Intel Broadwell (Gen8)

So new graphics features beyond dx12.

Nice.

 

Remember, API and hardware feature level are separate things in DirectX. I actually disagree with them being named so similarly... it would be better if feature levels were A,B,C or something The DirectX 12 API will run on a lot of feature level 11_x devices... new hardware "feature level" requirements obviously imply new hardware.

 

Hay all I said was now I have to destroy one on utube . I did not say I would give details how I ignite it. Many videos made with details left out. I just think it be cool to ignite one and than through Water on it . to show that reaction . But thanks for comment . I will leave comments off on the utube video . Don't want people getting the wrong idea. lol

 

Does intels present Gpu have a feature level that others don't have. I sure you guys know much more about than I . But the Smoke off the tires in Dirt looks pretty cool compared to what others offer. It adds a bit more realizism to the Game . Now if someone could do smell. so we could sense the smell of burning rubber or even a clutch burning up would be neat. Something like when i was a kid with train set, you drop something into smoke stack to cause smoke . I had that when i was 5 that was 58 years ago . I still have that full sized trainset in the box . I can buy 30 Pcs with what it valued at today. I even have the smoke drops.

 

Intel's PixelSync, or rather the pixel shader ordering that goes into it, is something I've not seen mentioned for competing GPUs.
One of the first things noted about DX12 was the introduction of a pixel-ordered UAV, for the purposes of efficient order independent transparency.
It's something that I've not seen a strong case made that the other big GPU makers can currently match.

 

Incidentally, Maxwell just announced that they support it now too (raster ordered views) Awesome that the industry is adopting such a useful feature.

 

The graphics hardware in Gen 8 can support cl_khr_fp64
From
https://software.intel.com/en-us/forums/topic/531928

"Our compute architects are having internal discussions about enabling cl_khr_fp64. If there are enough customer requests for this feature, we will do it."

I've started a thread in GPGPU section here for it to be discussed.

 

If their hardware supports it, why aren't they just flipping on the switch? Very strange!

 

There's probably a bit more to it than that. Do you think their GPU supports sin(double) for example?

 

Intel Releases Broadwell-U: New SKUs, up to 48 EUs and Iris 6100

Additionally, there are minor CPU architecture improvements:

 

I have not been following Broadwell too closely, but will there be desktop variants? I thought there were rumors of no.

 

All the new DX11.3 features are very useful (and will solve actual problems in modern graphics engines). 11.3 certainly seems like a bigger update than 11.1 and 11.2. ROVs are nice, but I also do like typed UAV loads and conservative rasterization a lot. Also conservative rasterization and ROVs combine nicely together.

 

Still no OpenGL 4.4 support (= no MDI support). It was expected, but makes me sad

 

I must say so far I'm disappointed by Broadwell graphics from the early numbers (tomshardware has some). Gen8 is a significant change in architecture compared to 7/7.5, yet the numbers show more of a scaling inline with just shader alu numbers increase.
Granted these are early numbers, it is also possible that (even both were 15W parts) the effective power consumption was lower. Driver might have some deficiencies too but I sure hope a ~20% performance increase within the same power envelope isn't all intel could muster in those Broadwells (would kinda indicate the architecture of Gen8 isn't really any better than Gen7 was). Sure it has other advantages (new features, might also scale down better to even lower TDPs) but still as far as I can tell perf/w wasn't quite class leading with Gen7 (not that it was terrible but not at maxwell level) so I hoped for some significant improvement there.

 

I'm quite aware of that but the units also should be rebalanced (those 24 EUs now have 3 quad tmus instead of just 2 for 20 EUs). But a 20% increase just looked like it was a bit more alus, some better manufacturing...

Well I'm not sure about the clocks (in the 15W units at least). The difference between max clock and what was actually achieved under a sustained load was significant, I have no idea if the clocks are going to be lower or higher now in practice. But well if the difference in actual applications is more like 30%-50% rather than 20% that's more like it. I guess won't have to wait too long for some real reviews...

 

Alright kids this isn't the console forums. In general, please don't present assumptions on unreleased products (or stuff you don't actually 100% know) as fact. This tends to bring out strong reactions and quickly dilutes the discussion.

In addition, any discussion about the quality of Anandtech (or any site really) should be done in a separate thread.

 

Good read: https://software.intel.com/sites/default/files/managed/71/a2/Compute Architecture of Intel Processor Graphics Gen8.pdf

If I understood correctly, this means that Intel Gen8 has 2x FP16 rate, just like Nvidia and Imagination. INT16 peak is also doubled (that is also nice for many purposes). I however wonder why the peak FP16/INT16 numbers are not shown in the peak compute throughput table (it only shows peak 32i, 32f, and 64f). I wonder if there are some other architecture limitations that cap the peak rates of 16 bit operations.

This is nice. Modern/future software will definitely benefit from this. We use lots of integer operations nowadays. Obviously older (last gen / cross gen) console ports will not see any gains (as last gen console GPUs didn't support integers).

Excellent. I wonder if there is additional information available about this? I would be interested to know what kind of penalties you pay for globally coherent pages (bandwidth, latency). Is there any examples already available showing tightly interleaved CPU & GPU work (both reading atomics written by each other and coordinating work)? With the huge shared L3 caches (and even L4 on some configurations) this would be totally awesome, as the CPU<->GPU traffic wouldn't need to hit memory at all.

Gen7.5 was already highly bandwidth starved (without the 128 MB EDRAM). 20%-30% gains without any BW improvements are quite nice. I don't see how they could have improved the performance much further with the same memory configuration (DDR3/LPDDR3 1600). Let's wait for the EDRAM equipped models (with 48 EUs). Desktop / server models likely support DDR4 as well.