To The Point
The good:
- Very nice performance for the $200 range.
- Improved filtering algorithms.
- Improved idle power consumption.
- Slightly improved.
- The architectural balance that Radeon HD 5800 series should have had.
- HDMI 1.4a, DisplayPort 1.2.
- 5 digital outputs.
- Tesselation performance improved.
The bad:
- No 64 bit FPU limits GPU computing.
- Compute performance is not great sometimes.
The ugly:
- These cards should have had model numbers Radeon HD 6700 something...
Specifications
AMD Radeon HD 6870:
- 900MHz core clock
- 255mm² die size
- 4200MHz GDDR5, 256bit bus
- 512 KiB L2 texture cache
- 1120 shaders, 14 SIMD clusters, 16 FPU each, 5 wide
- 56 TMUs
- 32 ROPs
- $240
- 775MHz core clock
- 255mm² die size
- 4000MHz GDDR5, 256bit bus
- 512 KiB L2 texture cache
- 960 shaders, 12 SIMD clusters, 16 FPU each, 5 wide
- 48 TMUs
- 32 ROPs
- $180
Architecture
This new "Barts" core is not very different from "Cypress", Radeon 5800 series cores:
DisplayPort 1.2 And 6 Monitor EyeFinity
The new cards support a yet unfinished specification, DisplayPort 1.2. DP 1.2 allows AMD to include EyeFinity 6 monitor support with only two mini DP 1.2 outputs, something previously impossible, which required 6 outputs and were present only in the very rare Radeon HD 5870 EyeFinity 6 cards.
It will be interesting to see if this support will also be present in lower end cards, as not everyone needs such graphics/compute power but might need multiple displays for work.
Performance
This new "Barts" core is not very different from "Cypress", Radeon 5800 series cores:
- Same core layout, cut down from 20 SIMD clusters to 14, still 5 wide units with a transcendental unit per SIMD unit.
- Loses double precision, 64 bit computing capabilities.
- Lower complexity memory controller derived from 5500 cards. Takes less die space but can't reach the same memory clock speeds as "Cypress".
- Tesselation performance has improved.
- Improved anisotropic filtering and anti aliasing algorithms.
There's not much more to add to this, aside that by keeping 32 ROPs, it stacks up pretty well against the old 5830 and the new GeForce GTX 460.
This new chip pretty much reinforces my previous opinion of the "Cypress" core, in the sense that this chip is a whole lot more balanced that the previous one(inquiring minds should have a look at the last four articles in the "Sources" section of this article). We now have no mentions of the hilarious "power viruses" problem that plagged previous x800 cards, while maintaining good performance for your buck.
Where "Cypress" was a bandwidth starved card with a massive overshoot in compute power, this one is balanced enough to provide around 10% performance drop in a 25% smaller core, sometimes being even more competitive, which is in deed a job well done in the performance department.
DisplayPort 1.2 And 6 Monitor EyeFinity
The new cards support a yet unfinished specification, DisplayPort 1.2. DP 1.2 allows AMD to include EyeFinity 6 monitor support with only two mini DP 1.2 outputs, something previously impossible, which required 6 outputs and were present only in the very rare Radeon HD 5870 EyeFinity 6 cards.
It will be interesting to see if this support will also be present in lower end cards, as not everyone needs such graphics/compute power but might need multiple displays for work.
Performance
Since we are mostly looking at less than 10% of performance difference between the main competitors, I say that price is then the deciding factor. If you look at the "Price Check" section, you'll see that the GTX 460 is very competitively priced right now, so that would make it my favorite pick, on par with the 5850 from MSI. If AMD can bring the price to par, it becomes a compute capability vs features vs performance game, one that can be better decided by the buyer himself.
If you're looking for the ultimate performance, you shouldn't be looking at this review at all - or maybe it will reinforce a thought that the GeForce GTX 480 is a powerhouse, CrossFire and 5970s be damned.
One can't argue that the new cards don't fit their pricing range, sadly it is not fitting of the 6800 nomenclature.
Compute Capabilities
The new "Barts" chips lack a 64 bit FPU. This is regrettable. This is not like a $200 chip should look like in 2010 and has pretty much guaranteed that these cards(or a similar 7 series one) may be on their way out from future Apple products. Sure Apple is selling 32 bit FPU only products today but will they also not realize that this is limiting the usage of OpenCL?
The matter of fact is that Nvidia, by releasing a full "Fermi" lineup with 64 bit support(thus far), will create an ecosystem likely to see even more adoption of GPU computing in the desktop, laptop and enterprise space. There is simply some situations where there is the need to calculate with 64 bit precision and having to move code back to the CPU for calculation will kill performance. The even worse alternative is to code alternate paths for GPUs with and without 64 bit FPU support, which would cause the need to maintain both CPU and GPU versions of said code - this is not the way to motivate people to write already hard enough parallel code.
There isn't much need for very advanced double precision performance, as low as 1/10 may be enough to satisfy most code, as these calculations are mostly done sparsely where strictly needed due to the obvious performance benefit of running 32 bit everywhere possible.
Model Numbers
The new "Barts" chips lack a 64 bit FPU. This is regrettable. This is not like a $200 chip should look like in 2010 and has pretty much guaranteed that these cards(or a similar 7 series one) may be on their way out from future Apple products. Sure Apple is selling 32 bit FPU only products today but will they also not realize that this is limiting the usage of OpenCL?
The matter of fact is that Nvidia, by releasing a full "Fermi" lineup with 64 bit support(thus far), will create an ecosystem likely to see even more adoption of GPU computing in the desktop, laptop and enterprise space. There is simply some situations where there is the need to calculate with 64 bit precision and having to move code back to the CPU for calculation will kill performance. The even worse alternative is to code alternate paths for GPUs with and without 64 bit FPU support, which would cause the need to maintain both CPU and GPU versions of said code - this is not the way to motivate people to write already hard enough parallel code.
There isn't much need for very advanced double precision performance, as low as 1/10 may be enough to satisfy most code, as these calculations are mostly done sparsely where strictly needed due to the obvious performance benefit of running 32 bit everywhere possible.
Model Numbers
I bring these issues about model numbers up plenty of times because if I was a system integrator, I'd be pretty pissed to have to explain customers why AMD wants to get to their wallet with deceiving model numbers and, by extension, myself. This is not something I take lightly: if you're selling a product, you should leave the customer with clear choices. You should only point them to some direction if they ask you too.
It's true that most people don't know what they're looking for in a computer, but a little dialog is extremely helpful and decent model numbers would go a long way for a person to be happy when upgrading and keep coming back for more. If you deceive them the first time, they're likely to defect to another vendor. It also doesn't help brick and mortar store sellers, because they don't know what they're selling most of the time, so you're deceiving two people in a row. (ah... society...)
There's no denying that this is not the right move for numbering on these parts. Performance is hovering around 10% less of equivalent 5800 parts, sometimes at par, so even though the price is lower, this is sure to be causing confusion among buyers. I had two other options, given that the new chips don't bring major features that justify the move to 6 series numbering, let alone that they fit the 800 bracket. I get the "bring the 800 models back to $200 range" argument but it would be better fit for a time when the new 800 cards would actually be faster than the old ones.
It's true that most people don't know what they're looking for in a computer, but a little dialog is extremely helpful and decent model numbers would go a long way for a person to be happy when upgrading and keep coming back for more. If you deceive them the first time, they're likely to defect to another vendor. It also doesn't help brick and mortar store sellers, because they don't know what they're selling most of the time, so you're deceiving two people in a row. (ah... society...)
There's no denying that this is not the right move for numbering on these parts. Performance is hovering around 10% less of equivalent 5800 parts, sometimes at par, so even though the price is lower, this is sure to be causing confusion among buyers. I had two other options, given that the new chips don't bring major features that justify the move to 6 series numbering, let alone that they fit the 800 bracket. I get the "bring the 800 models back to $200 range" argument but it would be better fit for a time when the new 800 cards would actually be faster than the old ones.
Re-branding The 5770 Another Way
The other issue is about rebranding. It still still uncertain if AMD will rebrand the 5770 and which way, although rumors were pointing to somewhere in the 6700 models. Lets look at a case that includes re-branding but doesn't use the 6700 models for "Juniper" but keeps it to "Barts", which would have left the 6800 series for a real replacement and not a card with inferior performance.
If "Turks" is faster than Juniper...
Barts XT - Radeon HD 6770
Barts Pro - Radeon HD 6750
Turks XT - Radeon HD 6670
Turks Pro/LE - Radeon HD 6650
Turks XT - Radeon HD 6670
Turks Pro/LE - Radeon HD 6650
Juniper XT, Radeon HD 5770 - Radeon HD 6630
Juniper LE, Radeon HD 5750 - Radeon HD 6610
There supposedly will be another set of cards, "Turks", with similar performance to the Radeon HD 5770, which would then fit the Radeon HD 6750, 6770 slots. This would make sense if the "Turks" cards have better performance, if they're slower than the 5700, move them to 6730 and 6710 and give the newcomers the 6650 and 6670 slots. Or go nuts if it fits the performance ladder better, if "Turks" is yet slower, which this mess seems to be pointing to:
Radeon HD 5770 - Radeon HD 6670
Turks XT - Radeon HD 6650
Radeon HD 5750 - Radeon HD 6630
Turks Pro/LE - Radeon HD 6610/55x0
As this solution fits better the problem that LE/PRO models probably feature lower speed GDDR3 memory that cripples their performance, though those cards have generally been sent to the x500 models.
Come on, there's plenty of slots to go around:
Radeon HD 5770 - Radeon HD 6670
Turks XT - Radeon HD 6650
Radeon HD 5750 - Radeon HD 6630
Turks Pro/LE - Radeon HD 6610/55x0
As this solution fits better the problem that LE/PRO models probably feature lower speed GDDR3 memory that cripples their performance, though those cards have generally been sent to the x500 models.
Come on, there's plenty of slots to go around:
- Radeon HD 6500
- Radeon HD 6400
- Radeon HD 6300
- Radeon HD 6200
Lots. It's plenty to keep the lineup consistent and when the 6500 is considerably faster than the 5500 cards, it's evolution, it's the way it should move, forward, not backwards. There's a reason the first number is reflective of the generation: the uninformed customer doesn't know it, they just look at the full number.
It would do so much to avoid this scenario where a 6900 card was dual chip on the previous gen and is now only single chip and with probably too less of a bump over the current 5800 to justify it.
It would do so much to avoid this scenario where a 6900 card was dual chip on the previous gen and is now only single chip and with probably too less of a bump over the current 5800 to justify it.
Don't Rebrand
Hell, it's an old chip, keep it that way. The performance is still good and companies like Apple will have no issue in selling them for some time, just like it happened with the Radeon HD 4800 series. Smart buyers may do the same, if the price is right, and nobody is being deceived by shady marketing practices.
Also, given the amount of difference between two series, especially features, none of these chips should be re-branded! What does calling a card range the "Radeon HD 6 series" if then you have lots of cards with completely different features? Gladly most people won't even notice, especially since DirectX 11 is supported either way, but what about the ones that specifically want UVD3, DP1.2, HDMI 1.4a and might end up purchasing the wrong card? Things like EyeFinity support can be confusing enough as is, it becomes worse when re-branding enters the scene.
Keep it simple.
Final Notes
In the end, this is not an exciting new launch I had anticipated, much less one hyped around the internet as a highly tweaked architecture that would deliver so big performance increases that the "Barts" core would deserve to be an 6800 model. Such has not been the case and although delivering a good amount of performance for the buck, the lack of compute capabilities and rudimentary PCB - compared to 5850's digital VRM - make me suggest that a Radeon HD 5850 at inventory clearance price will make for a much more decent upgrade, as is the Nvidia GeForce GTX 460 1GiB, now priced at $200. The moderately small drop in performance on both cards(disregarding few exceptions) more than makes up for it with enhanced compute capabilities that may prove very useful in the near future. The better part? The GTX 460 768 MiB is selling for $139 after a mail in rebate, $159 regular price, which pretty much rests my case, expect if you're really out for any new feature that the 6000 series brings.
If you want more performance, wait for the 6900 series cards or go with a GeForce GTX 480, although the GTX 470 is faster in some cases, it is not distinctively enough to call it a big win, which is expected since it's new price is $260, $20 more than the Radeon HD 6870.
Architecturally wise, the 6900 series is promising to be a very worthwhile upgrade to 5800 cards, although coming at a price, one that shouldn't surpass the prices seen during most of the early life of the 5800 series cards - possibly at $300 and $400 price points.
Price Check
Galaxy GeForce GTX 460 768 MiB (with that awesome cleanable cooler) - $139 after mail in rebate.
PNY GeForce GTX 470 1280MiB - $239 with Mafia 2 for free.
MSI Radeon HD 5850 Twin Frozer II 1GiB (higher quality hardware than most cards) - $195 after mail in rebate.
Sapphire Radeon HD 6850 1 GiB - $179
Sapphire Radeon HD 6870 1 GiB - $239
Sources
TechPowerUp
Radeon HD 5800, Built To Be Big
Radeon HD 5870 Reviews
Radeon HD 5850 Reviewed
Also, given the amount of difference between two series, especially features, none of these chips should be re-branded! What does calling a card range the "Radeon HD 6 series" if then you have lots of cards with completely different features? Gladly most people won't even notice, especially since DirectX 11 is supported either way, but what about the ones that specifically want UVD3, DP1.2, HDMI 1.4a and might end up purchasing the wrong card? Things like EyeFinity support can be confusing enough as is, it becomes worse when re-branding enters the scene.
Keep it simple.
Final Notes
In the end, this is not an exciting new launch I had anticipated, much less one hyped around the internet as a highly tweaked architecture that would deliver so big performance increases that the "Barts" core would deserve to be an 6800 model. Such has not been the case and although delivering a good amount of performance for the buck, the lack of compute capabilities and rudimentary PCB - compared to 5850's digital VRM - make me suggest that a Radeon HD 5850 at inventory clearance price will make for a much more decent upgrade, as is the Nvidia GeForce GTX 460 1GiB, now priced at $200. The moderately small drop in performance on both cards(disregarding few exceptions) more than makes up for it with enhanced compute capabilities that may prove very useful in the near future. The better part? The GTX 460 768 MiB is selling for $139 after a mail in rebate, $159 regular price, which pretty much rests my case, expect if you're really out for any new feature that the 6000 series brings.
If you want more performance, wait for the 6900 series cards or go with a GeForce GTX 480, although the GTX 470 is faster in some cases, it is not distinctively enough to call it a big win, which is expected since it's new price is $260, $20 more than the Radeon HD 6870.
Architecturally wise, the 6900 series is promising to be a very worthwhile upgrade to 5800 cards, although coming at a price, one that shouldn't surpass the prices seen during most of the early life of the 5800 series cards - possibly at $300 and $400 price points.
Price Check
Galaxy GeForce GTX 460 768 MiB (with that awesome cleanable cooler) - $139 after mail in rebate.
PNY GeForce GTX 470 1280MiB - $239 with Mafia 2 for free.
MSI Radeon HD 5850 Twin Frozer II 1GiB (higher quality hardware than most cards) - $195 after mail in rebate.
Sapphire Radeon HD 6850 1 GiB - $179
Sapphire Radeon HD 6870 1 GiB - $239
Sources
TechPowerUp
Radeon HD 5800, Built To Be Big
Radeon HD 5870 Reviews
Radeon HD 5850 Reviewed
No comments:
Post a Comment