Overclock.net › Forums › Industry News › Hardware News › [Computerbase] - 3gb vs 4gb vs 6gb vs 8gb GDDR5 VRAM Frametime Testing
New Posts  All Forums:Forum Nav:

[Computerbase] - 3gb vs 4gb vs 6gb vs 8gb GDDR5 VRAM Frametime Testing - Page 16

post #151 of 175
Quote:
Originally Posted by EightDee8D View Post

Iris pro and xbone has small but fast eDRAM, if being fast doesn't solve the size issue ( to some extent) than why are those 2 have pretty small eDRAM in size ? why is that when a cache is fast it's smaller, and when it is slow it's bigger ? deny all you want, but the faster speed does makes up for smaller size. yes it doesn't solve it completely but it does help. because no texture is a solid 4GB+ large in size.

you realize eDRAM is a "last level cache" right? its simply the same reason why CPUs have L2 and L3 caches.
it makes access to critical resources more readily available and more responsive than relying on the system ram.
and if you're comparing eDRAM to HBM, then those machines that uses eDRAM should work without any additional ram at all, much like HBM.

plus size has no relation with speed, you can simply stuff in more HBM chips at a higher cost, look at Nvidia's P100 cards.



and on this point, HBM should've just been used as a cache with GDDR5 as the main memory pool, with this it would be literally akin to how eDRAM is implemented.
2GB HBM1 (2048bit @ 256GB/s) + 6GB GDDR5 (384bit @ ~300GB/s) is a lot more cheaper and has more logical benefit.

furthermore this will allow the usage of an off-die GDDR5 IMC, making the more expensive GPU die smaller.
and the reasoning is simple, onboard GDDR5 is still much faster than DDR3/DDR4 through PCI-E.
Edited by epic1337 - 9/27/16 at 5:39am
post #152 of 175
Quote:
Originally Posted by EightDee8D View Post

why is that when a cache is fast it's smaller, and when it is slow it's bigger ?

Because fast usually means expensive. You use a smaller faster expensive cache to store the most commonly used assets so you don't have to repeatedly fetch them from slower larger cheaper storage.

You do realize that the Titan Xp has almost the same memory bandwidth as the Fury X, right? There's nothing magical about how fast HBM is.

And what does no texture is larger than 4GB in size have to do with anything? Of course it isn't, but the card isn't storing just one texture in VRAM.
post #153 of 175
Quote:
Originally Posted by neptunex View Post

You can ask as many questions as you like. There is a clear difference between being curious on the subject and seeking appropriate knowledge and arguing on the subject you have no knowledge about.

Quote:
Originally Posted by Forceman View Post

"AMD is hand-tuning games to ensure 4GB will not hold Fiji performance back". Which is exactly what Blameless and others are saying, they are managing the memory allocation through driver memory management - HBM is not somehow magically storing 6GB of data in 4GB of VRAM.

And that guy lost all credibility when he said you can't compare SSD and mechanical hard drive capabilities. Umm, what? Someone better tell all the SSD manufacturers. 4GB is 4GB, whether it is HBM, GDDR5, SSD or hard drive.

No one has ever said that AMD can't more intelligently manage VRAM to keep the 4GB limitation from being a problem, but the idea that there is something inherent in HBM that makes that possible needs to die. At launch, AMD stated that they had always taken a rather cavalier approach to memory management because they had essentially always had enough, and that there were better ways to allocate assets to reduce usage. Which is exactly what they are doing. But those efforts can be applied to GDDR as well as HBM.

Edit: and all VRAM is essentialy an embedded cache, storing on-card assets that otherwise would have to be fetched from system RAM when needed.


I will try one more time.

No one said AMD is not managing the memory pool with drivers. Every gpu does that. What no one is willing to admit is even possible, is that AMD is able to better manage that memory pool because the latency/distance is less with HBM. No one ever stated HBM gets rid of VRAM limits but you.

Prove that wrong and I will concede completely.

And not all "4GB" is "4GB"-just ask a GTX 970 owner.
Amelia
(13 items)
 
Professional
(13 items)
 
RCPC#1
(17 items)
 
CPUMotherboardGraphicsRAM
Phenom II X6 1100t MSI 890FX GD65 MSI Radeon HD5670 GSkill RipjawsX DDR3 PC3 12800 2x4GB CL8 
Hard DriveOptical DriveCoolingOS
WD Black 1TB SATA III Samsung BD Zalman 9900MAX Windows 7 64 Professional 
MonitorKeyboardPowerCase
AOC 22" LED Logitech Kingwin Lazer Platinum 500w Fractal Design R3 
Other
Samsung 470 SSD 128GB 
CPUMotherboardGraphicsRAM
AMD Phenom II X6 960T Asus M4A88T-VEVO Asus Strix R7 370 SuperTalent Perfomance 
RAMHard DriveHard DriveOptical Drive
GSkill Snipers Monster Daytona Seagate Barracuda 500GB 7,200 RPM 16Mb cache Memorex DVD/RW 
CoolingOSMonitorKeyboard
Corsair H60 Windows 8N IBM 9494 19" LCD IBM 
PowerCaseMouseMouse Pad
Corsair GS500 In Win H-Frame Wolfking OCZ Behemoth 
Audio
JBL Creature 
  hide details  
Reply
Amelia
(13 items)
 
Professional
(13 items)
 
RCPC#1
(17 items)
 
CPUMotherboardGraphicsRAM
Phenom II X6 1100t MSI 890FX GD65 MSI Radeon HD5670 GSkill RipjawsX DDR3 PC3 12800 2x4GB CL8 
Hard DriveOptical DriveCoolingOS
WD Black 1TB SATA III Samsung BD Zalman 9900MAX Windows 7 64 Professional 
MonitorKeyboardPowerCase
AOC 22" LED Logitech Kingwin Lazer Platinum 500w Fractal Design R3 
Other
Samsung 470 SSD 128GB 
CPUMotherboardGraphicsRAM
AMD Phenom II X6 960T Asus M4A88T-VEVO Asus Strix R7 370 SuperTalent Perfomance 
RAMHard DriveHard DriveOptical Drive
GSkill Snipers Monster Daytona Seagate Barracuda 500GB 7,200 RPM 16Mb cache Memorex DVD/RW 
CoolingOSMonitorKeyboard
Corsair H60 Windows 8N IBM 9494 19" LCD IBM 
PowerCaseMouseMouse Pad
Corsair GS500 In Win H-Frame Wolfking OCZ Behemoth 
Audio
JBL Creature 
  hide details  
Reply
post #154 of 175
Quote:
We just need to do a better job managing the capacities. We were getting free capacity, because with [GDDR5] in order to get more bandwidth we needed to make the memory system wider, so the capacities were increasing. As engineers, we always focus on where the bottleneck is. If you're getting capacity, you don't put as much effort into better utilising that capacity. 4GB is more than sufficient. We've had to go do a little bit of investment in order to better utilise the frame buffer, but we're not really seeing a frame buffer capacity [problem]. You'll be blown away by how much [capacity] is wasted."
- Joe Macri
post #155 of 175
So basically nothing to prove HBM faster vram doesn't solve anything. with eDRAM + huge G5 vram you are still getting less performance vs 4gb HBM + system ram. because you don't need to store large amounts of textures for immediate frame, and system ram can provide ample bandwidth to those textures which requires less frequent use.


there's a reason why 980 struggles in vram limited situations where Furyx doesn't, even 290x doesn't struggle in those situations even though it doesn't have HBM but it does have faster vram than 980's. BECAUSE faster read+write helps moving less important data very fast.
Edited by EightDee8D - 9/27/16 at 6:04am
post #156 of 175
Quote:
Originally Posted by Redwoodz View Post

What no one is willing to admit is even possible, is that AMD is able to better manage that memory pool because the latency/distance is less with HBM.

Because such statements are absurd and don't make any sense to anyone who has any idea what memory is used for.
Quote:
Originally Posted by EightDee8D View Post

So basically nothing to prove HBM faster vram doesn't solve anything. with eDRAM + huge G5 vram you are still getting less performance vs 4gb HBM + system ram. because you don't need to store large amounts of textures for immediate frame, and system ram can provide ample bandwidth to those textures which requires less frequent use.


there's a reason why 980 struggles in vram limited situations where Furyx doesn't, even 290x doesn't struggle in those situations even though it doesn't have HBM but it does have faster vram than 980's. BECAUSE faster read+write helps moving less important data very fast. but i guess it's too difficult to admit until nvidia does it.

No one is saying that memory performance doesn't matter, but the idea that memory performance can replace memory capacity...the idea that memory performance can benefit what isn't in memory...is idiotic.

I'm not sure why you are turning this into an AMD vs. NVIDIA thing either. Absurd statements should be called out, no matter what brand they appear to favor.
Quote:
Originally Posted by Marios145 View Post

- Joe Macri

Yep.

No one at AMD has ever implied that memory management with HBM is intrinsically different. Fury has shown us that a lot of VRAM is indeed wasted...but for some reason we have people coming away with the idea that HBM is magically changing what memory does.

HBM is faster memory, that is it.
Primary
(15 items)
 
Secondary
(13 items)
 
In progress
(10 items)
 
CPUMotherboardGraphicsRAM
5820K @ 4.2/3.5GHz core/uncore, 1.175/1.15v Gigabyte X99 SOC Champion (F22n) Gigabyte AORUS GTX 1080 Ti (F3P) @ 2025/1485, 1... 4x4GiB Crucial @ 2667, 12-12-12-28-T1, 1.34v 
Hard DriveHard DriveHard DriveCooling
Plextor M6e 128GB (fw 1.06) M.2 (PCI-E 2.0 2x) 2x Crucial M4 256GB 4x WD Scorpio Black 500GB Noctua NH-D15 
OSMonitorKeyboardPower
Windows 7 Professional x64 SP1 BenQ BL3200PT Filco Majestouch Tenkeyless (MX Brown) Corsair RM1000x 
CaseMouseAudio
Fractal Design Define R4 Logitech G402 Realtek ALC1150 + M-Audio AV40 
CPUMotherboardGraphicsRAM
X5670 @ 4.4/3.2GHz core/uncore, 1.36 vcore, 1.2... Gigabyte X58A-UD5 r2.0 w/FF3mod10 BIOS Sapphire Fury Nitro OC+ @ 1053/500, 1.225vGPU/1... 2x Samsung MV-3V4G3D/US @ 2000, 10-11-11-30-T1,... 
RAMHard DriveHard DriveHard Drive
1x Crucial BLT4G3D1608ET3LX0 @ 2000, 10-11-11-3... OCZ (Toshiba) Trion 150 120GB Hyundai Sapphire 120GB 3x Hitachi Deskstar 7k1000.C 1TB 
CoolingOSPowerCase
Noctua NH-D14 Windows 7 Pro x64 SP1 Antec TP-750 Fractal Design R5 
Audio
ASUS Xonar DS 
CPUMotherboardGraphicsRAM
i7-6800K @ 4.3/3.5GHz core/uncore, 1.36/1.2v ASRock X99 OC Formula (P3.10) GTX 780 (temporary) 4x4GiB Crucial DDR4-2400 @ 11-13-12-28-T2, 1.33v 
Hard DriveHard DriveCoolingOS
Intel 600p 256GB NVMe 2x HGST Travelstar 7k1000 1TB Corsair H55 (temporary) Windows Server 2016 Datacenter 
PowerCase
Seasonic SS-860XP2 Corsair Carbide Air 540 
  hide details  
Reply
Primary
(15 items)
 
Secondary
(13 items)
 
In progress
(10 items)
 
CPUMotherboardGraphicsRAM
5820K @ 4.2/3.5GHz core/uncore, 1.175/1.15v Gigabyte X99 SOC Champion (F22n) Gigabyte AORUS GTX 1080 Ti (F3P) @ 2025/1485, 1... 4x4GiB Crucial @ 2667, 12-12-12-28-T1, 1.34v 
Hard DriveHard DriveHard DriveCooling
Plextor M6e 128GB (fw 1.06) M.2 (PCI-E 2.0 2x) 2x Crucial M4 256GB 4x WD Scorpio Black 500GB Noctua NH-D15 
OSMonitorKeyboardPower
Windows 7 Professional x64 SP1 BenQ BL3200PT Filco Majestouch Tenkeyless (MX Brown) Corsair RM1000x 
CaseMouseAudio
Fractal Design Define R4 Logitech G402 Realtek ALC1150 + M-Audio AV40 
CPUMotherboardGraphicsRAM
X5670 @ 4.4/3.2GHz core/uncore, 1.36 vcore, 1.2... Gigabyte X58A-UD5 r2.0 w/FF3mod10 BIOS Sapphire Fury Nitro OC+ @ 1053/500, 1.225vGPU/1... 2x Samsung MV-3V4G3D/US @ 2000, 10-11-11-30-T1,... 
RAMHard DriveHard DriveHard Drive
1x Crucial BLT4G3D1608ET3LX0 @ 2000, 10-11-11-3... OCZ (Toshiba) Trion 150 120GB Hyundai Sapphire 120GB 3x Hitachi Deskstar 7k1000.C 1TB 
CoolingOSPowerCase
Noctua NH-D14 Windows 7 Pro x64 SP1 Antec TP-750 Fractal Design R5 
Audio
ASUS Xonar DS 
CPUMotherboardGraphicsRAM
i7-6800K @ 4.3/3.5GHz core/uncore, 1.36/1.2v ASRock X99 OC Formula (P3.10) GTX 780 (temporary) 4x4GiB Crucial DDR4-2400 @ 11-13-12-28-T2, 1.33v 
Hard DriveHard DriveCoolingOS
Intel 600p 256GB NVMe 2x HGST Travelstar 7k1000 1TB Corsair H55 (temporary) Windows Server 2016 Datacenter 
PowerCase
Seasonic SS-860XP2 Corsair Carbide Air 540 
  hide details  
Reply
post #157 of 175
Quote:
Originally Posted by epic1337 View Post

and on this point, HBM should've just been used as a cache with GDDR5 as the main memory pool, with this it would be literally akin to how eDRAM is implemented.
2GB HBM1 (2048bit @ 256GB/s) + 6GB GDDR5 (384bit @ ~300GB/s) is a lot more cheaper and has more logical benefit.

furthermore this will allow the usage of an off-die GDDR5 IMC, making the more expensive GPU die smaller.
and the reasoning is simple, onboard GDDR5 is still much faster than DDR3/DDR4 through PCI-E.

That is so much backwards thinking. Why would a cache that is slower and smaller than the main memory pool ever be a good thing? Why would moving the main pool memory controller off-die ever be good? We already saw the big improvements moving it on die has had for AMD and Intel CPUs. And why would having to manage a chip on an interposer along with doing a whole other set of memory and traces be a good thing? It would make things even more complicated and prone to failure in production and thus drastically raise costs for nothing but decreases in performance.
Gaming
(17 items)
 
Gaming PC
(20 items)
 
 
CPUMotherboardGraphicsRAM
7700K AS Rock Z170 OC Formula Titan X Pascal 2050MHz 64GB DDR4-3200 14-14-14-34-1T 
Hard DriveHard DriveHard DriveCooling
950 EVO m.2 OS drive 850 EVO 1TB games drive Intel 730 series 500GB games drive Custom water cooling 
OSMonitorKeyboardPower
Win 10 Pro x64 AMH A399U E-Element mechanical, black switches, Vortex b... EVGA G3 1kw 
CaseMouseAudioAudio
Lian-Li PC-V1000L Redragon M901 LH Labs Pulse X Infinity DAC Custom built balanced tube amp with SS diamond ... 
Audio
MrSpeakers Alpha Prime 
  hide details  
Reply
Gaming
(17 items)
 
Gaming PC
(20 items)
 
 
CPUMotherboardGraphicsRAM
7700K AS Rock Z170 OC Formula Titan X Pascal 2050MHz 64GB DDR4-3200 14-14-14-34-1T 
Hard DriveHard DriveHard DriveCooling
950 EVO m.2 OS drive 850 EVO 1TB games drive Intel 730 series 500GB games drive Custom water cooling 
OSMonitorKeyboardPower
Win 10 Pro x64 AMH A399U E-Element mechanical, black switches, Vortex b... EVGA G3 1kw 
CaseMouseAudioAudio
Lian-Li PC-V1000L Redragon M901 LH Labs Pulse X Infinity DAC Custom built balanced tube amp with SS diamond ... 
Audio
MrSpeakers Alpha Prime 
  hide details  
Reply
post #158 of 175
Quote:
Originally Posted by EightDee8D View Post

there's a reason why 980 struggles in vram limited situations where Furyx doesn't, even 290x doesn't struggle in those situations even though it doesn't have HBM but it does have faster vram than 980's. BECAUSE faster read+write helps moving less important data very fast.
any proof ?
post #159 of 175
Well, HBM can handle 4096bits in one cycle(1Hz), the 980 384bits and the 290/390 512bits.
So if you're having chunks of data moving from L2 cache<->VRAM, the fury can push 10 times more in a single cycle compared to the rest.
At its frequency which is too low, latency is waaaay up, but due to having the RAM physically closer on the die you end up with the same or lower latency.
The real gain from HBM1 vs GDDR5 was that it removed all memory speed bottlenecks at 80% less power, which allowed more CUs on die.

Fury is not bottlenecked by 4gb or hbm low speed, it's bottlenecked by the low ROP count(64 vs 96 for 980ti) and bad tesselation performance.




p.s. DOOM uses compute shaders instead of pixel shaders, we all saw how that played out for the Fury X.
Edited by Marios145 - 9/27/16 at 6:24am
post #160 of 175
Quote:
Originally Posted by Klocek001 View Post

any proof ?
No one does tests like that , but i have used both and talking from experience. ( mostly in WD,rom, fc4, som, on 1440p/4k mixed ultra) there are scenarios where 980 stutters but 290x doesn't.

Regardless, we have lots of people with 380x and 960 on this forum right ? so why don't we just test this and end this argument ?

run GTX 960 2GB on default, run same games and settings on GTX 960 4GB with vram frequency lowerd to half. same with 380x. basically same gpu but with different vram config.

there. if 2gb version ( faster vram one ) doesn't improve anything than it means faster vram doesn't solve low capacity problem even 1%. what kind of enthusiast community is this who instead of doing testing on their own, just happy being keyboard warriors ? why don't we just test this simple thing ?
Edited by EightDee8D - 9/27/16 at 6:34am
New Posts  All Forums:Forum Nav:
  Return Home
  Back to Forum: Hardware News
Overclock.net › Forums › Industry News › Hardware News › [Computerbase] - 3gb vs 4gb vs 6gb vs 8gb GDDR5 VRAM Frametime Testing