AMD Radeon R9 295X2 – NVidia Killer
AMD fights back… and seems to be successful. While NVidia is currently
promoting it’s TITAN Z series GPU (2x Titan Black), AMD goes ahead with
a dual 295X GPU card. The tremendous heat dissipation of 500 watts is
managed by a pre-mounted watercooling solution from Asetek.
What looks like an AMDized version of the Titan Z, is in fact currently fastest GPU that will become available in Mid April, 2014. Gobbling a whopping 500 Watts off the 12V rail, it means that the two 8pin connectors have to deliver 425 Watts together, meaning that one 12V rail has to deliver at least 20A stable. Breaking the standards here, this card maybe only targeted at enthusiasts, willing to fork over 1400 US$ (MSRP) for a graphics card.
Looking at the bare card, you can clearly see the two HAWAII GPUs rated at 1020 MHz core speed each! This means that the two GPUs are faster clocked than their single sister card, where hte HAWAII GPU is only clocked at 1000 MHz. While earlier Dual-GPU-solutions have been castrated in any way (less GOP/Shader units/less clock speeds), AMD this time seems to do the exact opposite using a better cooling to allow the card run at even higher specs now without the loss of the originate GPU design. While all other cards have been aircooled, the liquidcooling solution is the only way to tame this beast. The effort seems to have come good for AMD as first tests haven’t sown anything negative so far. The reached scores in 3DMark Vantage also show some outrageous values hereby: This puppy hits the 15K mark while NVidia’s Titan Black barely misses the 10K spot. Yet this card is only 400 US$ more expensive than the Titan Black. Okay, let’s say, that NVidia has also an ass-spanking dual-GPU solution at hands with the Titan Z card but honestly, who is willing to fork over 3000 US$ for a graphics card, that is expected to become outdated soon as it is based on the “old” KEPLER architecture. Maxwell is about to hit the road soon (Q3 2014) so why spend so much money if the upcoming GPU architecture is performing much better per watt against Kepler? The market seems to have such geeks that are already waiving their dollar bills in order to lay hands on a Titan Z.
Here’s the overview of the different HAWAII GPUs:
Radeon R9 295X2 | Radeon R9 290X | Radeon R9 290 | |
GPU | 2 × Hawaii | Hawaii | Hawaii |
Structure size | 28 nm | 28 nm | 28 nm |
Transistors | 12.4 billion | 6.2 billion | 6.2 billion |
Shader units | 5632 | 2816 | 2560 |
ROPs | 88 | 44 CU | 40 CU |
Texture units | 352 | 176 | 160 |
Texel per second | 359 GTex/s | 176 GTex/s | 152 GTex/s |
Rasterizing units | 128 | 64 | 64 |
GPU clock | 1020 MHz | 1000 MHz | 947 MHz |
FLOPs (SP/DP) | 11,49 / 2,87 TFlops | 5,63 / 1,4 TFlops | 4,85 / 1,2 TFlops |
Memory | 2 × 4 GByte GDDR5 | 4 GByte GDDR5 | 4 GByte GDDR5 |
Memory clock | 2500 MHz | 2500 MHz | 2500 MHz |
Memory interface | 2 × 512 Bit | 512 Bit | 512 Bit |
Data transfer speed | 640 GByte/s | 320 GByte/s | 320 GByte/s |
Power connectors | 2×8-pin | 1×6-pin, 1×8-pin | 1×6-pin, 1×8-pin |
Form factor | Dual-Slot | Dual-Slot | Dual-Slot |
Display connectors | DL-DVI, 4×MiniDP | 2 × DL-DVI, HDMI, DP | DL-DVI, HDMI, 2 × MiniDP |
Multimonitor operation | 5 | 4 | 4 |
4K ready | yes | yes | yes |
TDP | 500 Watts | 250 Watts | 250 Watts |
Direct3D | 11.2 | 11.2 | 11.2 |
Mantle | yes | yes | yes |
3DMark Firestrike score | 15862 | 9457 | 8910 |
Price (MSRP) | 1410 US$ | 499 US$ | 389 US$ |
The Titan series of NVidia seems to be also a very competing piece of GPU as they hit stores earlier this year but the MSRP of these cards is ways higher than of those made by AMD. The Comparision charts will follow later on. So what is the Titan Z all about? The Titan cards have been beating every GPU out to date then. Even the R9 280X has been overthrown by the Titan. The 780 TI took the step ahead making use of all available Shader units on the die. Yet the throughput on FLOPs has been artificially reduced from the drivers in order to promote the Titan cards. So it ain’t surprising much that a 780 TI, although better-equipped than a common Titan card is not as powerful as the Titan. Why NVidia takes this “cheating” in order to gain more money from the more-expensive cards, can be explained only this way: They simply cannot drag more out of the Kepler GPU and try to scale the GPU otherwise. he only way is via drivers and GPU identification. If there’s an option to raise FLOPs-power on DP using driver modifications, is unknown, but one thing is for sure: A 780 TI would beat a Titan without any problems and without any limitations as the TDP is the same (250W)
Having a look at the comparison table confirms that the 780 TI is in fact nothing else than a Titan Black with sligtly reduced clock speeds. The rest is identical. But as already stated: NVidia is NOT going to let 300 US$ go to waste. That’s why the driver limitations kick in. The only stupid in this game is, like always, the customer.
As for the Titan Z, the design is triple-slot and a giant air cooler. The big air fan is necessary as the nostrils in the third slot area so that this hot puppy can breathe. The TDP is a secret to date but expect anything near the 500 watts border as two Titan black GPUs want to be fed. And since one GPU already gobbles down 250W, well, you do the math here.
Here’s the comparison chart of the different NVidia cards:
GTX Titan Z | GTX Titan Black | GTX 780 Ti | GTX Titan | |
GPU | 2xGK110 Kepler | GK110 Kepler | GK110 Kepler | GK110 Kepler |
Structure size | 28nm | 28nm | 28nm | 28nm |
Transistors | 14.2 billion | 7.1 billion | 7.1 billion | 7.1 billion |
Stream Processors | 5760 | 2880 | 2880 | 2688 |
ROPs | 96 | 48 | 48 | 48 |
Texture Units | 480 | 240 | 240 | 224 |
Texel per second | 420.2GTex/s | 213.4GTex/s | 210.2GTex/s | 187.5GTex/s |
Core Clock | 700MHz(?) | 889MHz | 875MHz | 837MHz |
Boost Clock | 775MHz(?) | 980MHz | 928MHz | 876MHz |
FLOPs (SP/DP) | 8.0 / 2.6 TFLOPs | 5.1 / 1.7 TFLOPs | 5.0 / 0.2 TFLOPs | 4.5 / 1.3 TFLOPs |
Memory | 2x6GByte GDDR5 | 6GByte GDDR5 | 3GByte GDDR5 | 6GByte GDDR5 |
Memory Clock | 7000 MHz | 7000 MHz | 7000 MHz | 6000 MHz |
Memory Bus Width | 2 x 384-bit | 384-bit | 384-bit | 384-bit |
Data transfer speed | 674.0GB/s | 337.0GB/s | 336.0GB/s | 288.0GB/s |
Power connectors | 2x8Pin | 1x8Pin+1x6Pin | 1x8Pin+1x6Pin | 1x8Pin+1x6Pin |
Form factor | Triple-Slot | Dual-Slot | Dual-Slot | Dual-Slot |
Display connectors | 2 × DL-DVI, HDMI, DP | 2 × DL-DVI, HDMI, DP | 2 × DL-DVI, HDMI, DP | 2 × DL-DVI, HDMI, DP |
Multimonitor op. | 3 | 3 | 3 | 3 |
4K ready | yes | yes | partially | yes |
TDP | ? | 250W | 250W | 250W |
Direct3D | 11.2 | 11.2 | 11.2 | 11.2 |
CUDA | yes | yes | yes | yes |
3DMark Firestrike | 16-17K(?) | 9436 | 7437 | 8771 |
Price (MSRP) | 2999 US$ | 999 US$ | 699 US$ | 999 US$ |
Leave a comment