Mysterious NVIDIA 'GPU-N' Could Be Next-Gen Hopper GH100 In Disguise With 134 SMs, 8576 Cores & 2.68 TB/s Bandwidth, Simulated Performance Benchmarks Shown

A mysterious NVIDIA GPU known equally GPU-North which could possibly be the first expect at the next-gen Hopper GH100 fleck has been revealed in a new research paper published by the light-green team (as discovered past Twitter user, Redfire).

NVIDIA Research Paper Talks 'GPU-N' With MCM Design & 8576 Cores, Could This Be Next-Gen Hopper GH100?

The research paper 'GPU Domain Specialization via Composable On-Bundle Architecture' talks about a side by side-generation GPU design as the most practical solution for maximizing low-precision matrix math throughput to boost Deep Learning operation. The 'GPU-N' and its respective COPA designs have been discussed along with their possible specifications and simulated functioning results.

The 'GPU-N' is said to feature 134 SM units (vs 104 SM units of A100). This makes up a full of 8576 cores or a 24% increment over the current Ampere A100 solution. The bit has been measured at 1.4 GHz, the aforementioned theoretical clock speed of the Ampere A100 and Volta V100 (not to be dislocated as the final clocks). Other specifications include a 60 MB L2 cache, a 50% increment over Ampere A100, and a DRAM bandwidth of 2.68 TB/due south that tin can scale up to 6.3 TB/s. The HBM2e DRAM capacity is 100 GB and can be expanded upwards to 233 GB with the COPA implementations. It is configured effectually a 6144-scrap motorcoach interface at clock speeds of three.five Gbps.

Configuration	NVIDIA V100	NVIDIA A100	GPU-N
SMs	80	108	134
GPU frequency (GHz)	1.4	one.4	1.4
FP32 (TFLOPS)	xv.7	nineteen.5	24.2
FP16 (TFLOPS)	125	312	779
L2 cache (MB)	6	40	60
DRAM BW (GB/south)	900	i,555	2,687
DRAM Capacity (GB)	xvi	xl	100

Coming to the operation numbers, the 'GPU-North' (presumably Hopper GH100) produces 24.2 TFLOPs of FP32 (24% increment over A100) and 779 TFLOPs FP16 (2.5x increase over A100) which sounds really close to the 3x gains that were rumored for GH100 over A100. Compared to AMD's CDNA 2 'Aldebaran' GPU on the Instinct MI250X accelerator, the FP32 performance is less than one-half (95.7 TFLOPs vs 24.2 TFLOPs) but the FP16 performance is 2.15x college.

From previous information, we know that NVIDIA'south H100 accelerator would exist based on an MCM solution and use TSMC's 5nm process node. Hopper is supposed to have ii adjacent-gen GPU modules so we are looking at 288 SM units in total. We can't give a rundown on the core count still since we don't know the number of cores featured in each SMs just if it's going to stick to 64 cores per SM, so we get 18,432 cores which are 2.25x more than the total GA100 GPU configuration. NVIDIA could also leverage more FP64, FP16 & Tensor cores inside its Hopper GPU which would drive up performance immensely. And that's going to be a necessity to rival Intel's Ponte Vecchio which is expected to characteristic one:ane FP64.

It is likely that the concluding configuration will come with 134 of the 144 SM units enabled on each GPU module and as such, we are probable looking at a single GH100 die in action. Simply it is unlikely that NVIDIA would achieve the aforementioned FP32 or FP64 Flops as MI200's without using GPU Sparsity.

But NVIDIA may probable have a underground weapon in their sleeves and that would exist the COPA-based GPU implementation of Hopper. NVIDIA talks most two Domain-Specialized COPA-GPUs based on next-generation architecture, i for HPC and one for DL segment. The HPC variant features a very standard approach which consists of an MCM GPU blueprint and the respective HBM/MC+HBM (IO) chiplets but the DL variant is where things start to get interesting. The DL variant houses a huge cache on an entirely dissever die that is interconnected with the GPU modules.

Compages	LLC Capacity	DRAM BW	DRAM Chapters
Configuration	(MB)	(TB/s)	(GB)
GPU-N	60	2.7	100
COPA-GPU-1	960	2.7	100
COPA-GPU-2	960	4.5	167
COPA-GPU-3	1,920	ii.7	100
COPA-GPU-4	1,920	4.5	167
COPA-GPU-5	ane,920	6.3	233
Perfect L2	infinite	infinite	space

Various variants accept been outlined with up to 960 / 1920 MB of LLC (Terminal-Level-Cache), HBM2e DRAM capacities of upwards to 233 GB, and bandwidth of up to 6.iii TB/s. These are all theoretical but given that NVIDIA has discussed them now, nosotros may likely see a Hopper variant with such a blueprint during the full unveil at GTC 2022.

NVIDIA Hopper GH100 'Preliminary Specs':

NVIDIA Tesla Graphics Carte	Tesla K40 (PCI-Limited)	Tesla M40 (PCI-Express)	Tesla P100 (PCI-Express)	Tesla P100 (SXM2)	Tesla V100 (SXM2)	NVIDIA A100 (SXM4)	NVIDIA H100 (SMX4?)
GPU	GK110 (Kepler)	GM200 (Maxwell)	GP100 (Pascal)	GP100 (Pascal)	GV100 (Volta)	GA100 (Ampere)	GH100 (Hopper)
Procedure Node	28nm	28nm	16nm	16nm	12nm	7nm	5nm
Transistors	7.1 Billion	eight Billion	15.iii Billion	15.3 Billion	21.1 Billion	54.2 Billion	TBD
GPU Die Size	551 mm2	601 mm2	610 mm2	610 mm2	815mm2	826mm2	~1000mm2?
SMs	xv	24	56	56	lxxx	108	134 (Per Module)
TPCs	15	24	28	28	40	54	TBD
FP32 CUDA Cores Per SM	192	128	64	64	64	64	64?
FP64 CUDA Cores / SM	64	4	32	32	32	32	32?
FP32 CUDA Cores	2880	3072	3584	3584	5120	6912	8576 (Per Module) 17152 (Consummate)
FP64 CUDA Cores	960	96	1792	1792	2560	3456	4288 (Per Module)? 8576 (Complete)?
Tensor Cores	N/A	Due north/A	N/A	N/A	640	432	TBD
Texture Units	240	192	224	224	320	432	TBD
Boost Clock	875 MHz	1114 MHz	1329MHz	1480 MHz	1530 MHz	1410 MHz	~1400 MHz
TOPs (DNN/AI)	North/A	N/A	N/A	N/A	125 TOPs	1248 TOPs 2496 TOPs with Sparsity	TBD
FP16 Compute	N/A	N/A	eighteen.vii TFLOPs	21.two TFLOPs	xxx.four TFLOPs	312 TFLOPs 624 TFLOPs with Sparsity	779 TFLOPs (Per Module)? 1558 TFLOPs with Sparsity (Per Module)?
FP32 Compute	v.04 TFLOPs	6.viii TFLOPs	10.0 TFLOPs	ten.half-dozen TFLOPs	xv.7 TFLOPs	19.iv TFLOPs 156 TFLOPs With Sparsity	24.ii TFLOPs (Per Module)? 193.six TFLOPs With Sparsity?
FP64 Compute	ane.68 TFLOPs	0.2 TFLOPs	4.7 TFLOPs	5.30 TFLOPs	7.80 TFLOPs	xix.5 TFLOPs (nine.7 TFLOPs standard)	24.2 TFLOPs (Per Module)? (12.1 TFLOPs standard)?
Memory Interface	384-bit GDDR5	384-fleck GDDR5	4096-chip HBM2	4096-chip HBM2	4096-flake 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/due south 12 GB HBM2 @ 549 GB/south	sixteen GB HBM2 @ 732 GB/s	xvi GB HBM2 @ 900 GB/s	Up To 40 GB HBM2 @ 1.6 TB/s Up To 80 GB HBM2 @ 1.6 TB/south	Up To 100 GB HBM2e @ iii.v Gbps
L2 Cache Size	1536 KB	3072 KB	4096 KB	4096 KB	6144 KB	40960 KB	81920 KB
TDP	235W	250W	250W	300W	300W	400W	~450-500W

Source: https://wccftech.com/mysterious-nvidia-gpu-n-could-be-next-gen-hopper-gh100-in-disguise-with-134-sms-8576-cores-2-68-tb-s-bandwidth-simulated-performance-benchmarks-shown/

Posted by: julianmosurlow.blogspot.com