NVIDIA Ampere A100 PCIe GPU Launched, 20 Times Faster Than Volta

NVIDIA has added a third variant to its growing Ampere A100 GPU family unit, the A100 PCIe which is PCIe 4.0 compliant and comes in the standard full-length, full top course gene compared to the mezzanine board we got to see earlier.

NVIDIA's A100 Ampere GPU Gets PCIe 4.0 Ready Class Gene - Same GPU Configuration Only at 250W, Upwardly To 90% Operation of the Total 400W A100 GPU

Simply similar the Pascal P100 and Volta V100 before it, the Ampere A100 GPU was bound to get a PCIe variant sooner or later. At present NVIDIA has announced that its A100 PCIe GPU accelerator is available for a diverse ready of apply cases with system ranging from a single A100 PCIe GPU to servers utilizing 2 cards at the same time through the 12 NVLINK channels that deliver 600 GB/due south of interconnect bandwidth.

In terms of specifications, the A100 PCIe GPU accelerator doesn't modify much in terms of core configuration. The GA100 GPU retains the specifications we got to see on the 400W variant with 6912 CUDA cores bundled in 108 SM units, 432 Tensor Cores and twoscore GB of HBM2 memory that delivers the same memory bandwidth of 1.55 TB/south (rounded off to one.6 TB/due south). The master difference can be seen in the TDP which is rated at 250W for the PCIe variant whereas the standard variant comes with a 400W TDP.

At present nosotros can judge that the card would characteristic lower clocks to compensate for the less TDP input but NVIDIA has provided the peak compute numbers and those remain unaffected for the PCIe variant. The FP64 performance is still rated at nine.vii/19.5 TFLOPs, FP32 operation is rated at 19.v /156/312 TFLOPs (Sparsity), FP16 performance is rated at 312/624 TFLOPs (Sparsity) & INT8 is rated at 624/1248 TOPs (Sparsity).

Co-ordinate to NVIDIA, the A100 PCIe accelerator can evangelize ninety% the performance of the A100 HGX bill of fare (400W) in summit server applications. This is mainly due to the less fourth dimension it takes for the card to achieve the said tasks however, in complex situations which required sustained GPU capabilities, the GPU tin deliver anywhere from up to ninety% to down to 50% the functioning of the 400W GPU in the most extreme cases. NVIDIA told that the 50% drop will be very rare and just a few tasks can push button the card to such extend.

NVIDIA Ampere GA100 GPU Based Tesla A100 Specs:

NVIDIA Tesla Graphics Card	Tesla K40 (PCI-Limited)	Tesla M40 (PCI-Limited)	Tesla P100 (PCI-Express)	Tesla P100 (SXM2)	Tesla V100 (SXM2)	Tesla V100S (PCIe)	NVIDIA A100 (SXM4)	NVIDIA A100 (PCIe4)
GPU	GK110 (Kepler)	GM200 (Maxwell)	GP100 (Pascal)	GP100 (Pascal)	GV100 (Volta)	GV100 (Volta)	GA100 (Ampere)	GA100 (Ampere)
Process Node	28nm	28nm	16nm	16nm	12nm	12nm	7nm	7nm
Transistors	7.i Billion	8 Billion	15.3 Billion	15.3 Billion	21.1 Billion	21.i Billion	54.two Billion	54.two Billion
GPU Dice Size	551 mm2	601 mm2	610 mm2	610 mm2	815mm2	815mm2	826mm2	826mm2
SMs	fifteen	24	56	56	80	eighty	108	108
TPCs	15	24	28	28	40	40	54	54
FP32 CUDA Cores Per SM	192	128	64	64	64	64	64	64
FP64 CUDA Cores / SM	64	iv	32	32	32	32	32	32
FP32 CUDA Cores	2880	3072	3584	3584	5120	5120	6912	6912
FP64 CUDA Cores	960	96	1792	1792	2560	2560	3456	3456
Tensor Cores	N/A	Due north/A	N/A	North/A	640	640	432	432
Texture Units	240	192	224	224	320	320	432	432
Boost Clock	875 MHz	1114 MHz	1329MHz	1480 MHz	1530 MHz	1601 MHz	1410 MHz	1410 MHz
TOPs (DNN/AI)	N/A	N/A	N/A	Due north/A	125 TOPs	130 TOPs	1248 TOPs 2496 TOPs with Sparsity	1248 TOPs 2496 TOPs with Sparsity
FP16 Compute	N/A	North/A	18.7 TFLOPs	21.2 TFLOPs	thirty.iv TFLOPs	32.8 TFLOPs	312 TFLOPs 624 TFLOPs with Sparsity	312 TFLOPs 624 TFLOPs with Sparsity
FP32 Compute	5.04 TFLOPs	6.8 TFLOPs	10.0 TFLOPs	10.vi TFLOPs	xv.7 TFLOPs	16.4 TFLOPs	156 TFLOPs (xix.five TFLOPs standard)	156 TFLOPs (19.v TFLOPs standard)
FP64 Compute	1.68 TFLOPs	0.2 TFLOPs	iv.7 TFLOPs	5.30 TFLOPs	seven.80 TFLOPs	8.ii TFLOPs	19.v TFLOPs (ix.7 TFLOPs standard)	xix.5 TFLOPs (9.7 TFLOPs standard)
Memory Interface	384-bit GDDR5	384-chip GDDR5	4096-fleck HBM2	4096-bit HBM2	4096-bit HBM2	4096-bit HBM2	6144-bit HBM2e	6144-bit HBM2e
Memory Size	12 GB GDDR5 @ 288 GB/s	24 GB GDDR5 @ 288 GB/s	16 GB HBM2 @ 732 GB/south 12 GB HBM2 @ 549 GB/south	xvi GB HBM2 @ 732 GB/southward	16 GB HBM2 @ 900 GB/southward	16 GB HBM2 @ 1134 GB/s	Upwards To 40 GB HBM2 @ one.6 TB/south Up To 80 GB HBM2 @ 1.6 TB/s	Up To forty GB HBM2 @ one.6 TB/s Upwardly To 80 GB HBM2 @ ii.0 TB/southward
L2 Cache Size	1536 KB	3072 KB	4096 KB	4096 KB	6144 KB	6144 KB	40960 KB	40960 KB
TDP	235W	250W	250W	300W	300W	250W	400W	250W

There's a wide scale adoption beingness made possible already by NVIDIA and its server partners for the said PCIe based GPU accelerator which include:

ASUS will offer the ESC4000A-E10, which can exist configured with four A100 PCIe GPUs in a unmarried server.
Atos is offering its BullSequana X2415 organisation with iv NVIDIA A100 Tensor Core GPUs.
Cisco plans to back up NVIDIA A100 Tensor Cadre GPUs in its Cisco Unified Calculating Arrangement servers and in its hyperconverged infrastructure system, Cisco HyperFlex.
Dell Technologies plans to back up NVIDIA A100 Tensor Cadre GPUs beyond its PowerEdge servers and solutions that accelerate workloads from edge to core to deject, but every bit it supports other NVIDIA GPU accelerators, software and technologies in a wide range of offerings.
Fujitsu is bringing A100 GPUs to its PRIMERGY line of servers.
GIGABYTEwill offer G481-HA0, G492-Z50 and G492-Z51 servers that support up to ten A100 PCIe GPUs, while the G292-Z40 server supports up to eight.
HPE will support A100 PCIe GPUs in the HPE ProLiant DL380 Gen10 Server, and for accelerated HPC and AI workloads, in the HPE Apollo 6500 Gen10 Organization.
Inspur is releasing eight NVIDIA A100-powered systems, including the NF5468M5, NF5468M6 and NF5468A5 using A100 PCIe GPUs, the NF5488M5-D, NF5488A5, NF5488M6 and NF5688M6 using viii-way NVLink, and the NF5888M6 with 16-fashion NVLink.
Lenovo will support A100 PCIe GPUs on select systems, including the Lenovo ThinkSystem SR670 AI-ready server. Lenovo will expand availability across its ThinkSystem and ThinkAgile portfolio in the fall.
One Stop Systemsvolition offer its OSS 4UV Gen 4 PCIe expansion organization with up to eight NVIDIA A100 PCIe GPUs to permit AI and HPC customers to scale out their Gen 4 servers.
Quanta/QCT will offering several QuantaGrid server systems, including D52BV-2U, D43KQ-2U and D52G-4U that back up upwards to 8 NVIDIA A100 PCIe GPUs.
Supermicro will offer its 4U A+ GPU organisation, supporting up to eight NVIDIA A100 PCIe GPUs and up to two additional high-performance PCI-E iv.0 expansion slots along with other 1U, 2U and 4U GPU servers.

NVIDIA hasn't announced any release engagement or pricing for the card nevertheless only considering the A100 (400W) Tensor Cadre GPU is already existence shipped since its launch, the A100 (250W) PCIe will be post-obit its footsteps soon.