Review: Gigabyte GeForce RTX 3090 Gaming OC

by Tarinder Sandhu on 24 September 2020, 14:01

Tags: NVIDIA (NASDAQ:NVDA), Gigabyte (TPE:2376)

Quick Link: HEXUS.net/qaeoda

Add to My Vault: x

Introduction

Nvidia came out swinging with the Ampere bat by launching a trio of RTX 30-series graphics cards recently. We took a good look at the underlying architecture and the RTX 3080 Founders Edition in the launch-day review. It was no secret that whist handsome in the benchmark stakes and based on the performance GA102 die, it wasn't the fastest gaming Ampere around. That honour fell to the GeForce RTX 3090. Let's take a closer look at the behemoth in table form.

Turing to Ampere

 
RTX 3090
RTX 3080
RTX 3070
RTX 2080 Ti
RTX 2080 Super
RTX 2080
RTX 2070 Super
Launch date
Sep 2020
Sep 2020
Oct 2020
Sep 2018
July 2019
Sep 2018
July 2019
Codename
GA102
GA102
GA104
TU102
TU104
TU104
TU104
Architecture
Ampere
Ampere
Ampere
Turing
Turing
Turing
Turing
Process (nm)
8
8
8
12
12
12
12
Transistors (bn)
28.3
28.3
17.4
18.6
13.6
13.6
13.6
Die Size (mm²)
628.4
628.4
392.5
754
545
545
545
PCIe
4.0
4.0
4.0
3.0
3.0
3.0
3.0
Base Clock (MHz)
1,400
1,440
1,500
1,350
1,650
1,515
1,605
Boost Clock (MHz)
1,695
1,710
1,725
1,545
1,815
1,710
1,770
Founders Edition Clock (MHz)
1,695
1,710
1,725
1,635
1,815
1,800
1,770
Streaming Multiprocessors
82
68
46
68
48
46
40
Shaders
10,496
8,704
5,888
4,352
3,072
2,944
2,560
GFLOPS
35,581
29,768
20,314
13,448
11,151
10,068
9,062
Founders Edition GFLOPS
35,581
29,768
20,314
14,231
11,151
10,598
9,062
Tensor Cores
328
272
184
544
384
368
320
RT Cores
82
68
46
68
48
46
40
Memory Size
24GB
10GB
8GB
11GB
8GB
8GB
8GB
Memory Bus
384-bit
320-bit
256-bit
352-bit
256-bit
256-bit
256-bit
Memory Type
GDDR6X
GDDR6X
GDDR6
GDDR6
GDDR6
GDDR6
GDDR6
Memory Clock
19.5Gbps
19Gbps
14Gbps
14Gbps
15.5Gbps
14Gbps
14Gbps
Memory Bandwidth
936
760
448
616
496
448
448
ROPs
112
96
64
88
64
64
64
Texture Units
328
272
184
272
192
184
160
L2 cache (KB)
5,120
5,120
4,096
5,632
4,096
4,096
4,096
SLI
Yes
No
No
Yes
Yes
Yes
Yes
Power Connector (FE)
12-pin
12-pin
8-pin
8-pin + 8-pin
8-pin + 6-pin
8-pin + 6-pin
8-pin + 6-pin
TGP (watts)
350
320
220
250
250
215
215
Founders Edition TDP (watts)
350
320
220
260
250
225
215
Suggested MSRP
$1,499
$699
$499
$999
$699
$699
$499
Founders Edition MSRP
$1,499
$699
$499
$1,199
$699
$799
$499

You may recall that the GA102 die is a silicon monster. Comprised of 28.3 billion transistors with a 628.3mm² footprint and built on Samsung 8nm process, it carries 84 streaming multiprocessors (SMs) that each contain 128 Cuda-capable cores. That's 10,752 shaders all told, but like previous generations, Nvidia leaves a little wiggle room for its top-tier regular card. The RTX 3090 uses 82 out of those 84, offering 10,496 shaders, 328 Tensor cores and texture units, 112 ROPs, 82 RT cores, and a 384-bit memory bus connecting to GDDR6X operating at 19.5Gbps. The absolute full-fat GA102 is likely to be the as-yet-unannounced Titan model, left in the wings in case AMD's upcoming 'Big Navi' is better than expected.

RTX 3090 is certainly no poorer cousin of probable Titan, mind, as a 1,695MHz boost frequency combines to offer over 35TFLOPs of compute and almost 1TB/s of memory bandwidth. Those numbers compare favourably to the RTX 3080 by offering 19.5 percent and 23.2 percent more on those two metrics, respectively, but what really sets the RTX 3090 apart is its 24GB framebuffer.

Let's be clear. Having 2.4x the onboard memory pool - the same as RTX Titan from the 20-series - doesn't readily help in gaming as titles are tuned to use 11GB and below. Rather, Nvidia advances the same argument for RTX 3090 as Titan; the massive framebuffer is useful in rendering applications that make liberal use of large datasets within card memory, rather than spooling them out to much slower system RAM and hampering time to completion. Applications such as OctaneRender, AutoDesk, Blender, and DaVinci Resolve are put forward for the RTX 3090/Titan argument. Gamers, on the other hand, are beguiled by the promise of being able to play in 8K60 through adoption of this silicon might and DLSS technology, though such monitors are rarer than retail RTX cards right now.

The GeForce RTX 3090 uses almost all of this GA102 die - 82 out of 84 SMs.

Oh, and in case you missed it, unlike the other two Ampere cards, RTX 3090 can also be teamed with another board for SLI usage through third-generation 112GB/s NVLink connector present on every board. Sounds good on paper, but we reckon the experience will be poor given the lack of resources games developers are investing in multi-GPU optimisations.

Having more of everything that matters pushes the power consumption out to 350W, up from the RTX 3080's 320W. That's no biggie, and Nvidia recommends a 750W PSU as a minimum. Going faster on the core, memory and providing 2.4x the footprint escalates price to $1,499. That's a 114 percent increase of the '80. Nvidia is acutely aware of two things: that those who hanker for the absolute best will pay for it, and stock is sure to be so limited that demand far, far outstrips supply. Nvidia could probably charge double and still sell what we expect to be a meagre allocation.

Nvidia is taking the same tack as all recent launches. There's a good-looking Founders Edition plumbed with the specifications reported in the table above, and a host of partner cards from the likes of Gigabyte, MSI and Asus using a reference board design paired with company-specific coolers that purport improved thermals and acoustics.

With that in mind, we're investigating the Gigabyte GeForce RTX 3090 Gaming OC's credentials.