I saw an empty solder space for power regulation on the 5870.Originally posted by: Janooo
The latest rumor mill says that Cypress has more than 20 SIMDs.
It's hard to believe but apparently there is no Cypress die picture in the wild wide web. AMD could be hiding something.
Originally posted by: BFG10K
This thread has been started in anticipation of the 5870/5850 reviews.
Please keep all review discussion for the 5870/5850 in this thread.
Please link to reviews as you find them and I?ll keep the front page updated.
Video Mod BFG10K.
5850
AnandTech 5850 Review
http://techreport.com/articles.x/17652
http://www.hardocp.com/article...5850_video_card_review
http://hothardware.com/Article...850-Performance-Review
http://www.firingsquad.com/har...0_performance_preview/
http://www.computerbase.de/art...au_ati_radeon_hd_5850/
http://www.guru3d.com/article/...5850-review-crossfire/ (crossfire)
5870
AnandTech
http://www.tomshardware.com/re...deon-hd-5870,2422.html
http://www.xbitlabs.com/articl...lay/radeon-hd5870.html
http://techreport.com/articles.x/17618
http://www.computerbase.de/art.../#abschnitt_einleitung
http://www.hardware.fr/article...md-radeon-hd-5870.html
http://www.hardocp.com/article...5870_video_card_review
http://www.techpowerup.com/rev.../Radeon_HD_5870/1.html (includes 5850).
http://www.techpowerup.com/rev...0_PCI-Express_Scaling/ (PCIe scaling).
http://www.firingsquad.com/har...0_performance_preview/
http://www.guru3d.com/article/...n-hd-5870-review-test/
http://www.legitreviews.com/article/1080/1/
http://www.tweaktown.com/revie...aphics_card/index.html
http://www.pcper.com/article.php?aid=783
http://www.legionhardware.com/document.php?id=858
http://www.hardwarecanucks.com...-1gb-gddr5-review.html
http://www.hexus.net/content/item.php?item=20289
http://www.hexus.net/content/item.php?item=20354 (crossfire)
http://www.pcgameshardware.com...graphics-card/Reviews/
http://www.driverheaven.net/reviews.php?reviewid=842
-----
Guys, comments like this are not permitted in the Video forum:
Originally posted by: OneOfTheseDays
It's sad to see the Nvidia fanboys in full force here. Unfortunately it seems that even the moderators cannot be free from their own inherent bias.
If you're coming in here for the 5 series launch as a regular of the social forums or are new to AnandTech, please be aware that personal attacks and call outs are not allowed in the technical forums -- see the stickied Moderator message at the top of this forum. If you're a Video regular, you shouldn't need to be reminded of these guidelines .
AmberClad
[the other] Video Moderator
I tend to agree with you. Whether it?s intentional or whether the drivers simply need work, I don?t think we?ve seen the full performance of the 5870/5850 yet.Originally posted by: Idontcare
I really think AMD is being wiley this time around, I have high expectations of them "dropping the other shoe" once Nvidia fully lays down their hand.
But how? I mean, I'm always for more performance/features, but I really can't tell where AMD could be hiding something. The number of transistors, units, die size all line up against previous gen. 850MHz is also a jump from previous generation, and 5870's power consumption is also comparable to that of previous gens.Originally posted by: BFG10K
Whether it?s intentional or whether the drivers simply need work, I don?t think we?ve seen the full performance of the 5870/5850 yet.
Originally posted by: lopri
But how? I mean, I'm always for more performance/features, but I really can't tell where AMD could be hiding something. The number of transistors, units, die size all line up against previous gen. 850MHz is also a jump from previous generation, and 5870's power consumption is also comparable to that of previous gens.Originally posted by: BFG10K
Whether it?s intentional or whether the drivers simply need work, I don?t think we?ve seen the full performance of the 5870/5850 yet.
1.0 GHz core sounds a bit too optimistic, and there wasn't a precedent of 'magic drivers' that I know of from AMD.
Only thing I can reasonably think of is a different grade/amount of GDDR5, and I'd think that'd be far short of being a wild card?
Originally posted by: Idontcare
Well stated HurleyBird, given the specs of Cypress versus those of 2xRV770, we have no choice but to conclude that either HD5870 (the product) has a driver-based limitation (intentional or not) or Cypress (the GPU) has an architecture bottleneck (intentional or not, presumably not) that is not present in RV770.
There are some interesting ramifications to the outcomes of such a logic tree.
For example, if it is an unavoidable consequence of the inherent scaling limitations of the architecture itself then that would speak to the likelihood of Cypress being the end of the road for this underlying architecture, the next shrink would not likely be a simple doubling yet again of the varying architectural components as that would yield an even lower percentage increase in performance.
If it is an unintentional consequence of the design, i.e. an accidental handicap or bottleneck within the design itself that wasn't present in RV770 then it might be possible for AMD to rectify the "bug" by investing resources into doing a respin. Depends just how deeply integrated the handicap is with the existing architecture and design.
If it is driver related then presumably at some point over time the performance will eventually be extracted by future driver releases, free performance for existing customers.
Originally posted by: BFG10K
I tend to agree with you. Whether it?s intentional or whether the drivers simply need work, I don?t think we?ve seen the full performance of the 5870/5850 yet.Originally posted by: Idontcare
I really think AMD is being wiley this time around, I have high expectations of them "dropping the other shoe" once Nvidia fully lays down their hand.
Originally posted by: Janooo
A PC is very complex system. Bottlenecks are dynamic. They shift based on applied changes.
It just might be that by doubling the specs, 5870 shifted some bottlenecks outside of video card. Let's not forget we don't have 100% CPU increase. It's actually 0% if the cards were tested on the same system.
Originally posted by: Janooo
A PC is very complex system. Bottlenecks are dynamic. They shift based on applied changes.
Originally posted by: Janooo
It just might be that by doubling the specs, 5870 shifted some bottlenecks outside of video card. Let's not forget we don't have 100% CPU increase. It's actually 0% if the cards were tested on the same system.
Originally posted by: HurleyBird
Originally posted by: Janooo
A PC is very complex system. Bottlenecks are dynamic. They shift based on applied changes.
It just might be that by doubling the specs, 5870 shifted some bottlenecks outside of video card. Let's not forget we don't have 100% CPU increase. It's actually 0% if the cards were tested on the same system.
Except that the performance numbers don't look like they are coming from a bottleneck outside the graphics card. If that were the case, we wouldn't see the 4870X2 or 4890CF outperforming the 5870, we would see them perform the same.
Originally posted by: Janooo
The driver can take better advantage of multicore in AFR mode than with a single GPU.
Originally posted by: Idontcare
Originally posted by: Janooo
The driver can take better advantage of multicore in AFR mode than with a single GPU.
How do we (you) know this to be the case? That's not a rhetorical question, I was unawares we had access to this information.
Originally posted by: Janooo
Originally posted by: Idontcare
Originally posted by: Janooo
The driver can take better advantage of multicore in AFR mode than with a single GPU.
How do we (you) know this to be the case? That's not a rhetorical question, I was unawares we had access to this information.
The AFR renders two frames in parallel. The driver is multithreaded and each card has its own threads. Does it make sense?
In the drivers.Originally posted by: lopri
But how? I mean, I'm always for more performance/features, but I really can't tell where AMD could be hiding something.
This is true to some degree (though not as much as you think), but it?s not really relevant since it equally applies to previous generation video cards, and it also doesn?t apply in what are clearly GPU limited situations, which is what we?re discussing.Originally posted by: Janooo
The driver can take better advantage of multicore in AFR mode than with a single GPU. The CPU load is different between AFR and non AFR modes.
Sure there is, given overclocking the 5870 still increase performance. And because the core raises performance more than memory by a factor of 2.5 to 1, that means the core is the primary bottleneck. This lends credibility to my two possible driver scenarios (intentional vs unintentional) that I listed earlier.So I am not sure if it's that easy to compare and say that there is no CPU bottleneck.
Originally posted by: Idontcare
Originally posted by: Janooo
Originally posted by: Idontcare
Originally posted by: Janooo
The driver can take better advantage of multicore in AFR mode than with a single GPU.
How do we (you) know this to be the case? That's not a rhetorical question, I was unawares we had access to this information.
The AFR renders two frames in parallel. The driver is multithreaded and each card has its own threads. Does it make sense?
It does make sense, but cypress is multithreaded as well and has the resource capability to process twice the threads on its own...so why should the driver care what hardware is processing those threads? 2xRV770 or 1xCypress.
It is like arguing that a dual-core single-socket cpu would be slower than a dual-socket single-core cpu just because the software is multithreaded...an argument that makes no sense to me. It is six of one, half dozen of the other.
So what am I missing here?
Originally posted by: Janooo
Originally posted by: Idontcare
Originally posted by: Janooo
Originally posted by: Idontcare
Originally posted by: Janooo
The driver can take better advantage of multicore in AFR mode than with a single GPU.
How do we (you) know this to be the case? That's not a rhetorical question, I was unawares we had access to this information.
The AFR renders two frames in parallel. The driver is multithreaded and each card has its own threads. Does it make sense?
It does make sense, but cypress is multithreaded as well and has the resource capability to process twice the threads on its own...so why should the driver care what hardware is processing those threads? 2xRV770 or 1xCypress.
It is like arguing that a dual-core single-socket cpu would be slower than a dual-socket single-core cpu just because the software is multithreaded...an argument that makes no sense to me. It is six of one, half dozen of the other.
So what am I missing here?
Let me explain in a simplified way.
Let's say when a frame is rendered the driver executes A, B, C stages in a sequence. A, B, C are serial by definition. I am not an expert and there could be more things that have to go in a sequence but I am sure these serial stages exist.
When we have gpu1 and gpu2 in AFR mode then frame1 A1, B1, C1 can be executed on core1 for gpu1 and frame2 A2, B2, C2 can run on core2 for gpu2. They run in parallel.
When you have single gpu then A1, B1, C1, A2, B2, C2 will always go in a sequence. There is no A2 till C1 is finished.
I hope it makes sense to you.
Therefore comparing dual gpus solution to a single gpu one is trickier in regards to the cpu bottleneck. If somebody tested with a single core I am sure we would see the difference in performance.
SIMD two groups work seamlessly
AMD GPU, the processor is a SIMD (Single Instruction, Multiple Data) is a unit to perform the "SIMD cores (SIMD engine, also known as)" They are presented as. RV770 does this total was for 10 SIMD cores, Cypress has become pieces in 20. Cypress is a dual-core entities, RV770 equivalent to 1 as ten SIMD cores, 20 cores seems to be two things to manage the split core.
Cypress elements have shown that the structure of the dual-core, Cypress can be seen throughout the architecture. For example, Cypress offers the rasterizer to convert the polygon pixel is increased to two each. This is acceptable given per core and configuration of taking the rasterizer. In addition, the full set of threads allocated to each processor, and also to control Sureddoshikensa branching and thread switching are two pieces of the program. Cypress total of 20 SIMD cores, the 10 that are controlled Sureddoshikensa one by one. AMD officials have described as follows.
"Sureddoshikensa conducted by the branch control. RV770 in the 10's was one of SIMD. Cypress, the SIMD 10 and 20 of each group, each managed by Sureddoshikensa"
However, each was divided into two core Cypress (10 SIMD) is, CPU is much more tightly connected to the dual core.
"The software has two work seamlessly as a single core. For example, (thread) dispatch, and has been able to make load balancing across both cores. Therefore, here are the vertices of the polygon processing core, the core can dispatch and the other after processing pixel rasterization. in order to transfer data from one core from the core to the other, no need to write once memory.
The texture (data path) are (two cores) are shared. Texture is prevalent in two cores go. Texture (data path) are two core services across "(Kruger's)
In other words, the processor Sureddodisupatcha control group, seemed to be some control can be conducted across two cores. Therefore, the two can be considered to keep the load balancing operation processor cores busy.
Original in Japanese
Google translated to English
Originally posted by: Idontcare
Cypress dual-core
Compute Unit Comparison: R600 vs. RV770 vs. Cypress
Originally posted by: Janooo
^
Let's have a look at AFR and a single GPU time sequence. 'x' will represent time spent on GPU, A1, B1, ... is time spent on CPU.
AFR:
A1xxxxB1xxxxC1xxxx
A2xxxxB2xxxxC2xxxx
Single 2x faster GPU:
A1xxB1xxC1xxA2xxB2xxC2xx
I know this is an extreme situation but what I am trying to say is that A1, B1, C1 go in parallel with A2, B2, C2 for AFR and they have to go in sequence for a single GPU even though it's 2x faster. Therefore under this conditions 2x faster GPU can not be 2x faster overall.
Originally posted by: Idontcare
Originally posted by: Janooo
^
Let's have a look at AFR and a single GPU time sequence. 'x' will represent time spent on GPU, A1, B1, ... is time spent on CPU.
AFR:
A1xxxxB1xxxxC1xxxx
A2xxxxB2xxxxC2xxxx
Single 2x faster GPU:
A1xxB1xxC1xxA2xxB2xxC2xx
I know this is an extreme situation but what I am trying to say is that A1, B1, C1 go in parallel with A2, B2, C2 for AFR and they have to go in sequence for a single GPU even though it's 2x faster. Therefore under this conditions 2x faster GPU can not be 2x faster overall.
But Janooo isn't Cypress a wider GPU? Meaning it can process those A1/A2, etc, in parallel just as readily as two RV770's would? Everything was doubled up.
From what I can determining the only thing you are doing here to make your argument true in print is enforcing an artificial constraint that Cypress is a serial processing GPU...what is it about Cypress that is making you think it can't handle multithreading?
The part I bolded, why must this be true? Why can't a wider GPU capable of handling parallel threads also be capable of doing the AFR sequence in parallel the same has having two discrete GPU's dispatching the sequence in parallel?
No, it?s doubled up but it can still only work on one frame at a time, unlike an AFR system. This is the inherent difference between a single GPU and AFR. With AFR, each GPU in the system can be working on a different frame concurrently.Originally posted by: Idontcare
But Janooo isn't Cypress a wider GPU? Meaning it can process those A1/A2, etc, in parallel just as readily as two RV770's would? Everything was doubled up.