I stopped reading when he mentioned Amdahl's law in this context...
Llano IGP vs SNB IGP vs IVB IGP
- Thread starter AnarchX
- Start date
I stopped reading when he mentioned Amdahl's law in this context...
Now that SB reviews are all over the net, anyone else here who fails to see why the IGP is praised so much in some of the reviews?
The way I see it, the only reason why Intel was able to catch up in performance is not that SB's IGP is so great, it's simply that both low-end discrete and AMD IGP performance have stagnated for a long, long time.
The 790GX chipset was released in August 2008, the Radeon 4550 September 2008. That's 2 1/4 years.
Furthermore, right now Intel still lags behind in [driver] stability and image quality. Performance hit from enabling AA/AF is much heavier than on a 5450 as well.
So Intel is getting closer, but I think they're not quite there yet.
The way I see it, the only reason why Intel was able to catch up in performance is not that SB's IGP is so great, it's simply that both low-end discrete and AMD IGP performance have stagnated for a long, long time.
The 790GX chipset was released in August 2008, the Radeon 4550 September 2008. That's 2 1/4 years.
Furthermore, right now Intel still lags behind in [driver] stability and image quality. Performance hit from enabling AA/AF is much heavier than on a 5450 as well.
So Intel is getting closer, but I think they're not quite there yet.
This is certainly a reason. But you can't fault intel for that.
Haven't seen AA results yet, but from the looks of it intel isn't interested in having somewhat fast graphics for the desktop at all. Why else would they only enable all 12 EUs in the K editions, which have unlocked cpu multipliers you can't use in the H67? Also, only support DDR3-1333 despite the iGPU could likely benefit from faster ram (which you can only use with P67) and despite mobile cpus actually supporting DDR3-1600?
I agree though the 6 EU version isn't really that exciting. Well it's not surprising with half the cores of the old Ironlake (no matter how improved they are and that they are clocked higher - clearly the architecture improvements are there).
EUs are still four-wide (128bit, as confirmed by anandtech), so 12 EUs are good for 48 flops per cycle. Or you can say twice that if you accept counting multiply-accumulate using accumulator reg (contrary to AMD/NVIDIA these intel igps can't do MAD only MAC) as two flops.
(So the gflops rating per clock did not actually change since last gen for the 12 EU version - though it should be much more easily possible to use the accumulator with Sandy Bridge, as you can now enable/disable it per instruction.)
Last edited by a moderator:
To be honest, I would consider a 60% hit rate for an 8 MB cache as piss poor. Intel uses to claim much higher hitrates for their CPUs (90+%).
In some twisted sense, you could already claim a 50% hitrate for a tiny buffer storing just a burst read from memory when you are doing sequential accesses. With an 128Bit interface this would be 128 Bytes and when working with a cache line size of 64 bytes you could claim that each second line comes from this buffer and is therefore a hit. But does it help anything?
Sharing the last level cache may have benefits of lower latency draw calls or can serve as an extended buffer for operations which write an intermediate amount of data to memory and reuse it later. But for most traditional graphics workloads it's probably hardly worth the added effort compared to old fashioned texture and color caches of GPUs (which serves mainly as "bandwidth amplifiers").
ltcommander.data
Regular
Well Tech Report showed the SB's IGP is twice as fast as AMD's 890GX, Anandtech showed the SB's IGP is comparable to the current fastest IGP which was introduced in April 2010, nVidia's 320M, and SB's IGP is faster than the HD 5450 so it is as superlative as you can get for an IGP. Anand's review did mention that Intel is now putting real levels of funding behind game and driver testing, which as you say is where we'll have to keep watching.
Unfortunately it's only really true for the 12 EU version, but the one you'd get on the desktop (for anything else than the K parts which are unlikely to be paired with H67) is the 6 EU version which isn't really a whole lot faster than AMD's IGP (which clearly shows its age). The situation seems to be much better with the mobile chips though indeed (not only do they have 12 EUs but the quad cores even support faster memory which is probably going to help a bit).
In any case, for the people saying SNB could challenge Llano in 3d performance, they need to think again. While I don't think Llano will be THAT fast, it should imho perform about the same as a HD5550 (ddr3 version - don't get confused by the ddr2 and gddr5 versions of this card...). From the looks of it, not even Ivy Bridge will even try to get there (with the rumors saying 16 EUs - hopefully for a bit more than only the "K" parts for desktop chips then...).
Yeah, in low settings in every game. Cranking it up to medium settings and it gets apparent that Sandy Bridge is still lackluster.
Imagine how nVidia's 320M would do with Sandy Bridge, compared to the built-in IGP.
On average maybe (although not by much), certainly in some games, but definitely not everywhere. In many games the 5450 is either on par or a bit faster.
We also shouldn't forget that AMD has a Mobility 5470, clocked 100 MHz higher than the desktop 5450.
Unfortunately I don't remember which review and game it was, but it showed HD 3000 dropping from 55ish frames to 22-23 with 4xAA/16AF, whereas the 5450 only dropped from 43 to 30 fps (game was tested in 1024x768 with low settings). Depending on the type of game, 30 fps can be quite playable, while 22-23 is a borderline case (although admittedly, even that can be enough for some games).mczak said:
Point taken, it really doesn't look like Intel cares much for the desktop graphics market right now.
Well, I guess they simply don't need to, all but the absolute cheapest PCs usually come with discrete solutions anyway, and most consumers who want the best CPU on the market will go for Intel, at least until BD hits (maybe even then).
It's a bit different in mobile space, where Intel's IGPs used to be - or, if we ignore SB for now since notebooks with SB are yet to reach the market in significant quantities, still are - a reason to avoid any notebook without a GPU from AMD or Nvidia.
Based on that (assuming the perf hit is mostly from AA, not AF) it looks like indeed MSAA is more of a checkbox feature (needed for DX10.1) instead of a serious implementation. I guess either the chip lacks buffer compression, and/or the chip can't do enough z tests/updates (either due to lack of capabilities in the rops or even limitations in message passing for the threads). Changing memory clock could give some hints.
In any case, interesting to see the performance difference vanishes between HD2000 and HD3000 - not unexpected, since the HD2000 simply lacks shader units so its performance drop isn't that large with MSAA. For HD3000 the drop is just huge, nearly approaching SSAA levels... Last chip I've seen with such large MSAA performance hit was probably GF4, though it remains to be seen if it fares any better with other games...
Strange, these AA results http://ht4u.net/reviews/2011/intel_sandy_bridge_sockel_1155_quadcore/index27.php show a very different picture. Performance drop quite comparable to HD 5450.
Going to 4x MSAA and 16x AF requires a large increase in memory bandwidth. These things are already BW starved as it is so it's not surprising to see discrete cards - even slow ones - pull away when you simply jack up the memory bandwidth requirements.
Blazkowicz
Legend
maybe the reviewer used a 5450 with gddr3 in the first link, and a 5450 with ddr2 was used in yours one.
And when AMD and Nvidia finaly arives at next process node(SB is 32nm) and decide to put more than 1-2 SIMD on those entry chips and use gddr5 only (not shity ddr3) than its prety much done.
But i have a feeling that the entry chips were designed to be so slow just because it was still lightyears before intels GMA. And prety much with no effort the same situation could happen if AMD and Nvidia decides to do so again and SB-s igp will be forgotten.(not like they didnt killed it by themself with the H and P mobos, and mainly 6EU gpus on desktop
)Well I mean they are forgotten because the reality of their 3D gaming non-potential sinks in. IGPs are largely used to run Aero and maybe video. They are there to make a cheap but "complete" budget PC package. The 3D hardware isn't completely worthless but it might as well be.
They will never be competitive with discrete cards because doing so requires increasing their cost considerably. They need much more bandwidth than they have and more GPU hardware. The vast majority of IGP users would not care to pay for that stuff because they won't use it. Most IGP users wouldn't notice a GMA 950 compared to the latest greatest. Personally I think that all they need to be able to do is suffice for HD video playback because then you can make a HTPC very cheaply. 780G, GF8200 and even GMA4500 already do that.
Also, the sideport-style secondary RAM is there to save power in mobile applications, not to improve performance tangibly (which it doesn't).
I could only see a considerably faster "gamer IGP" happening on some sort of premium motherboard and why on earth would anyone want to buy that instead of a discrete, replaceable gaming card? It would probably end up more like a discrete card integrated onto the mobo because it would need its own fast RAM supply.
They will never be competitive with discrete cards because doing so requires increasing their cost considerably. They need much more bandwidth than they have and more GPU hardware. The vast majority of IGP users would not care to pay for that stuff because they won't use it. Most IGP users wouldn't notice a GMA 950 compared to the latest greatest. Personally I think that all they need to be able to do is suffice for HD video playback because then you can make a HTPC very cheaply. 780G, GF8200 and even GMA4500 already do that.
Also, the sideport-style secondary RAM is there to save power in mobile applications, not to improve performance tangibly (which it doesn't).
I could only see a considerably faster "gamer IGP" happening on some sort of premium motherboard and why on earth would anyone want to buy that instead of a discrete, replaceable gaming card? It would probably end up more like a discrete card integrated onto the mobo because it would need its own fast RAM supply.
Last edited by a moderator:
The "smaller" version is fine for desktop work and anyone who wants to do 3D/gaming needs a discrete GPU still anyways. Gaming on any IGP is a neat little party trick but not a reasonably experience for the time being.(not like they didnt killed it by themself with the H and P mobos, and mainly 6EU gpus on desktop
And "throwing GDDR5" on an integrated part is not so obvious. If you're not sharing the memory bus with the CPU then you're basically a chipset-integrated/discrete part anyways with none of the advantages of being on-die. No I think it's far more likely that we have to shift how we do rendering to more bandwidth-friendly techniques liking binning/tiling, similar to what the phones and soon tablets are doing.
OCW oced HD Graphics 3000 to 1.9GHz: http://en.ocworkbench.com/tech/asrock-h67m-geht-reviewed-with-core-i5-2500k-with-gpu-overclocked/
So Intel may counter Llano with a clock upgrade?
So Intel may counter Llano with a clock upgrade?
A 0.2V voltage bump is pretty huge. I'm not sure what could be achieved at the standard voltage, or at least with a more reasonable increase. Honestly, even at 1.9GHz, I doubt "HD Graphics 3000" could match Llano's GPU.