocol. We also have aa total memory bandwidth of 4.9 TB/s, which hints that the machine uses AMD’s HBM2e DRAM dies. That brings us to single GPU memory bandwidth of about 1.225 TB/s.
Each GPU has 32 GB of memory, which gives the entire machine 128 GB. AMD is using 4 HBM2 DRAM stack technology with each stack housing 8-hi DRAM dies. Hence the XGMI technology will not be available to standard configurations of the server. We will only see it in specialized 1U racks.
AMD plans to launch the 1U server with AMD EPYC (Rome / Milan) CPUs sometime around December 2020. Team blue users will have to wait as the Intel Xeon variant is also expected to launch in February 2021.
Instinct MI100 based 3U Server:
The second product in AMD’s server lineup 3U is expected to arrive by March 2021. This is the more advanced model as compared to the previous 1U system. This machine will have 8 Radeon Instinct MI100 GPUs linked to two EPYC CPUs. The four cards per CPU will be connected with an XGMI (100 GB/s bi-directional) and a quad bandwidth of 1.2 TB/s. In terms of performance, the four Instinct compute units would equal a total of 272 TFLOPs of FP32 compute.
The PCIe bandwidth we have is 512 GB per second combined with 9.8 TB/s of HBM2 memory bandwidth and a total of 256 GB of memory DRAM capacity. With this much horsepower, this machine is one power-hungry beast the rated power draw of the 3U server machine is 3kW.
Radeon Instinct MI100 Performance:
The leaked slides compare AMD Radeon Instinct MI100 to NVIDIA Volta V100 and Ampere architecture A100 GPU accelerators. One thing to note is that the slides benchmark MI100 against a 300W Ampere A100 GPU, but there is no such variant in production.
This means that these values were estimated from the actual model to put them in perspective with the 300W Radeon card. In the real market, we see A100 in two variants only. There is a 400W SXM form factor device, and then there is a 250W PCIe card.
In the performance benchmarks, we see that the Radeon MI100 delivers about 13 percent improvement in performance over the Apmere A100. It also crushes the Volta V100 GPUs by double the performance. We also have comparisons in performance to the value of cards.
The MI100 offers about 2.4 times better value as compared to NVIDIA’s V100s and about 50 percent better value from the Apere A100. In terms of performance scaling, the MI100 astonishingly shows almost a linear behavior of up to 32 GPU configs in Resenet.
The slides also explain how AMD’s HPC GPU will offer better performance in specialized segments like Oil & Gas, Academia, and HPC & Machine Learning. IIn other HPC workloads like FP64 compute, AI, and Data analysis, the A100 accelerator from NVIDIA offers much better performance.
Nvidia also has the upper hand because of Multi-Instance GPU architecture. In comparison, we get 2.5 times better FP64 performance, two times FP16 performance, along with a double tensor performance with tensor cores on Ampere A100 GPU.
Interestingly, there is no mention of NVIDIA’s sparsity numbers in the slides or the benchmarks by AMD. With the help of sparsity, Ampere A100 for NVIDIA goes up to 156TFLOPs in performance. AMD had a very specific benchmark comparison in mind while making these slides.
Can It Game?
While we keep on mentioning AMD Radeon Instinct MI100 as a GPU, it is not because it is made for gaming. At the same time, it is built on the 7nm+ process and has some very interesting set of specifications that would fly through modern titles. This is a compute card, and there is no gamer feature in it. There is Direct X support in the card, and since all modern games rely on Direct X video libraries, they might not run.
There is also no display output on the card, but that is something we see a lot on the mining cards. We can easily rout the output to the motherboard display output with driver adjustments. However, Even if you do get a game running on this card, it will not perform as good as a gaming GPU. In addition, you’ll be stuck with custom drivers made from untrusted sources. Because of no driver support from the manufacturer.
Lastly, the Radeon Instinct MI100 looks excellent on paper and offers great HPC performance and value. Let’s hope that these numbers hold up in the launch of the device. Needless to say, the above information is just a leak. The final product might vary from these specifications and performance numbers. On the off chance that they are the actual specifications, we are looking at a great future for computing area as well.
AMD will also Launch their consumer gaming cards soon, and we’re excited about those as well.