Focusing on competing in the HPC market, as NVIDIA has warned, this GPU is the largest designed to date and as we will see below, the numbers are really fast and no comparisons are possible.
Designed as a scientific research card, artificial intelligence or neural networks to highlight some of the tasks that will be used, NVIDIA is approaching a high-profit field and is passionate about power.
NVIDIA Ampere GA100, the creation of a beast that will destroy AMD
Although this version will have nothing to do with its expected gaming card, NVIDIA has originally released all the muscle it has and put all the flesh on its head to surprise the world and shock its rivals. First of all, this Ampere GA100 being the largest chip and in its advanced configuration will hit the market in several ways: Mezzanine, PCIe 4.0 and maybe even MXM.
This is not really a surprise, because NVIDIA made it with Volta, but when we look at its size … it's amazing. It has an area of 826 mm2 and surpasses its previous one by 11 mm2, which may not be the norm, but in that extra space to make it its perfect destination, NVIDIA has managed to install 54 billion transistors.
In other words, 54,000,000,000 transistors were stored in 826 mm2 silicon, which is more than twice as much as its predecessor. So, we're talking about the smallest GPU ever created and created so far.
Technical specification
| GPU | Kepler GK110 | Maxwell GM200 | Pascal GP100 | Volta GV100 | Ampere GA100 |
|---|---|---|---|---|---|
| Calculation capacity | 3.5 | 5.3 | 6.0 | 7.0 | 8.0 |
| Threads / Warp | 32 | 32 | 32 | 32 | 32 |
| Max Warps / Multiprocessor | 64. | 64. | 64. | 64. | 64. |
| String threads / multiprocessor | 2048 | 2048 | 2048 | 2048 | 2048 |
| Max Thread / Multiprocessor Blocks | 16. | 32 | 32 | 32 | 32 |
| Max 32bit / SM registers | 65536 | 65536 | 65536 | 65536 | 65536 |
| Big records / block | 65536 | 32768 | 65536 | 65536 | 65536 |
| Records / main thread | 255 | 255 | 255 | 255 | 255 |
| The maximum size of the cord block | 1024 | 1024 | 1024 | 1024 | 1024 |
| CUDAs / cores | 192 | 128 | 64. | 64. | 64. |
| Shared Memory Friendly / SM Settings (bytes) | 16K / 32K / 48K | $ 96K | 64K | $ 96K | 164K |
If these numbers are already dull, it should be noted that it contains 128 SM with 8192 CUDA cores, distributed to 8192 Shaders capable of working with FP32, 4096 capable of operating on FP64 and 512 Tensor Cores, all at 8 GPCs where each contains 16 SM and 8 TPC.
These impressive numbers are also reflected in its usage, as NVIDIA has revealed that each Ampere GA100 will go up to 400 watts on its various Tesla A100s.
If this didn't seem good enough, this GPU would integrate a 6144-bit bus that would be compatible with 48 GB of HBM2E memory with six batteries, all in one translation space.
Each stack has a capacity of 2 GB, so each die will have 4 Hi at an average of 8 GB where each individual will have a speed of each pin over 2 Gbps. This would also leave us with a total number of true bandwidth: 1.6 Tbps.
The Ampere GA100 will have its limited edition Tesla A100
As expected, NVIDIA will introduce its Tesla GPUs in some ways set to a chip that, in this case, yes, can start giving us clues as to whether the rumors about GA102 are true or not.
These Tesla A100s will come with reduced GA100 chips, offering only 108 SM of 6912 cores of FP32, 3456 cores that will be able to interact with FP64 and 432 Tensor Cores. So the bus will be downgraded, straight to the equally impressive 5,120-bit values that will result in a VRAM value of 40GB.
Rape will vary between 400 watts and 300 watts, depending on the structure and frequency they install, something that NVIDIA has not been properly exposed to.
The DGX teams will also use the Tesla V100 models, but this time under Ampere
There is nothing spectacular here, much less since Huang released the infamous fire video. The companies of its company DGX and HGX will focus on the responsibilities of AI and HPC on the one hand, and cloud computing and data centers on the other.
In any case, they will be third-generation machines that can be configured up to 10U, a real madness today. Each server will have 8 of these V100 boards imported by NVIDIA and will use the well-known PCIe 4.0 interfaces to maximize performance.
Huang gave some concrete power calculations, specifically, and said the DGX-A100 will reach 5 Petaflops through 8 Tesla A100 Ampere GPUs, a figure over 20 times the performance of its previous Volta-based version.
The systems start at a cost of $ 199,000 and are available from today via NVIDIA.
Table of Contents
